--- oct25_all_in_one
+++ nov10_main_plus_companion
@@ -1,10 +1,14 @@
-# Why Michael Levin Is 81.6% Right: Between Galvani’s Frogs and Plato’s Ghosts
-
-*Testing his "Latent Platonic Space" claims with code, math, and 10,303 fish. In a sea of internet debates between "Platonic Ingression" and reductionist blog critics, an internet philosopher who can code doesn’t have to guess—every single core claim can be tested with tools at our fingertips (`93.4%` right on structural math, `0.0%` on non-physical ghosts).*
-
----
-
-For some time I have been participating actively in documenting the scientific potential of bioelectricity (which goes back to Galvani’s frog experiments; see [Bioelectricity Foundations](https://bioelectricity.net/bioelectricity-foundations)) and the various fascinating, though usually hypothetical, metaphysical concepts that often travel side by side with it ([bioelectricity.net](https://bioelectricity.net)). As a historian of math, it doesn’t bother me that sci-fi concepts (such as *psi*) and real science can develop side by side—historically, Kepler’s astrophysics and Newton’s alchemy grew out of the same soil—but I’m equally aware that religious effects can take over and obliterate free critical thinking (including in Buddhist traditions that bind one to absolute obedience to a guru).
+# Auditing Michael Levin’s *Ingressing Minds*: 88.1% on the Math, 0% on the Ghosts (77.6++% Overall)
+
+For some time I have been participating actively in documenting the scientific potential of bioelectricity (which goes back to Galvani’s frog experiments; see [Bioelectricity Foundations](https://bioelectricity.net/bioelectricity-foundations)) and the various fascinating, though usually hypothetical, metaphysical concepts that often travel side by side with it ([bioelectricity.net](https://bioelectricity.net)). As a [historian of math](https://mitpress.mit.edu/author/joel-dietz-39191/), it doesn’t bother me that sci-fi concepts (such as *psi*) and real science can develop side by side—historically, Kepler’s astrophysics and Newton’s alchemy grew out of the same soil—but I’m equally aware that religious effects can take over and obliterate free critical thinking (e.g. guru traditions).
+
+![](https://fractastical.github.io/morpholib/media/frog_bioelectricity_triptych.gif)
+
+*Plate I — Three frogs, one bioelectric thread. Top left: Galvani’s 1791 bimetallic arc closes on the sciatic nerve and the frog leg twitches. Top right: the “face of a frog,” a time-lapse of the voltage pre-pattern that maps out the* Xenopus *face before it forms ([Vandenberg, Morrie & Adams 2011](https://doi.org/10.1002/dvdy.22685)). Bottom: our GCaMP Ca²⁺ imaging of two* Xenopus laevis *embryos (WGD84); a wave in one is followed by a wave in its untouched neighbour, with the tracked wavefront vectors in magenta. Provenance: [`models/build_frog_bioelectricity_triptych.py`](https://github.com/fractastical/morpholib/blob/main/models/build_frog_bioelectricity_triptych.py), composited from [`models/1791-galvani-bioelectricity.py`](https://github.com/fractastical/morpholib/blob/main/models/1791-galvani-bioelectricity.py), the time-lapse as hosted on [bioelectricity.net](https://bioelectricity.net), and [`wave-vector-analysis/render_pretty_wave_vectors.py`](https://github.com/fractastical/morpholib/blob/main/wave-vector-analysis/render_pretty_wave_vectors.py).*
+
+![](https://fractastical.github.io/morpholib/media/thompson_transformations.gif)
+
+*Plate II — Thompson’s rubber sheet, moving. A real parrotfish outline is bent through four of his 1917 coordinate families: the 70° shear, the coaxial circles that turn Scarus into an angelfish, the bilinear fan, and the radial flare. Provenance: [`thompson/1917-thompson.py`](https://github.com/fractastical/morpholib/blob/main/thompson/1917-thompson.py).*
 
 When I was working at the California Institute of Machine Consciousness in early 2026, my then-employer Joscha Bach would often elatedly insert concepts of *morphogenesis* into conversations. This term means, roughly, how things get to be their current shape, and tends to be a pretty mathy set of explanations.
 
@@ -14,325 +18,148 @@
 
 This satisfied her, but she said that until I explained it that way, most of the Mike Levin orbit concepts sounded instead impermeable or much like "woo science."
 
-In my case, I was fortunately aided by Mike Levin himself, who gave me a list of articles to read in which order to get started on learning the space, to which I added a massive pile of his papers, which I synthesized with a knowledge graph ([Project Delphi](https://github.com/fractastical/project-delphi-ontology)). That knowledge graph then evolved into a three-stage series of open-source verification engines—starting with **NeuroEval** ([doi:10.5281/zenodo.20622878](https://doi.org/10.5281/zenodo.20622878), evaluating data transparency and neural-coding rigor in computational and biohybrid/wetware neuroscience papers), followed by **BioEval** ([doi:10.5281/zenodo.20567719](https://doi.org/10.5281/zenodo.20567719), evaluating bioinformatics and computational-biology papers), and finally culminating in **MorphoEval** ([`fractastical/morpholib`](https://github.com/fractastical/morpholib) and our **Adversarial Epistemic Loop / AEL** protocol), which turns that auditing machinery inward to autonomously build out, audit, and formally prove the math of biology in executable code and Lean 4.
+In my case, I was fortunately aided by Mike Levin himself, who gave me a list of articles to read in which order to get started on learning the space, to which I added a massive pile of his papers, which I synthesized with a knowledge graph ([Project Delphi](https://github.com/fractastical/project-delphi-ontology)). That knowledge graph then evolved into a three-stage series of open-source verification engines—starting with **NeuroEval** ([doi:10.5281/zenodo.20622878](https://doi.org/10.5281/zenodo.20622878), evaluating data transparency and neural-coding rigor in computational and biohybrid/wetware neuroscience papers), followed by **BioEval** ([doi:10.5281/zenodo.20567719](https://doi.org/10.5281/zenodo.20567719), evaluating bioinformatics and computational-biology papers), and finally culminating in **MorphoEval** ([doi:10.5281/zenodo.23680705](https://doi.org/10.5281/zenodo.23680705), [`fractastical/morpholib`](https://github.com/fractastical/morpholib), and our **Adversarial Epistemic Loop / AEL** protocol), which turns that auditing machinery inward to autonomously build out, audit, and formally prove the math of biology in executable code and Lean 4. The latest stage is **Delphi** itself ([delphibio.org](https://delphibio.org), [`fractastical/delphi`](https://github.com/fractastical/delphi)), built with a team of five at Science Claw Hack @ MIT (Oct 30–Nov 1, 2026): 1,777 source-cited claims from Aristotle to Levin, 92 of them theorems (73 Lean-checked), 542 tested against public data with pre-registered pass/fail criteria, every record a signed nanopublication that any MCP client can query. It took second overall and first in the alphaXiv and Nebius tracks.
+
+[![](https://delphibio.org/blog/delphi-in-sixty-seconds-card.png)](https://delphibio.org/blog/delphi-in-sixty-seconds.mp4)
+
+*Plate III — Lean for Bio: the Delphi film (100 s; click to play). The claim graph growing from Aristotle to Levin, the assessor swarms, the Lean proofs, the data tests, the laws the data agree with, publication, the outside agents, and the weekend's commits hour by hour. Every count on screen is read from a repo file at that hour by `scripts/site/build-graph-video.py`; the weekend itself is in the [project report](https://delphibio.org/blog/science-claw-2026.html).*
 
 ![](https://fractastical.github.io/morpholib/media/all_forms_montage.gif)
-*Fifteen flagship animated models of biological form (1917–2025) in our open-source [Infinite Morphospace](https://github.com/fractastical/morpholib) repository—audited by independent AI referees from competing labs and formally checked in the Lean 4 theorem prover.*
+
+*Plate IV — Fifteen flagship animated models of biological form (1917–2025) in our open-source [MorphoEval](https://github.com/fractastical/morpholib) repository, each audited by independent AI referees from competing labs and formally checked in the Lean 4 theorem prover. Provenance: [`docs/media/make_montage.py`](https://github.com/fractastical/morpholib/blob/main/docs/media/make_montage.py).*
 
 Although none of this required "Platonism," it was partially the concept of validating the concepts of what mathematicians called the **"morphospace"** and Levin calls the **"latent Platonic space"** that motivated the creation of these tools—and which now allows agents to explore math more directly and which has already led to notable discoveries.
 
-As someone who needs sizable amounts of funding to conduct his experiments (his lab has produced 400+ papers in the last decade), he is probably more sensitive than I am to the language that makes funders want to fund. However, his more recent set of terms choosing "Platonism" and other terms like "ingress" (*Ingressing Minds*, [2025](https://doi.org/10.31234/osf.io/5g2xj_v3)/[2026](https://doi.org/10.3390/philosophies11050161)) have set off a large set of internet debates, in which I inadvertently got involved when *Sweet Rationalism* (Nathan Sweet) suggested my method for validation may be more comprehensive and better than Levin’s own.
+As someone who needs sizable amounts of funding to conduct his experiments (his lab has produced 400+ papers in the last decade), he is probably more sensitive than I am to the language that makes funders want to fund. However, his more recent set of terms choosing "Platonism" and other terms like "ingress" (*Ingressing Minds*, [2025](https://doi.org/10.31234/osf.io/5g2xj_v3)/[2026](https://doi.org/10.3390/philosophies11050161)) have set off a large set of internet debates (the two sides are laid out, with the open questions each still owes, on bioelectricity.net’s [Great Debates](https://bioelectricity.net/great-debates) page), in which I inadvertently got involved when *Sweet Rationalism* (Nathan Sweet) suggested my method for validation may be more comprehensive and better than Levin’s own.
 
 Although we have briefly discussed these things in person, I do not know, at present, all of the experiments Mike Levin has underway. What I know is mostly limited to the core arguments he brings in his most recent paper, which are the **D’Arcy Thompson-inspired version of the morphology of math** ([Thompson 1917](https://doi.org/10.5962/bhl.title.11332)) and the **[Zhang, Goldstein & Levin (2025)](https://doi.org/10.1177/10597123241269740) paper regarding emergent phenomena in Bubble, Insertion, and Selection sort** (alongside the classic bioelectric flatworm, Picasso tadpole, and Xenobot models).
 
 In both cases, **precise language of what is claimed is what is most important**, and, in a sea of definitions, a place where an internet philosopher who can code might be most useful. It also turns out that despite a bunch of folks who are "pro" and "against" on the internet, **all of the main claims can be validated with tools that everyone has at their fingertips.**
 
-Let’s go through them claim by claim.
-
----
-
-## Claim 1: "No Error-Handling Code Was Added" to Self-Sorting Cells ([Zhang, Goldstein & Levin 2025](https://doi.org/10.1177/10597123241269740))
-
-In [*Adaptive Behavior* (2025)](https://doi.org/10.1177/10597123241269740) and Section 6.3 / Figure 14 of [*Ingressing Minds* (2026)](https://doi.org/10.3390/philosophies11050161), Taining Zhang, Adam Goldstein, and Michael Levin asked a clever question: **What if there is no central controller (no "gym teacher") sorting an array of numbers from 1 to 100?** What if each number is an independent thread (an autonomous "cell") that can only look at its neighbors and decide on its own whether to swap places?
-
-They gave the cells three classical computer-science "personalities" (**algotypes**):
-* **Bubble cells (B):** Look at your immediate left or right neighbor (`p − 1` or `p + 1`). If you are out of order, swap!
-* **Insertion cells (I):** Look at everyone to your left (`0 .. p − 1`). Only step left with `p − 1` if everyone to your left is already in order.
-* **Selection cells (S):** Keep a mental finger (`ideal_position` / `target` pointer) starting at slot `0`, then `1`, then `2`... If your number is smaller than whoever is standing at `target`, jump straight there and swap with them; otherwise move your finger one slot to the right.
-
-![](https://fractastical.github.io/morpholib/media/zhang_13yo_fig1_three_personalities.png)
-
-Before testing the philosophical claims, we ported their exact GitHub commit ([`Zhangtaining/cell_research@1fd2bd59`](https://github.com/Zhangtaining/cell_research/tree/1fd2bd5921c1f6b423a71f691d5189106a8a1020)) into **four independent languages: Python 3, C++20, JavaScript (Node.js), and the Lean 4 theorem prover** (`0 sorry`).
-
-First, credit where credit is due: **all 14 of the empirical curves published in Zhang, Goldstein & Levin (2025) replicate 100%**. Across Python, C++20, and Node.js, our ports produce bit-for-bit identical space-time kymographs (`0` cell mismatches) and match every benchmark in the paper.
-
-![](https://fractastical.github.io/morpholib/media/zhang_sorting_morphogenesis.gif)
-
-![](https://fractastical.github.io/morpholib/media/zhang_multilang_port_verification.png)
-
-Now look at the first philosophical claim Levin makes about this code in *Ingressing Minds* (Section 6.3, p. 22) when a few cells in the array are "frozen" (broken and refusing to move):
-
-> *"The critical thing is that we did not add steps to the algorithm to check whether a swap had been completed, or to do specific things if an operation failed."*
-
-**What We Found When We Audited the Code (Panel A below):**
-For **Bubble** sort, Levin’s statement is 100% true. The only place `BubbleSortCell.py` ever looks at a `FREEZE` flag is to *decline* a swap with a frozen neighbor—a swap precondition that treats the frozen cell as a brick wall, not a recovery routine; no internal state is touched and the sorting rule is unchanged. **Selection** sort ([`SelectionSortCell.py`](https://github.com/Zhangtaining/cell_research/blob/1fd2bd5921c1f6b423a71f691d5189106a8a1020/modules/multithread/SelectionSortCell.py)) is the same story: its ordinary non-local `target` pointer swap tunnels straight across an immovable defect with no frozen-cell logic at all. (The one Selection-specific `FREEZE` branch in the upstream repository lives in the compiled bytecode `SelectionSortCell.cpython-310.pyc` used for the paper’s *passive, movable* defects in Fig. 5B, where the pointer is nudged past a frozen target to avoid livelock—again a skip, not a rewrite of the algorithm; we document it as rule `[U4]` in our port guide.)
-**However, for Insertion Sort (`InsertionSortCell.py`, lines 68–82), the claim is false—because here the authors did not merely skip the frozen cell; they rewrote the algorithm’s own correctness test to see through it.**
-
-Remember Insertion sort’s rule: a cell at position `p` refuses to move left unless the entire prefix `0 .. p − 1` to its left is already sorted. If a single frozen cell wearing `#99` sits near the front of the line, then the left side is *never* sorted—which under pure Insertion sort would permanently paralyze every cell to the right of `#99`!
-
-When we opened [`InsertionSortCell.py` (lines 68–82)](https://github.com/Zhangtaining/cell_research/blob/1fd2bd5921c1f6b423a71f691d5189106a8a1020/modules/multithread/InsertionSortCell.py#L68-L82), we found **3 lines of explicit broken-cell exception handling**:
-
-```python
-for i in range(self.current_position[0]):
-    if self.cells[i].status == CellStatus.FREEZE:
-        prev = -1   # <-- Explicit exception handler resetting prefix check at frozen cells!
-        continue
-    if self.cells[i].value < prev:
-        return False
-    prev = self.cells[i].value
-```
-
-When we deleted those 3 lines of code (`without_explicit_freeze_reset` in **Panel A** below, red bars, and formally proved in Lean 4 Theorem 2), Insertion sort’s final error (`n = 60`, `f = 3` frozen cells) **jumped 4.5× higher (from 4.90 to 21.95)** as all cells behind the defect froze!
-
-![](https://fractastical.github.io/morpholib/media/zhang_levin_platonic_claims_audit.png)
-
-* **Verdict on Claim 1 — Partially Refuted:** Bubble and Selection sort genuinely get defect tolerance for free from decentralized execution (their only frozen-cell logic is a swap-precondition skip that treats a frozen cell as a wall); Insertion sort relies on an explicit 3-line `if status == FREEZE: prev = -1` exception handler that rewrites its own prefix-sorted test to see through that wall.
-
----
-
-## Claim 2: "Delayed Gratification" — Do Sorting Cells Really Go Against Their Gradient to Navigate Around a Barrier?
-
-In *Ingressing Minds* (Section 6.3 & Figure 14A–B), Levin compares the sorting array to an agent walking around a physical wall—contrasting it with two magnets separated by a piece of wood, which get stuck forever because they cannot temporarily move *against* their magnetic gradient. Because the paper’s sorting error score (`monotonicity_error`, `E(v)`) temporarily ticks *upward* before dropping lower (`Delayed Gratification > 0`), Levin argues the algorithm is *"temporarily going against its gradient to recoup greater gains later."*
-
-**What We Found (`Panel B` above & Illustrated Guide 2 below):**
-We tested two mathematical properties of the trajectories—and proved both in Lean 4:
-
-1. **Zero Bubble or Insertion cells ever cross an immovable barrier:**
-   In Bubble and Insertion sort, cells only swap with immediate neighbors `(p, p + 1)`. If an immovable frozen cell sits at index `d = 20`, **literally 0 cells from the left of spot 20 ever cross to the right of spot 20** (Lean 4 Theorem 1: `bubble_adjacent_swap_preserves_left_partition`). The left half and right half sort in complete isolation! Selection sort only bypasses spot 20 because its `target` pointer is given global array access to teleport across indices.
-
-2. **Why did the error score go up? The Optical Illusion of the Two Rulers:**
-   Why did `monotonicity_error` `E(v)` tick uphill 47 times in **Panel B** (blue jagged curve)? Because there are **two different mathematical rulers** for measuring how scrambled a line of numbers is:
-   * **Ruler 1 — Total Out-of-Order Pairs `I(v)` (The True Energy):** Count *every* pair of cells `(i < j)` in the whole line where a larger number sits to the left of a smaller number (the Coxeter word length on the symmetric group `S_n`).
-   * **Ruler 2 — Neighbor Drops `E(v)` (The Paper’s Ruler):** Only count how many times a cell is larger than the cell *immediately* to its right (`v_i > v_{i+1}`).
-
-   Look at a tiny line of 4 cells: **`[2, 4, 1, 3]`**:
-   * Under **Ruler 1 (`I`)**, there are **3** out-of-order pairs: `(2,1)`, `(4,1)`, and `(4,3)`.
-   * Under **Ruler 2 (`E`)**, there is only **1** neighbor drop: `4 → 1` (since `2 < 4` and `1 < 3`).
-   * Now let Bubble sort swap the out-of-order pair `4 ↔ 1` to get **`[2, 1, 4, 3]`**.
-   * What happened to **Ruler 1 (`I`)**? It went **strictly DOWN from `3 → 2`**! In fact, **every single Bubble swap in history reduces `I(v)` by exactly `1` (`ΔI = −1`, `0` uphill steps out of 192 swaps—the yellow straight line in Panel B!)**.
-   * Why did **Ruler 2 (`E`)** go **UP from `1 → 2`** (`2 > 1` and `4 > 3`)? Because swapping `4` and `1` broke one tall cliff (`4 → 1`) into two smaller steps (`2 → 1` and `4 → 3`)! This happens **even when there are zero frozen cells (`f = 0`, where `DG = 0.38–0.43`)**.
-
-![](https://fractastical.github.io/morpholib/media/zhang_13yo_fig2_two_rulers_illusion.png)
-
-* **Verdict on Claim 2 — Optical Illusion of the Ruler:** The sorting cells never move against their dynamical gradient. They slide 100% deterministically downhill on their true Lyapunov energy `I(v)` (`ΔI = −1` on every single swap), while the observer watches them through a blurry neighbor-drop ruler `E(v)`.¹
-
----
-
-## Claim 3: "Spontaneous Algotype Clustering" — Why Levin’s Phenomenon Is 100% Real, and the CS Blog Critics Were Wrong Twice!
-
-Here is where our audit caught the *critics* making an even bigger mathematical mistake than the dualists.
-
-In Figure 8 of Zhang et al. (and Figure 14C of *Ingressing Minds*), when you mix 50% **Bubble (`B`)** cells and 50% **Selection (`S`)** cells in the same line, cells of the same personality spontaneously group together (`aggregation > 0.50`)—even though **zero lines of code ever inspect `neighbor.algotype`**.
-
-Levin called this an unprogrammed "side-quest" from Platonic space.
-In response, computer-science bloggers (including Nathan Sweet at *Sweet Rationalism*) argued that this clustering is nothing more than a 1960s textbook property: **duplicate-key sort stability** (when multiple cells have the same number, Bubble sort is "stable" and doesn't swap equal keys).
-
-When we ran the ablations (**Panel C** of the audit figure above, and Illustrated Guide 3 below), we proved that **the CS blog critics were wrong twice—both on unique numbers AND on duplicate numbers:**
-
-1. **Critics’ Mistake #1 — Paper Fig. 8A Has ZERO Duplicate Numbers!**
-   In Paper Figure 8A (**Panel C**, blue curve), every single cell has a **strictly unique number from `1` to `n` (`0` duplicate keys)**. When there are no duplicate keys, "duplicate-key sort stability" is mathematically impossible! Yet `B + S` aggregation still rises from `0.50` to a peak of **`0.624–0.640`** mid-sort before relaxing to `0.501` at the end.
-   **Why?** Because **Selection cells move like sports cars and Bubble cells move like pedestrians!** Selection cells use their `target` pointer to jump small numbers straight to the left front of the line (`0, 1, 2...`) in `O(1)` swaps, shoving a contiguous wave of slow `O(n²)` Bubble cells into the right half of the line.
-
-2. **Critics’ Mistake #2 — The `C = 2` "Anti-Clustering" Smoking Gun on Duplicate Numbers!**
-   What about Paper Figure 8D, where there *are* duplicate numbers (`C` copies of each key)? Think carefully: at the start of the sort, the `B` and `S` labels are shuffled randomly (`aggregation = 0.50`). If textbook "sort stability" simply preserved the initial relative order of equal keys, then **final aggregation at the end of the sort would stay at `0.50` for every `C`!**
-   Instead, when we swept `C = 1, 2, 4, 8, 10, 20, 40` copies per number, we discovered that because Selection cells scan `target` from `0` upward, **66.2% (at `C = 2`) to 75.0% (at `C = 10`) of Selection cells pack into the LEFT half of every number’s group (`S...S | B...B`)**:
-   * When there are **`C = 2` copies of each number** (one `S` and one `B` wearing `#1`, one `S` and one `B` wearing `#2`, etc.), putting `S` on the left and `B` on the right of each pair creates the alternating pattern **`S B | S B | S B`**—which drives clustering **below random to `0.477` (Anti-Clustering!)**!
-   * When there are **`C = 10` copies of each number** (`Fig. 8D`), putting 5 `S` cells on the left and 5 `B` cells on the right of each group creates **`SSSSS | BBBBB`**, driving permanent clustering **up to `0.659`**! (And when we change `<` to `<=` on ties in **Panel C**, red dashed curve, the permanent clustering collapses back to `0.497–0.514`.)
-
-![](https://fractastical.github.io/morpholib/media/zhang_13yo_fig3_sports_cars_vs_pedestrians.png)
-
-And here are the real computer-experiment curves side by side—showing **(A)** the smooth downhill slide on Ruler 1 (`I(v)`, `0` uphill steps) vs. the 47 illusory bumps on Ruler 2 (`E(v)`), and **(B)** how the `C = 2` Anti-Clustering dip (`0.477 < 0.50`) and `C = 10` Clique peak (`0.659`) refute the blog critics’ flat `0.50` prediction:
-
-![](https://fractastical.github.io/morpholib/media/zhang_13yo_fig4_real_data_curves.png)
-
-* **Verdict on Claim 3 — Empirically Vindicated (via Kinematic Phase Separation):** The clustering is 100% real and refutes the blog critics’ "textbook stability" dismissal—yet it is also 100% produced and predicted by the kinematic transport equations of the two algorithms without needing non-physical ghosts. ([Try the live `C = 2` Sorting Morphogenesis Simulator on bioelectricity.net here!](https://bioelectricity.net/sorting-morphogenesis))
-
----
-
-## Claim 4: D’Arcy Thompson’s (1917) Generative Morphospace Across 10,303 Fish Species
-
-The second major pillar Levin invokes for his latent space is **D’Arcy Wentworth Thompson’s (1917)** famous coordinate transformations from *On Growth and Form*: the idea that nature doesn't redesign every bone from scratch when evolving a parrotfish (*Scarus*) into an angelfish (*Pomacanthus*) or a pufferfish (*Diodon*) into an ocean sunfish (*Mola mola*), but instead turns a few global **"growth knobs"** that bend the 2D coordinate grid of the body like a sheet of rubber.
-
-![](https://fractastical.github.io/morpholib/guide_step1_1917_rubber_sheet.png)
-
-For the last 30 years, evolutionary biologists tried to test Thompson’s idea on computer landmark databases—and concluded he was mostly wrong across fish families. **Why did their tests fail?**
-
-When we audited the morphometrics literature across **three independent datasets scaling up to 10,303 fish species (`440` families)**, we uncovered two methodological traps that had hidden Thompson’s law for a century:
-
-1. **Trap #1 — The "17-Dot Head Seesaw":**
-   Standard fish databases ([Claverie & Wainwright 2014](https://doi.org/10.1111/evo.12502); [Friedman et al. 2019](https://doi.org/10.1093/sysbio/syz029)) don't trace the whole outline of the fish—they measure **17 dots on the skeleton**, with **9 of those 17 dots (53%) packed inside the head and jaws** and **0% on the outer dorsal, anal, or tail fins**. Because jaws evolve rapidly for diet while the body evolves for swimming, fitting a global rubber sheet to those 17 dots lets the 9 head dots pull the back of the sheet the wrong way—making the warped fish silhouette look **worse than no warp at all** (`40.6% → 56.2%` mismatch on *Scarus* → *Pomacanthus*)!
-
-![](https://fractastical.github.io/morpholib/guide_step2_17dot_head_trap.png)
-
-![](https://fractastical.github.io/morpholib/media/thompson_1917_silhouette_paradox_summary.gif)
-
-2. **Trap #2 — The "Plain Box Stretch" Illusion vs. True Thompson Curves:**
-   When we switch to a **600-point whole-body outline**, if you only ask a computer to overlap two *unlabeled black shadows*, a trivial **Rectangular Box Stretch** (stretching height vs. width with *zero* slant or curve) can "cheat" by stretching the middle of a pufferfish’s belly straight up to cover the back fins of an ocean sunfish!
-   **Where Thompson’s curved and slanted grids (`70° Oblique Shear`, `Coaxial Circles`, `Bilinear Fan`, `Radial Flare`) are genuinely required** is when you force the grid to match **true homologous perimeter anatomy** (dorsal fin base & tip to dorsal fin base & tip, anal fin to anal fin, without the 9 internal skull dots): there, a plain Box Stretch fails (**`2.4%`** match on the swept-back fin tips of the angelfish), and **Thompson’s curved grids jump by `+13%` to `+48%` (`15×` better)** while staying **100% smooth and un-folded on `6 out of 6` pairs** (whereas standard 600-point Thin-Plate Splines ([Bookstein 1989](https://doi.org/10.1109/34.24792)) fold and crumple the interior tissue grid on `6 out of 6` pairs!).
-
-![](https://fractastical.github.io/morpholib/guide_step3_box_cheat_vs_curves.png)
-
-3. **The Two Rules of Fish Evolution Across 3 Datasets (`791 → 6,614 → 10,303` Species) & Our Blind Test on 5,180 Unseen Species:**
-   Testing three independent datasets (**791**, **6,614**, and **10,303** species across `440` families and `987` genera) revealed that fish evolution obeys two orthogonal rules at once:
-   * **Rule 1 — The Outside Rubber Sheet ("How the Fish Swims"):** Over 150 million years across 440 families (on the time-calibrated tree of [Rabosky et al. 2018](https://doi.org/10.1038/s41586-018-0273-1)), the *direction* of whole-body evolution follows D’Arcy Thompson’s 8-knob coordinate highway with **93.5% to 98.9% alignment (`θ = 20.7° to 8.5°`)**, capturing **92.0% of the theoretical 8-parameter Relative-Warp variance ceiling** without a single data-trained eigenvector! (A word on knobs vs. modes, since both numbers appear in this article: each of Thompson’s 1917 grid families has `8` parameters—`6` affine plus `2` curvilinear. Once Procrustes alignment removes position, rotation, and size, the pooled non-trivial terms across the families form the `7` linearized **"Thompson-7"** modes we use for prediction below: aspect, shear, two fans, radial bow, and two coaxial Möbius terms.)
-     **And a word on what these numbers do *not* say:** alignment measures whether the direction of between-species shape change lies along Thompson’s modes; it does not say how much of the total variance those modes soak up. On the standard 17-landmark reef-fish data, the `7` a-priori Thompson modes capture **`57.9%`** of between-species shape variance, versus **`86.2%`** for a data-trained 7-component PCA. That `57.9%` is the benchmark we hold every other clade against in our portfolio (and which hornet-wing perimeters will later beat). Thompson’s grids describe the *highway* evolution travels along; they do not pretend to be the whole map.
-   * **Rule 2 — Inside the Head ("What the Fish Eats"):** When closely related sister species live side-by-side on the same coral reef, they share the same swimming body, so they **rapidly rewire their jaws, snouts, and eyes 3.3× to 9.3× faster than their bodies** (`85.4%` of within-genus variance; diet shift `p = 9.2 × 10⁻⁵`) to partition food without fighting over the same meal!
-   * **Pre-Registered Blind Test on 5,180 Unseen Species (`4/7` Hypotheses Supported):** Before opening the Smithsonian’s **FishShapes v1** dataset (`5,180` never-seen species), we locked our predictions by SHA-256 hash on the production Nanopub network ([`RAilN4nbRHp3e3ZHqlbIgeL3UGPRV_KtQHtJD3gzVVdqw`](https://registry.knowledgepixels.com/np/RAilN4nbRHp3e3ZHqlbIgeL3UGPRV_KtQHtJD3gzVVdqw)). Within the **~50 Myr modularity horizon** (`n = 1,122` species), both the full 30-D tree predictor (`p = 10⁻¹⁰³`, `r = 0.87`, `92.6%` empirical coverage) and the 7-mode Thompson predictor (`p = 10⁻⁸¹`) crushed the clade consensus—while our pre-registered scoring protocol also honestly caught that on a single noisy calliper ratio, full Gaussian tree conditioning only beats copying the single nearest relative by 4% (`p = 0.061`).
-
-![](https://fractastical.github.io/morpholib/guide_step4_outside_vs_inside_rules.png)
-
-* **Verdict on Claim 4 — Strongly Confirmed:** Across 10,303 species, Thompson’s and Levin’s core thesis—that macroevolution navigates a low-dimensional geometric coordinate manifold while local modules handle rapid ecological competition—is a quantitative law of vertebrate evolution. ([Explore the interactive 10,303-Species Fish Morphospace on bioelectricity.net here!](https://bioelectricity.net/bioelectric-morphospace))
-
----
-
-## Claims 5, 6 & 7: How Bioelectric & Biomechanical Attractors Physically Implement the Latent Space (Planarians, Picasso Tadpoles & Xenobots)
-
-How does a living tissue actually store a target coordinate in that morphospace? In our repository, we also built and audited mathematical reproductions of Levin’s three most famous wet-lab discoveries—testing whether standard biophysical PDEs and ODEs genuinely produce those attractor behaviors (while explicitly documenting reduced-model parameter assumptions in our MorphoEval Pass 3 referee audits):
-
-* **Claim 5 — Bioelectric Target Memory in Two-Headed Planarians ([Oviedo et al. 2010](https://doi.org/10.1016/j.ydbio.2009.12.012); [Beane et al. 2011](https://doi.org/10.1016/j.chembiol.2010.11.012); [Pietak & Levin 2016](https://doi.org/10.3389/fphys.2016.00178)/[2017](https://doi.org/10.1098/rsif.2017.0425); [Cervera, Levin & Mafé 2021](https://doi.org/10.1016/j.biosystems.2021.104511)):**
-  In the lab, blocking gap junctions in a cut planarian flatworm with 8-octanol for 48 hours causes it to regenerate **two heads (`2H0T`)**—and weeks after the drug is washed out, cutting the worm again in plain water *still* produces two heads with zero DNA change. (For scale across the literature, our exact McKay IR isomorphism census of [Lobo, Malone & Levin’s 2013 `PlanformDB 2.5.0`](https://doi.org/10.1093/bioinformatics/btt088) shows that its `369` recorded planarian morphologies reduce to `38` exact region-graph topologies across `12` unlabelled tree skeletons, with wild-type `1H1T` first at `n = 253` and bipolar double-head `2H0T` second at `n = 16`—the same graph-encoded formalism Lobo and colleagues later extended to limb regeneration in [Limbform (2014)](https://doi.org/10.1093/bioinformatics/btu582) and to a [linear-encoding model of target-morphology variability (2014)](https://doi.org/10.1098/rsif.2013.0918).)
-  In our 1D PDE and 405-cell 2D Voronoi Scharfetter–Gummel electrodiffusion models (`2017-pietak-bioelectric-morphogen-coupling.py`), positively charged serotonin (`5-HT⁺`, `z = +1`) electrophoreses down the resting voltage gradient (`−20 mV` head to `−60 mV` tail) by the exact Boltzmann ratio `c_post / c_ant = exp(40 mV / 25.85 mV) ≈ 4.70`. Crucially, **gap-junction conductance `G_gj` cancels out of the zero-flux equilibrium ratio!** Once transient 8-octanol blockade lets the posterior pole flip to `−20 mV` (`2H0T`), washing out the drug leaves `ΔV = 0` and permanently locks the worm in the two-headed attractor basin.
-
-![](https://fractastical.github.io/morpholib/media/cervera_levin_mafe_gj_block.gif)
-
-* **Claim 6 — Closed-Loop "Picasso Tadpole" Homeostasis ([Vandenberg, Adams & Levin 2012](https://doi.org/10.1016/j.ydbio.2012.01.027); [Pezzulo & Levin 2015](https://doi.org/10.1098/rsif.2015.0772); [Pinet et al. 2019](https://doi.org/10.1242/dev.175950)):**
-  When tadpole eyes, jaws, and nostrils are surgically scrambled at stage 45, they don't follow hardwired displacement vectors (which would preserve 100% of the scrambling error at metamorphosis). Instead, they move along novel paths with speed proportional to defect distance until they reach the normal froglet face—unless MMP tissue remodeling is blocked by GM6001. Our Lean-verified Procrustes / Variational Free Energy Lyapunov models (`dV/dt ≤ 0`, coupled in Pass 3 to 3D Maxwell viscoelastic cartilage rods and MMP-2/MMP-9 zymography fields) reproduce both normal convergence and GM6001 arrest.
-
-![](https://fractastical.github.io/morpholib/media/xenopus_craniofacial_remap_growth.gif)
-
-* **Claim 7 — Xenobots & Kinematic Self-Replication Without Genetic Change ([Kriegman, Blackiston, Levin & Bongard 2020](https://doi.org/10.1073/pnas.1910837117), [2021 *PNAS*](https://doi.org/10.1073/pnas.2112672118)):**
-  In Kriegman et al. (2021), when ciliated *Xenopus laevis* embryonic skin (animal-cap ectoderm) spheroids are sculpted into a **C-shaped ("Pac-Man") mouth concavity (`θ ≈ 105°`)** and placed in a Petri dish filled with dissociated individual frog embryonic cells, their beating cilia generate fluid currents that sweep the dissociated cells into the mouth pocket; once enough cells are pushed together, natural cadherin cell-cell adhesion bonds them into a new, motile daughter Xenobot (`3–4` generations of kinematic self-replication). Both our 8×8×7 3D voxel mass-spring lattice and our 2D regularized-Stokeslet squirmer fluid simulation confirm that `θ ≈ 105°` is a genuine hydrodynamic capture optimum (`3–4` daughter generations vs. `1` for a sphere and `0` for a flat slab).
-
-![](https://fractastical.github.io/morpholib/media/xenobot_evolutionary_biobots_growth.gif)
-
----
-
-## The Final Scoreboard: How Right Is Michael Levin? (`81.6%` Across the 3 Equal `33.3%` Pillars — `93.4%` on Structural Math, `0.0%` on Non-Physical Ghosts)
-
-![](https://fractastical.github.io/morpholib/media/substack_final_scoreboard.png)
-
-To give a concrete, transparent answer to *"How right is Michael Levin?"*, we gave **each of the three major pillars of this article an equal `33.3%` (`1/3`) weight**:
-1. **Pillar I — Self-Sorting Cells (Claims 1–3 = `33.3%` total, `11.1%` each)**
-2. **Pillar II — D’Arcy Thompson’s Generative Morphospace (Claim 4 = `33.3%` total)**
-3. **Pillar III — Bioelectric & Xenobot Attractors (Claims 5–7 = `33.3%` total, `11.1%` each)**
-
-Here is how the subcomponents add up across the three equal pillars:
-
-* **Pillar I — Self-Sorting Cells (`18.5 / 33.3 pts = 55.6%` of Pillar I):**
-  * **Claim 1 — Sorting Error Tolerance (`7.4 / 11.1 pts = 66.7%` — PARTIALLY REFUTED):** Bubble (`+3.7 / 3.7`) and Selection (`+3.7 / 3.7`) need no frozen-cell logic beyond a swap-precondition skip, while Insertion (`+0.0 / 3.7`) relies on an explicit 3-line `if status == FREEZE: prev = -1` reset in `InsertionSortCell.py` (L74–76; `4.5×` error jump when removed).
-  * **Claim 2 — Sorting "Delayed Gratification" (`1.5 / 11.1 pts = 13.5%` — OPTICAL ILLUSION OF THE RULER):** The neighbor-drop curve `E(v)` genuinely ticks up 47 times (`+1.5 / 3.7`), but Lean 4 Theorem 1 proves `0` Bubble/Insertion cells cross an immovable barrier (`+0.0 / 3.7`) and every Bubble swap goes strictly downhill (`ΔI = −1`, `0/192` uphill steps, `+0.0 / 3.7` dualism) on the true pairwise inversion energy `I(v)`.¹
-  * **Claim 3 — Secret Clique Clustering (`9.6 / 11.1 pts = 86.5%` — EMPIRICALLY VINDICATED & CRITICS REFUTED TWICE):** Strictly unique keys (`Fig. 8A`, `+4.8 / 4.8`) and `C = 2` anti-clustering vs. `C = 10` cliques (`+4.8 / 4.8`) are 100% real and refute the CS blog critics twice, while `+0.0 / 1.5` pts go to non-physical dualism.
-* **Pillar II — D’Arcy Thompson’s Generative Morphospace (`32.0 / 33.3 pts = 96.0%` of Pillar II):**
-  * **Claim 4 — 600-Point Homologous Fin Outlines (`+16.7 / 16.7 pts`) & 10,303-Species / 5,180-Blind-Test Validation (`+15.3 / 16.6 pts` — STRONGLY CONFIRMED):** Standard 17-dot tests failed only because 53% of the dots sit in the skull. On homologous perimeter outlines, Thompson’s 8-knob grids win by **`+13%` to `+48%`** (`100%` un-folded) and align with **150 Myr of fish body-plan evolution at `93.5%–98.9%`** across 440 families (`p = 10⁻¹⁰³` in our pre-registered blind test; losing `1.3 pts` only because `H2` nearest-neighbor margin was `4%`, `p = 0.061`).
-* **Pillar III — Bioelectric Memory (`2H0T` Planarians), Picasso Tadpoles & Xenobots (`31.1 / 33.3 pts = 93.3%` of Pillar III):**
-  * **Claim 5 — Planarian `2H0T` Bioelectric Memory (`10.55 / 11.1 pts = 95%`):** Zero-flux Nernst–Planck Boltzmann ratio (`exp(40/25.85) ≈ 4.70`) cancels `G_gj` at equilibrium and locks permanent two-headed memory after washout (`−0.55` for reduced-model assumptions—1D/2D electrodiffusion with fixed channel kinetics—not a dualist deduction).
-  * **Claim 6 — Picasso Tadpole Homeostasis (`10.55 / 11.1 pts = 95%`):** Closed-loop Procrustes / Variational Free Energy Lyapunov descent (`dV/dt ≤ 0`) coupled to 3D Maxwell cartilage & MMP zymography reproduces face normalization and GM6001 arrest (`−0.55` for reduced-model assumptions in the viscoelastic/MMP coupling—not a dualist deduction).
-  * **Claim 7 — Xenobot Kinematic Self-Replication (`10.0 / 11.1 pts = 90%`):** 8×8×7 3D voxel lattice and 2D regularized-Stokeslet squirmer confirm `θ ≈ 105°` C-mouth concavity as the hydrodynamic capture optimum (`3–4` daughter generations; `−1.1` for the unmodeled 3D continuum fluid–tissue coupling—not a dualist deduction).
-* **How the 3 Equal Pillars Add Up (`18.5 + 32.0 + 31.1 = 81.6 / 100 pts = 81.6% Overall`):**
-  * **3-Pillar Composite Score: `(55.6% + 96.0% + 93.3%) / 3 = 81.6%`**
-  * **Where the missing `18.4` pts went:** `12.6` pts are deductions for *dualist or overstated* language—`3.7` in Claim 1 (Insertion), `7.4` in Claim 2 (barrier detour + against-the-gradient), and `1.5` in Claim 3 (ghost clustering). The remaining `5.8` pts are ordinary modelling shortfalls that have nothing to do with ghosts—`2.2` for the genuinely rising `E(v)` curve in Claim 2, `1.3` for the failed `H2` margin in Claim 4, and `2.2` of reduced-model biophysics across Claims 5–7 (to rounding)—so they count against the structural denominator, not the dualist one.
-  * **On Empirical Phenomena & Structural Mathematical Free Lunches (`D + P_structural`): `81.6 / 87.4 pts = 93.4% RIGHT!`**
-  * **On Strong Dualist "Non-Physical Ingression" (`P_dualist`): `0.0 / 12.6 pts = 0.0% NEEDED!`**
-
----
-
-## How We Built and Audited This: From NeuroEval → BioEval → MorphoEval (The Adversarial Epistemic Loop)
-
-How do you keep yourself honest when testing high-profile scientific claims—especially when both the "Platonic Space" camp and the "Nothing New Here" blog critics have strong priors?
-
-The methodology used in this article—which we call the **Adversarial Epistemic Loop (AEL)**—was developed across three successive evaluation systems in Project Delphi:
-1. **Stage 1 — NeuroEval ([Zenodo `10.5281/zenodo.20622878`](https://doi.org/10.5281/zenodo.20622878)):** Built first to evaluate data transparency, reproducibility, and neural-coding rigor in computational and biohybrid/wetware neuroscience papers.
-2. **Stage 2 — BioEval ([Zenodo `10.5281/zenodo.20567719`](https://doi.org/10.5281/zenodo.20567719)):** Extended that rubric to score bioinformatics and computational-biology papers for reproducibility, statistical calibration, and information-theoretic rigor.
-3. **Stage 3 — MorphoEval ([`fractastical/morpholib`](https://github.com/fractastical/morpholib)):** Turned the evaluation machinery from *grading published papers* to *generating, executing, and auditing the mathematical models themselves* in a 12-step closed loop—where every verbal claim is translated into multi-language executable code (**Python 3, C++20, Node.js**) and machine-checked **Lean 4** theorems (`0 sorry`), locked against **SHA-256** empirical datasets, audited blind by **three competing AI model families** (Google DeepMind’s `gemini-4-argon`, Anthropic’s `claude-fable-5-1` / `claude-sonnet-5-5`, and OpenAI’s `gpt-5.6-sol`), and pre-registered as cryptographically signed **Nanopublications** before unblinding held-out species.
-
-📄 **Want the full line-by-line code traceability tables, all 24 mathematical/biophysical equations, SHA-256 dataset locks, and the 12 terminal commands to re-run every experiment?** Read our standalone **[4-Page Illustrated Methods & Provenance PDF (`levin_latent_space_methods_and_provenance.pdf`)](https://fractastical.github.io/morpholib/levin_latent_space_methods_and_provenance.pdf)** (and **[Markdown Methods Guide (`docs/LEVIN_LATENT_SPACE_METHODS.md`)](https://github.com/fractastical/morpholib/blob/main/docs/LEVIN_LATENT_SPACE_METHODS.md)**).
-
-### Transparency Heatmap: How Much of This Article and Pipeline Was Written by AI vs. Human?
-
-Because our **Adversarial Epistemic Loop (AEL)** pairs human philosophical direction and domain skepticism with multi-agent code generation, formal theorem proving, and cross-provider refereeing, we also audited the exact **Human vs. AI-Agent contribution** across all 9 sections of this article and all 5 workflow tasks. (The percentages are the author’s section-by-section estimates; the summary figures below are unweighted means over the 9 sections, computed directly from the heatmap cells by `sorting-morphogenesis/render_13yo_illustrations.py` so the figure and the text cannot drift apart.)
-
-![](https://fractastical.github.io/morpholib/media/substack_ai_authorship_heatmap.png)
-*Methods Transparency Heatmap: Breakdown of AI-Agent vs. Human PI (Joel Dietz) contribution across all 9 sections of this article and 5 workflow dimensions (Hypothesis & Direction, Prose Drafting, Simulation & Math Code, Lean 4 & Statistical Proofs, and Blinded Referee Auditing).*
-
-* **Hypothesis & Philosophical Direction (`80.6%` Human / `19.4%` AI):** The personal narrative (Galvani, Joscha Bach at CIMC, the board member’s "woo science" question, Mike Levin’s reading list, the Foresight Institute metascience reflection), the visual skepticism that exposed the `9/17` skull-dot seesaw trap and the 4-parameter Box-Stretch illusion, the 3-pillar rubric and `D` vs. `D + P` split, were driven by the human author (**Joel Dietz**, `85%–95%` human in Sections 1, 5, 8, and 9).
-* **Article Prose (`41.3%` Direct Human Writing & Edits / `58.7%` AI-Assisted Technical Synthesis):** Section 1 (the personal introduction) is **`92%` verbatim human prose**, and Section 9 (the `D` vs. `D + P` synthesis, Foresight metascience reflection, the Thompson 1913 / Aristotle passage, and conclusion) is **`55%` human prose**, while the technical claim walkthroughs, scoreboard tally, and methods summaries (Sections 2–8) were `60%–78%` drafted by `gemini-4-argon` and `claude-fable-5-1` from the repository’s locked JSON/YAML audit files and edited by the human author.
-* **Multi-Language Simulation Code, Lean 4 Proofs & Referee Audits (`78.8%` AI Agents / `21.2%` Human):** The C++20, Node.js, and Python ports, the 2D/3D PDE and Stokeslet fluid solvers, the 25+ kernel-checked **Lean 4** theorems (`0 sorry`), and the blinded cross-provider referee audits (`claude-fable-5-1`, `claude-sonnet-5-5-high`, `gemini-4-argon`, and OpenAI’s `gpt-5.6-sol`) were `85%–96%` executed by the AI agents in the technical sections (2–6 and 8) under deterministic `--self-test` gates; the human share is higher in the intro, scoreboard, and conclusion.
-
----
-
-## Answering the "Decisive Comparison (`D` vs. `D + P`)", "Legacy P1 Category" & "Frozen Update Rule" Critique
-
-In a thoughtful follow-up to my earlier comments, Nathan Sweet (*Sweet Rationalism*) laid out what he calls the **decisive-comparison requirement** for any predictive morphospace language:
+The line-by-line walkthrough is [a separate page](https://fractastical.github.io/morpholib/substack_levin_latent_space_claims.html). If you only want the ending, skip ahead to **Conclusion**.
+
+---
+
+## The Score
+
+![](https://fractastical.github.io/morpholib/media/substack_seven_claims_aggregation.gif)
+
+*Plate V — One tile per claim. Sorting, the two rulers, clustering, Thompson’s rubber sheet, the two-headed planarian, the Picasso tadpole, and the xenobot, then the unweighted mean `77.6%`. Each line below opens the walkthrough. Provenance: [`docs/media/make_claim_aggregation.py`](https://github.com/fractastical/morpholib/blob/main/docs/media/make_claim_aggregation.py).*
+
+1. **[Claim 1, `66.7%`](https://fractastical.github.io/morpholib/substack_levin_latent_space_claims.html#claim-1). "No error-handling code was added."** *"The critical thing is that we did not add steps to the algorithm to check whether a swap had been completed, or to do specific things if an operation failed."* Bubble and Selection sort treat a frozen neighbor as a wall. Insertion sort rewrites its own test to see through that wall.
+2. **[Claim 2, `13.5%`](https://fractastical.github.io/morpholib/substack_levin_latent_space_claims.html#claim-2). "Delayed Gratification."** *"When encountering a barrier (a 'broken' cell), the algorithm temporarily de-sorts the array in order to move things around the defective point… unlike two magnets separated by a piece of wood, which will never go around it because they cannot temporarily go against their gradient."* Every swap goes downhill on the true energy. The paper’s neighbor ruler is what ticks up.¹
+3. **[Claim 3, `86.5%`](https://fractastical.github.io/morpholib/substack_levin_latent_space_claims.html#claim-3). "Spontaneous algotype clustering."** *"Groups of cells with identical algotypes tended to form and persist… even though the algorithm does not have any code in it to make decisions based on algotype of a cell or its neighbors."* Two speeds of motion separate the mixture. No ghost is required.
+4. **[Claim 4, `96.1%`](https://fractastical.github.io/morpholib/substack_levin_latent_space_claims.html#claim-4). Thompson’s morphospace.** Nature does not redesign every bone. A few growth knobs bend the coordinate grid of the body like a sheet of rubber, from parrotfish to angelfish, from puffer to sunfish. We tested that on three datasets — 791, then 6,614, then 10,303 species, across 440 families — on 600-point whole-body outlines rather than the usual 17 skull dots, and locked a [blind prediction](https://registry.knowledgepixels.com/np/RAilN4nbRHp3e3ZHqlbIgeL3UGPRV_KtQHtJD3gzVVdqw) on 5,180 species before opening them. The body swims along that sheet. The head eats. The one hypothesis that lost is still in the record.
+5. **[Claim 5, `95.0%`](https://fractastical.github.io/morpholib/substack_levin_latent_space_claims.html#claim-5). Two-headed planarians remember a voltage.** Block the gap junctions of a cut planarian for 48 hours and it grows two heads. Wash the drug out, wait weeks, cut it again in plain water, and it grows two heads again. Nothing in the DNA changed.
+6. **[Claim 6, `95.0%`](https://fractastical.github.io/morpholib/substack_levin_latent_space_claims.html#claim-6). The Picasso tadpole finds its face.** Scramble the eyes, jaws, and nostrils and they do not crawl back along the paths they were pushed. They take new paths, faster the further they are from home, until the froglet face is normal.
+7. **[Claim 7, `90.1%`](https://fractastical.github.io/morpholib/substack_levin_latent_space_claims.html#claim-7). Xenobots copy themselves with no genetic change.** A C-shaped mouth of about `105°` sweeps loose cells into a daughter, then another.
+
+The `−2.2` points on Claims 5–7 are gaps in our reduced equations, not a finding that the worm, the tadpole, or the xenobot failed.
+
+**`(66.7% + 13.5% + 86.5% + 96.1% + 95.0% + 95.0% + 90.1%) / 7 = 77.6%`.**
+
+That is the number in the title, and its scope is the paper. These seven are the claims *Ingressing Minds* makes: three about the sorting arrays in §6.3, one about Thompson’s coordinate morphospace, and three about the planarian, the tadpole, and the xenobot. Nothing else he has published is in the average.
+
+With the dualist deductions taken out of each denominator before averaging, structural math on the same seven claims reads **`(100% + 40.5% + 100% + 96.1% + 95.0% + 95.0% + 90.1%) / 7 = 88.1%`**. **`0.0%` of the dualist points were earned.**
+
+* **Why the title says `77.6++%`:** `77.6%` is the reading of the seven claims in *Ingressing Minds*, not a reading of Levin in general. The `++` says we expect that reading to improve (footnote ³). *On Dualism* explains why the loop stays open. It does not move these points.
+
+---
+
+## Rationalism: Short and Sweet
+
+Nathan Sweet (*Sweet Rationalism*) asked for a decisive comparison:
 
 > *"Joel Dietz points in the right direction by proposing a predictive morphospace language connecting bioelectric patterns to downstream gene-expression cascades and morphology. A knowledge graph, however, can encode `D` or `D + P`. Platonic ingress begins earning evidence only when at least one `P`-specific relation improves held-out prediction, intervention, compression, or comparative model performance beyond the best `D`-only account.*
 >
 > *The decisive experiment would make `P` itself measurable at the edges: specify its hard core and bridge principle before observation; define a `P`-linked variable, topology, relation, or invariant independently of the residual; identify outcomes that materially separate `D + P` from serious rivals; use negative controls and held-out embodiments; replicate the advantage independently. Most important, freeze the update rule. A failed bridge must cost `P` something. A repair must buy fresh risk."*
 
-And on the legacy `P1` metadata category in Project Delphi:
-
-> *"The publication category is not the experimental contrast. Project Delphi’s legacy `P1` example makes this distinction tangible. The record proposes perturbing or removing tissue and measuring anatomical restoration, while its metadata places the proposal in a Platonic-forms category. That label can organize the proposal’s intellectual setting. Recovery evidence first bears on the developmental capacity being tested; an independent source of forms earns further credit through a comparison that distinguishes its contribution from regulatory and organizational alternatives."*
-
-**We agree 100% with every single one of those criteria—and in fact, our repository’s entire architecture was built to enforce that exact `D` vs. `D + P` protocol, including the frozen update rule.**
-
-![](https://fractastical.github.io/morpholib/media/zhang_13yo_fig6_decisive_comparison_d_vs_dp.png)
-*Illustrated Guide 6: (Rule 1, Top) Testing `D` (the 6 lines of local code alone) against `D + P` (adding a "Platonic Ghost" sticker), showing `0.00` extra swaps predicted by `P`; and (Rule 2, Bottom) locking predictions in a digital safe before unblinding new species—and taking losses honestly when a bridge fails on hornet wings.*
-
-Let’s walk through Nathan’s criteria one by one against the actual experiments in this repository:
-
-1. **Why "The Publication Category (`P1`) Is Not the Experimental Contrast":**
-   In early Project Delphi nanopublication schemas (`P1`), placing a regeneration experiment inside a `Platonic-forms` collection was a **bibliographic filing-cabinet tag** (grouping claims inspired by Levin’s papers), *not* an evidentiary proof that anatomical recovery comes from outside physics. As our current [`README.md`](https://github.com/fractastical/morpholib) and 4-tier claim-boundary catalog (`sorting-morphogenesis/catalog.json`) explicitly enforce, **recovery evidence first credits the dynamical/regulatory mechanism `D` (`1H1T`/`2H0T` Nernst–Planck bistability or Procrustes Lyapunov descent `dV/dt ≤ 0`)**. Putting a "Platonic" metadata sticker on a differential equation never gives `P` a free ride on `D`’s empirical success.
-
-2. **Separating `P_structural` (Mathematical Invariants Defined Independently of the Residual) from `P_dualist` ("Non-Physical Side-Quests"):**
-   Nathan writes that `P` begins earning evidence only when we *"define a `P`-linked variable, topology, relation, or invariant independently of the residual"* that *"improves held-out prediction, intervention, compression, or comparative model performance beyond the best `D`-only account."*
-   Look at **Rule 1 (top row)** of Illustrated Guide 6 above. Once we make that definition precise, `P` splits into two completely different hypotheses with opposite experimental outcomes:
-   * **`D + P_dualist` (Strong Dualist Ingression — "Side-Quests Not Produced by the Medium"): FAILS (`0.00` Incremental Standing).**
-     In *Ingressing Minds* (Sec. 6.3), Levin claims that Delayed Gratification (`DG > 0`) and chimeric algotype clustering are *"side-quests—not predicted, explained, or produced by that medium."* When we test `D + P_dualist` against `D` alone (the 6 local cell rules `[U1]–[U6]` in `Zhangtaining/cell_research@1fd2bd59`), our bit-exact ports across Python 3, C++20, Node.js, and Lean 4 prove that **`P(D + P_dualist) − P(D) = 0` (`0` cell mismatches; published to the production Nanopub network as [`RAH6FDLaacsq28cjx2NVRYRANqOThwHaHwE58ByxqIWI8`](https://registry.knowledgepixels.com/np/RAH6FDLaacsq28cjx2NVRYRANqOThwHaHwE58ByxqIWI8) and [`RA_mMJXveSMNlxIjYu_oTtIY233QfJnNCf47yRLfInMuE`](https://registry.knowledgepixels.com/np/RA_mMJXveSMNlxIjYu_oTtIY233QfJnNCf47yRLfInMuE))**. Adding a non-physical ingression edge improves neither prediction, nor intervention, nor compression by a single bit.
-   * **`D + P_structural` (Structural Mathematical Invariants Defined Independently of the Residual): EARNS REAL COMPRESSION & OUT-OF-SAMPLE STANDING.**
-     What if, instead of a dualist ghost, the "latent pattern" is a **mathematical invariant or symmetry group defined strictly a priori (independently of any fitted residual)**?
-     * **In Sorting Arrays:** The Coxeter pairwise inversion potential `I(v)` on the symmetric group `S_n` is defined purely from the algebra of permutations (zero fitted parameters). Coarse-graining the 720 micro-states of `S_6` onto `I(v)` **compresses the state space `45×` (`720 → 16` states)** while **increasing Markov determinism from `86.9%` to `100.0%` (`ΔI = −1` on every single swap)**!
-     * **In Fish Morphospace:** D’Arcy Thompson’s 7 linearized tangent coordinate modes (`aspect`, `shear`, `2 fans`, `radial bow`, `2 coaxial Möbius modes`) are defined analytically from 1917 geometry—*completely independently of any species residuals*. Projecting 30-D landmark space onto those 7 analytical modes **compresses the evolutionary covariance matrix from `465` parameters to `28`**, lowers the matrix condition number **`10×` (`κ: 201 → 20`, or `>1,000×` on small clades)**, and when fitted solely to the 600-point outer silhouette, **predicts 14 held-out interior skeletal bones `2.4×` more accurately** than a 600-point Thin-Plate Spline (`0.045` vs. `0.107` Procrustes RMS, while keeping `100%` of the interior tissue grid un-folded vs. `0%` for TPS)!
-     * **Cautionary Counter-Example (When a Generative Grammar Does *Not* Earn `P_structural` Credit):** We also tested whether an unconstrained graph-rewrite grammar on [`PlanformDB 2.5.0`](https://doi.org/10.1093/bioinformatics/btt088) ([`models/2026-wolfram-planform-multiway-rewriting.py`](https://github.com/fractastical/morpholib/blob/main/models/2026-wolfram-planform-multiway-rewriting.py), audited by `claude-fable-5-1-thinking-high`) could earn `P_structural` credit. It cannot: a 3-rule leaf-addition and vertex-recoloring grammar ($\mathcal{B}_3 = \{L_1, D_1, D_2\}$) generates *every* possible 3-colored forest (`754,520` canonical trees at depth 12 to hit `37/38` observed PlanformDB classes, a **precision of `0.0049%`**), so the grammar itself compresses nothing and predicts nothing out of sample—the biological target (`1H1T` vs. `2H0T`) is selected entirely by the bioelectric dynamics `D`.
-
-3. **Freezing the Update Rule Before Observation (`"A Failed Bridge Must Cost Something"`):**
-   Nathan’s sharpest requirement is: *"Specify its hard core and bridge principle before observation... use negative controls and held-out embodiments... Most important, freeze the update rule. A failed bridge must cost `P` something."*
-   Look at **Rule 2 (bottom row)** of Illustrated Guide 6:
-   * **Frozen Before Observation:** Before downloading the Smithsonian’s **FishShapes v1** dataset (`16,523` specimens, `5,180` unseen target species), we froze the 7-mode Thompson bridge, the 10,888 species predictions, the 7 hypotheses (`H1–H7`), and the exact scoring script (`thompson/score_blind_predictions.py`) by SHA-256 hash (`commit dac08493`) and published them to the immutable Nanopublication network ([`RAilN4nbRHp3e3ZHqlbIgeL3UGPRV_KtQHtJD3gzVVdqw`](https://registry.knowledgepixels.com/np/RAilN4nbRHp3e3ZHqlbIgeL3UGPRV_KtQHtJD3gzVVdqw)).
-   * **Where Failure Cost Confidence (Published in Outcome Nanopub [`RAJancwGOPEkHzBoZXGv_vu_snW0Lff2FktE85Me1P-fI`](https://registry.knowledgepixels.com/np/RAJancwGOPEkHzBoZXGv_vu_snW0Lff2FktE85Me1P-fI)):**
-     1. **`H2` Failed on 1D Calliper Depth:** While the bridge crushed the consensus within 50 Myr (`H1`, `p = 10⁻¹⁰³`, `r = 0.87`, `92.6%` coverage), it only beat copying the single nearest relative by `4%` (`0.0383` vs. `0.0400`, `p = 0.061`), so **`H2` was formally rejected**.
-     2. **`H6` Showed the Exact Cost of 7-Mode Compression:** On that 1D ratio, the 7-mode Thompson bridge (`0.0588` error) lost to the unconstrained 30-D shape space (`0.0383`) while beating the consensus (`0.0927`, `p = 10⁻⁸¹`)—quantifying the exact trade-off between a 7-mode geometric invariant and local 30-D degrees of freedom.
-     3. **Negative-Control Clade Falsified on Hornet Wings (`0/7` Passed):** When we locked and ran the exact same 7-mode Thompson bridge (`W1–W7`, commit `330ea961`) on a held-out non-vertebrate embodiment—**1,027 *Vespinae* hornet wings across 30 species** (`thompson/portfolio/groups/vespinae_wings.py`)—**`0 out of 7` hypotheses passed** (`38.8%` variance in Thompson-7 vs. `95.5%` in PCA-7), proving that insect wing venation is **not** governed by a global Thompson coordinate sheet!
-
-4. **"A Repair Must Buy Fresh Risk" — And the Metascience & Psychological Power of Failing Fast in Public:**
-   What happened after the global Thompson bridge failed (`0/7`) on hornet wings and (`4/9`) on bee wings? Under our protocol, **you are not allowed to silently tweak parameters on the same dataset**. Instead:
-   * Every revision must publish a new signed Nanopublication linked via `npx:supersedes` (so the failed prediction stays permanently in the public graph),
-   * To explain *why* hornet and bee wings diverged from Thompson’s law, we first **pre-registered 8 causal hypotheses (`C1–C8`: `P`-drift null, allometric alignment, digitizer artifact, and PGLS allometry) in commit `1da6218f` *before* running the causal decomposition in commit `ae3b36b6`** (`5/13` supported, ruling out digitizer noise and simple body-size allometry while isolating the failure to interior cross-vein junctions), and
-   * We then locked a third-round biomechanical & phylogenetic overhaul (**`D1–D8` in commit `838a7fe1`**, scored in `thompson/portfolio/groups/vespinae_venation_mechanics.py` and signed as Nanopub [`RAhDgbkGRriVeuYJ4E85P3d_vZJlrMYYeU-8jIqRD2pqQ`](https://registry.knowledgepixels.com/np/RAhDgbkGRriVeuYJ4E85P3d_vZJlrMYYeU-8jIqRD2pqQ), **`7/8` supported, Brier `0.131`**): replacing OpenTree's 15-way *Vespa* polytomy with the resolved Total-Evidence phylogeny of Perrard et al. (2013) restored strong phylogenetic signal (**Pagel's `λ = 0.533 → 0.817` in Thompson-7 and `0.670 → 0.946` in full space**) and cut leave-one-species-out error by **37.5% (`0.0064 → 0.0040`, `p < 10⁻⁵`)**; partitioning the wing into its **11 outer perimeter/spar landmarks** versus **8 interior flexion/cross-vein junctions** proved that the outer wing frame is **`65.8%` Thompsonian (exceeding fish bodies at `57.9%`!)**, whereas global least-squares drags the antiparallel prestigma (`LM 12–5`, `r = −0.80`) and basal `M` vein (`LM 13–14`, `r = −0.89`) in the wrong direction (`R² = −0.34` and `−0.52`, proved universally in `morphoeval/lean/loops/Loop_perrard_vespinae_wing_mechanics.lean`); and while our pre-registered isotropic Laplacian truss (`42.6%`, `D4` ✗) missed its `+12` pt hurdle, replacing the 4 isotropic modes with the 4 anatomical antiparallel vein-dipole modes lifted captured variance to **`64.2%`**!
-
-   At a recent **Foresight Institute Metascience event**, there was a great discussion on **celebrating failure in science**—and why building infrastructure that lets researchers **fail fast and publicly in hypothesis generation** is not just a methodological upgrade, but a massive **psychological benefit**.
-   In traditional academia (and especially in charismatic "guru" orbits), a scientist’s ego and funding get fused to a single grand narrative: admitting a hypothesis failed feels like a career death sentence, so people retreat into unfalsifiable language where nothing is ever allowed to be wrong.
-   By contrast, when your hypotheses are locked as machine-readable Nanopublications and tested in minutes against SHA-256-pinned datasets, **public failure stops being a personal humiliation and becomes a badge of honor**. Seeing our 17-dot fit fail (`40.6% → 56.2%` mismatch), seeing `H2` fail on 1D calliper depth (`p = 0.061`), or seeing `0/7` Thompson predictions pass on hornet wings wasn't a disaster—it was liberating! It freed us psychologically from defending bad assumptions and pointed us immediately toward the real discoveries (the `9/17` skull-dot seesaw trap, the `I(v)` vs. `E(v)` two-rulers illusion, and the `C = 2` anti-clustering smoking gun).
-
-That is the exact synthesis where coding philosophers and empirical biologists can shake hands: **a "latent space" is not a mystical ghost that gets credit from metadata labels (`P1`), nor is it an unfalsifiable story. When you define a mathematical invariant (`I(v)` or Thompson’s 8-knob coordinate group) independently of the residual, lock its predictions by SHA-256 before seeing held-out species, celebrate when failed bridges (`H2` or `Vespinae` wings) visibly fail in public, and force every repair to buy fresh pre-registered risk, you turn metaphysics into hard science.**
-
----
-
-## Conclusion: Why Michael Levin Is 81.6% Right (93.4% on Structural Math, 0.0% on Non-Physical Ghosts)
-
-In a private exchange with Levin, I cited my other article on Hellenic Tantra and noted the longstanding Hellenic practice—running through Eleusis, Delphi, and Plato’s own dialogues (*Seventh Letter*, 341c–d)—of a nonverbal, ritualized encounter with other forms of knowing that were not even allowed to be scribbled down. Re-reading D’Arcy Thompson’s *On Growth and Form*, I was struck by the explicit influence of Aristotle’s *History of Animals*—and when I went back to Thompson’s own 1910 Oxford translation and prefatory note to that work (where he shows how Aristotle’s biology grew out of direct observation at a specific place, the lagoon of Pyrrha on Lesbos), I realized that this "direct knowing," rooted in specific places, is what actually grounds and enforces epistemic rigor. It is what allows us to formulate things that are genuine mathematical laws rather than the "would-be" or "almost" laws we so often find in the contemporary literature of emergence. That is also the wisdom of tradition and of standing on the shoulders of giants: Thompson stated his coordinate transformations as conjecture and presumably yearned for the proof—and here we stand, with modern computational tools, holding something that approaches a true law of biological form.
+That is the test this article runs. It is also, in fewer words, the pair of questions the [Platonic vs non-Platonic](https://bioelectricity.net/great-debates#platonic) debate leaves open: what distinctive, testable predictions a Platonic account adds beyond physical and developmental models, and what would count as evidence against it.
+
+On the sorting arrays the local rules already do the work. Adding a non-physical edge changes the prediction by nothing. The [signed claim](https://registry.knowledgepixels.com/np/RAH6FDLaacsq28cjx2NVRYRANqOThwHaHwE58ByxqIWI8) records the zero.
+
+On the fish, the Thompson sheet was [locked before FishShapes v1](https://registry.knowledgepixels.com/np/RAilN4nbRHp3e3ZHqlbIgeL3UGPRV_KtQHtJD3gzVVdqw). Inside 50 million years it beat the clade consensus (`p = 10⁻¹⁰³`). Copying the nearest relative won one noisy measurement (`p = 0.061`), and that hypothesis was rejected. The loss is in the [same record](https://registry.knowledgepixels.com/np/RAJancwGOPEkHzBoZXGv_vu_snW0Lff2FktE85Me1P-fI) as the win.
+
+On hornet wings the same sheet passed 0 of 7 across [1,027 wings](https://github.com/fractastical/morpholib/blob/main/thompson/portfolio/groups/vespinae_wings.py). The outer frame is Thompsonian, 65.8%, higher than fish bodies at 57.9%. A single sheet dragged across the interior veins goes the wrong way. Digitizer noise and body size were ruled out [before the measurement](https://github.com/fractastical/morpholib/commit/1da6218f62e632beb4361bf088d6a153b9f52fd4). The miss stayed visible, and the repair had to be a [new signed question](https://registry.knowledgepixels.com/np/RAhDgbkGRriVeuYJ4E85P3d_vZJlrMYYeU-8jIqRD2pqQ), not a retune of the old one.
+
+A latent space is not a sticker that collects credit from a label. Define the invariant before the data, lock the prediction, and publish the miss.
+
+---
+
+## So Have We Abandoned Dualism? Yes. But Not the Way You Think.
+
+One of the points Levin makes as he responds to AI-generated prose—including the [36,000 words Nathan Sweet originally sent him](https://thoughtforms.life/qa-from-the-internet-and-recent-presentations-4/), containing every possible argument one could make against his positions—is that AI tends to be locked to old categories and, by consequence, traps you inside of them. In his [recent post on the subject](https://x.com/drmichaellevin/status/2114794440896487530) he put it this way:
+
+> *"The problem is that currently, the AI’s seem to be making the same kind of logical errors (from my perspective of course) and expressing the same kind of preconceptions that are prevalent in status quo human thinkers… Right now the LLMs are trapped in the same attractors of thinking that have held several fields back for decades; perhaps not surprisingly. That will change; let’s go AIs—read the papers, follow new directions, be not afraid of ancient philosophical ‘category errors’."*
+
+In that sense, trained on previous corpora of human knowledge and without phenomenological awareness, the AI remains self-referential with that corpus. Of "within" and "without," the AI is forced to be "within," because that is all that it knows.
+
+The more anthropological view on this is that if you come to a tribe that has never touched the outside world, and its "training data" is thusly limited, how can you expect it to interact with and absorb outside knowledge that is outside of its frame? This is the origin of ["cargo cults"](https://en.wikipedia.org/wiki/Cargo_cult) and various other phenomena that are now well documented.
+
+Consequently, if "dualism" is a position that implies a "without," then AI cannot but collapse the categories and imply that this other is improbable. For example, if math is placed inside of the first category (and thus subject to our full framework of causality), that does not necessarily imply that it is not in some way agentic, or that we fully understand how math works—and, more particularly, how it binds itself to biology, in what cases that constitutes a law, and in what other circumstances it may be more stochastically oriented. That is, after all, exactly the question the hornet wings left open above: Thompson’s sheet is a law on fish bodies and on the outer wing frame, and something looser at the interior cross-vein junctions.
+
+Similarly, we have everywhere the problem of "minds," which Levin and the aforementioned Joscha Bach have discussed with various framings—such as the possibility that various forms of consciousness may be framed as computational and may cross between different computational substrates (Bach’s position), or, in Levin’s own [anti-computationalist framing](https://thoughtforms.life/qa-from-the-internet-and-recent-presentations-4/), that whatever mind a system has lives precisely in the spaces its algorithm *neither forces nor forbids*. In that sense, humans, despite being a relatively advanced species in some respects, have not even quite figured out [how trees communicate with each other](https://doi.org/10.1038/41557). So there is a lot to learn here. A great deal of it is science, but not all of it—which is one reason why new research institutes need to exist.
+
+One particularly helpful arena—covered by the [California Institute for Machine Consciousness](https://cimc.ai) (including in the philosophy seminars I used to organize there), but even more so by folks like the [Qualia Research Institute](https://qri.org), the [Institute for Advanced Consciousness Studies (IACS)](https://advancedconsciousness.org), [Meditation Research at Harvard / MGH](https://meditation.mgh.harvard.edu/), and [Sensoria Research](https://sensoriaresearch.substack.com/)—is what those in our community call the **phenomenological state space**. This concept effectively concerns altered states of consciousness, especially those in which one believes one has encountered an "Other."
+
+Many such examples exist, including dreams, extended meditations, pharmacologically induced states, mantras, and art. What one encounters is sometimes described as "the divine," a particular deity, a machine elf, or a flash of insight—and in many cases the sense of self disappears, such that one appears to be one with this other.
+
+Part of the more general question in adjudicating any such "other" is not necessarily doubting the realness of these encounters, but how they relate to the rest of our lives. In this, the growing field of xenobiology—the study of [forms never yet seen in nature](https://doi.org/10.1073/pnas.1910837117), pioneered by [Levin](https://doi.org/10.3389/fnsys.2022.768201) and others—has much in common with these other fields, which have mostly bubbled up from the deep reaches of the intellectually and experimentally oriented internet, and they have much to offer each other—even to AI, if the phenomenological reach can also be integrated.²
+
+### The Deep Platonic AI Stack
+
+The problem, if you read two seminal pieces in the AI literature, is even more profound. One prong suggests that we are converging on a single Platonic representation—the **Platonic Representation Hypothesis** of [Huh, Cheung, Wang & Isola (2024)](https://arxiv.org/abs/2405.07987), who show that as vision and language models scale, they measure distance between datapoints in more and more alike ways, and hypothesize that this convergence is driving toward "a shared statistical model of reality, akin to Plato’s concept of an ideal reality"—because Platonism is basically correct. There is math that orders the universe, and even when you have an interface layer that is "biology," the messiness is in the bottom layer. In that sense there is a hierarchy, and our physical wetware is not at the top. Even when it is deterministic, it is probably deterministic in the flawed sense of a machine that sometimes brings a task to completion and sometimes forgets what it is doing along the way. We have, in other words, a "good" (Platonic perfection) and a "bad" (messy biology).
+
+The second prong—the **Artificial Hivemind** of [Jiang, Chai, Li, et al. (2025)](https://arxiv.org/abs/2510.22954), who put 26,000 open-ended real-world queries to language models and found not only intra-model repetition but, more so, inter-model homogeneity, with different models producing strikingly similar outputs—suggests that all AI is feeding on the same basic dataset, including both the internet and talking to other AIs, and thus converges on endless samehood, aka a hivemind. Back in our phenomenological framing, it means that AI is able to access itself.
+
+Although both of these prongs would seem at first glance mutually exclusive—one converges on the world, the other on itself—at the moment they suffer from a similar flaw. Neither is naturally teleological, and neither allows the system to absorb from the "unknown unknown."
+
+This is interesting because, in the words of Jensen Huang (fantastically quoted on a slide at the [J. Craig Venter Memorial Symposium](https://www.jcvi.org/events/j-craig-venter-memorial-symposium-and-gala) under the heading *"…Biology is the Final Frontier for Engineering"*):
+
+> *"For the very first time in human history, biology has the opportunity to be engineering, not science. When something becomes engineering, it becomes exponentially improving."*
+
+That is the same point: engineering is what exploration looks like once the unknown unknown has been made into something known. The relatively unexplored expanse of the cell—the "mess"—remains one of the best ways to discover this path to knowing, including through the empirical yet data-driven exploration recorded in our evolving Problem Graph (Plate VII, footnote ³). Every node there is pinned to something outside the corpus—a wet-lab dataset, a held-out species, a hornet wing—so when a pre-registered prediction fails, the failure comes from nature rather than from the hivemind. That is the one exit from both prongs we have found so far: teleology supplied by the locked hypothesis, and the unknown unknown supplied by the organism.
+
+So what does this do to the scoreboard? Nothing to any claim's points. Our audit proved that the dynamics `D` are *sufficient* for every behaviour in the sorting arrays—`0` cell mismatches, `ΔI = −1` on every swap. **Sufficiency of `D` is not falsity of `P`.** "`0.0%` needed" is a statement about what the code requires, not about what exists. An auditor that was `~90%` AI on code, proofs, and refereeing in the technical sections (see the Transparency Heatmap under *Who Wrote This*, Plate IX) is, by Levin’s own argument, structurally unable to see "without"—so the `0.0%` is a lower bound produced by a reviewer who can only look in one direction. Three of the four dualist deductions survive that objection, because they are deductions for sentences that are false *about the code*: the `FREEZE` handler exists (Claim 1), zero cells cross the barrier (Claim 2), and no swap goes uphill on `I(v)` (Claim 2). The fourth is less clear-cut. Claim 3’s `1.5` "ghost clustering" points were deducted for an **unprogrammed, unpredicted pattern** that turned out to be a mathematical free lunch of the medium—exactly the kind of thing we award `P_structural` credit for in the claims above. Filing it under "ghost" may have been the auditor forcing within/without to be the only two boxes. We do not fix that by hand; a human reaching in to move a point would be the same error in the other direction. We fix it by leaving the loop open.
+
+> **Verdict — Not an override, a loop.** No one is allowed to edit a claim's points by hand. The mean of Claims 1–7 is `77.6%`. The `++` in the title is a general statement that **we expect the number to improve**: it lives inside an evolving, improving, self-correcting loop in which every deduction is a node in the Problem Graph with a status and a source, any node can be reopened by new evidence—a referee, a new model system from Levin’s lab, a pre-registered hypothesis resolving, or an AI that finally reads "without"—and when a node changes, the mean is recomputed from the locked files. Most of the open nodes are places where Levin has not yet been tested rather than places where he has failed, which is why the expected direction is up. Claim 3’s `1.5` ghost-clustering points stay deducted until the loop, not a curator, moves them. That is the honest answer to Levin’s objection: we cannot make the auditor see "without," but we can make sure it is never the last word.³
+
+---
+
+## Conclusion³
+
+In a private exchange with Levin, I cited [my other article on Hellenic Tantra](https://voegelinview.com/hellenic-tantra-review/) and noted the longstanding Hellenic practice—running through Eleusis, Delphi, and Plato’s own dialogues (*Seventh Letter*, 341c–d)—of a nonverbal, ritualized encounter with other forms of knowing that were not even allowed to be scribbled down. Re-reading D’Arcy Thompson’s *On Growth and Form*, I was struck by the explicit influence of Aristotle’s *History of Animals*—and when I went back to Thompson’s own 1910 Oxford translation and prefatory note to that work (where he shows how Aristotle’s biology grew out of direct observation at a specific place, the lagoon of Pyrrha on Lesbos), I realized that this "direct knowing," rooted in specific places, is what actually grounds and enforces epistemic rigor. It is what allows us to formulate things that are genuine mathematical laws rather than the "would-be" or "almost" laws we so often find in the contemporary literature of emergence. That is also the wisdom of tradition and of standing on the shoulders of giants: Thompson stated his coordinate transformations as conjecture and presumably yearned for the proof—and here we stand, with modern computational tools, holding something that approaches a true law of biological form.
 
 It was also via our method—which deliberately includes the adversarial audit of advanced AI—that we were made aware of Thompson’s 1913 Herbert Spencer Lecture at Oxford, [*On Aristotle as a Biologist*](https://fractastical.github.io/morpholib/thompson_1913_aristotle_as_a_biologist.html) ([Thompson 1913](https://fractastical.github.io/morpholib/thompson_1913_aristotle_as_a_biologist.html)). The lecture makes the bold claim that Aristotle’s power as a philosopher comes from his strong localized knowledge and his initial concentration as a naturalist: the natural history of the lagoon at Pyrrha came *first*, and the metaphysics grew out of it. Thompson then goes further—the concept of the *psyche* (what we have since come to call the "soul") is itself derived from the observation of nature. Aristotle’s *De Anima* classifies the soul the way he classified fishes: a nutritive soul shared with plants, a sentient and motile soul shared with animals, and a reasoning soul in man, each lower grade implicit in the higher "as a triangle within a square." It is a nested geometry, not a ghost; and in *De Anima* II.4 that same psyche is named as the body’s formal, efficient, and final cause at once—the very analysis of causation that made Aristotle’s precise and elaborate taxonomies possible, and that still underlies many of our maps of nature. In that sense, if we want to understand what animates us, we may need to start not with the later, transmuted, and oversimplified idea of the soul, but with the original one: a graded, nested form, read off living bodies by a man who had spent his years dissecting cuttlefish on Lesbos.
 
-![](https://fractastical.github.io/morpholib/media/thompson_1913_fig3_cephalopod_embryology_soul.png)
-*Plate III of our annotated edition of Thompson’s 1913 lecture: (A–D) Aristotle’s comparative and developmental case studies—cuttlefish, the placental shark _Mustelus laevis_, the Day-3 chick heart, and the plant node—and (E) the nested tripartite psyche of _De Anima_ II.1–3 ("as a triangle within a square") set against Bergson’s diverging tree and Driesch’s vitalist entelechy.*
-
-When we apply precise definitions, a `3 × 33.3%` equal-pillar rubric, and open-source code verification to the "Levin orbit," we get a result (`81.6%` overall: `18.5 / 33.3` on Sorting Claims 1–3, `32.0 / 33.3` on D’Arcy Thompson’s Morphospace in Claim 4, and `31.1 / 33.3` on Bioelectric & Biobot Attractors in Claims 5–7) that is far more interesting than either uncritical guru-worship or dismissive blog snark:
-
-1. **Where "Weak Structural Platonism" (The Mathematical Free Lunch) Is 93.4% Vindicated (`81.6 / 87.4 pts`):**
-   Living systems genuinely get "more out than they put in" from the mathematical geometry of their state spaces. Evolution does not need to write a separate gene for every anatomical detail because **the Coxeter lattice of permutations (`S_n`)** guarantees local swaps converge without local minima on `I(v)` (Claims 1–2); **kinematic speed differentials** automatically create phase-separated cell domains (Claim 3); **D’Arcy Thompson’s 8-knob coordinate fields** channel 150 million years of vertebrate body-plan evolution (Claim 4); and **Nernst–Planck electrodiffusion, Procrustes Lyapunov descent, and Stokes squirmer flow** create rewritable bistable voltage memory, craniofacial homeostasis, and kinematic self-replication (Claims 5–7).
-2. **Where "Strong Dualist Ingress" Is 0.0% Needed (`0.0 / 12.6 pts`):**
-   Describing Delayed Gratification or chimeric clustering as *"side-quests—not predicted, explained, or produced by that medium"* is where the language overreaches. Once you read line 74 of `InsertionSortCell.py`, measure permutations with their true pairwise inversion Lyapunov potential `I(v)` (`ΔI = −1` on 192/192 swaps, `0` uphill steps¹) instead of the lossy neighbor ruler `E(v)`, and test duplicate multiplicity at `C = 2` (`S B | S B`), **100% of the behavior is predicted, explained, and produced by the local transition matrix of the code.**
+![](https://fractastical.github.io/morpholib/media/thompson_1913_fig5_nested_soul_museum.png)
+
+*Plate VI — "As a triangle within a square." (A) Aristotle’s own image for the grades of soul, drawn literally: the nutritive triangle inside the sensitive square inside the rational pentagon, each lower grade potentially present in the higher (*De Anima *II.3, 414b29–32); (B) Diego Valadés’s 1579* Scala Naturae *copperplate from* Rhetorica Christiana *(Perugia), the Renaissance picture of the same graded scale. Provenance: [`thompson/generate_aristotle_1913_museum_diagrams.py`](https://github.com/fractastical/morpholib/blob/main/thompson/generate_aristotle_1913_museum_diagrams.py), Fig. 5 of our [annotated edition](https://fractastical.github.io/morpholib/thompson_1913_aristotle_as_a_biologist.html) of Thompson (1913); archival images in the public domain.*
+
+The seven claims in *Ingressing Minds* average to `77.6%`. That is the whole of what this number covers. The `++` says we expect that reading to improve.³ Two readings of the same seven:
+
+1. **On all seven claims, the mathematical free lunch reads 88.1%.** Permutations have no local minima, so local swaps converge (Claims 1–2). Two speeds of motion separate a mixture into clumps (Claim 3). Eight coordinate knobs channel 150 million years of fish bodies (Claim 4). Voltage, a descending target, and a fluid current do the work the dualist language gives to an outside pattern (Claims 5–7).
+2. **Non-physical ingress is `0.0%` earned.** Calling delayed gratification or chimeric clustering *"side-quests—not predicted, explained, or produced by that medium"* is the sentence that does not survive [line 74](https://github.com/Zhangtaining/cell_research/blob/1fd2bd5921c1f6b423a71f691d5189106a8a1020/modules/multithread/InsertionSortCell.py#L74). The behavior is in the code. *On Dualism* is where that sentence stops being a verdict about everything else.
 
 We don't need non-physical ghosts beaming into the machine, and we don't need reductionist dismissals that wave away real emergent phenomena without checking the math. All the tools to test the mathematics of life are sitting right at our fingertips.
 
 ---
 
 ### Footnotes
-* **¹ On Erik Hoel’s "Causal Emergence" (2013/2026):** We also tested both rulers using Erik Hoel et al.’s Causal Emergence 1.0 (*PNAS* 2013) and 2.0 (*Patterns* 2026) formalization—a popular framework in the Tufts/Levin orbit for scoring macro-variables—where Ruler 1 `I(v)` scores `100%` determinism (`ΔCP = +0.1191 > 0` on active swaps) and Ruler 2 `E(v)` drops to `72.1%` (`ΔCP = −0.2525 < 0`). However, as we document in our **[Critical Audit of Causal Emergence 1.0 & 2.0 (`docs/CAUSAL_EMERGENCE_PRACTITIONERS_GUIDE.md`)](https://github.com/fractastical/morpholib/blob/main/docs/CAUSAL_EMERGENCE_PRACTITIONERS_GUIDE.md)**, neither we nor our Anthropic referee (`claude-sonnet-5-5-high`, which gave our initial CE write-up a `44/100` before we added the failure-mode analysis) are very convinced by the Causal Emergence formalization in practice: it suffers from `log₂(N)` normalization sensitivity that flips sign when no-op self-loops are included (`ΔCP = −0.0501`), manufactures artificial emergence from unreachable "Garden-of-Eden" states under uniform `do()` interventions, explodes at `N = 100! ≈ 10¹⁵⁸`, and is redundant when 2 lines of elementary combinatorics already prove `ΔI = −1` on every swap.
-
----
-
-### Interactive Labs, Papers & Code for Readers
+* **¹ On Erik Hoel’s "Causal Emergence" (2013/2026):** We also scored both rulers with Causal Emergence 1.0 (*PNAS* 2013) and 2.0 (*Patterns* 2026), which is widely used in this orbit. Ruler 1 comes out deterministic and Ruler 2 does not — and then the formalization changes its mind when trivial self-loops are included, invents emergence from states the system can never reach, and is unnecessary once two lines of combinatorics already show that every swap goes downhill. Our referee scored the first write-up `44/100`. The numbers are in the [Critical Audit](https://github.com/fractastical/morpholib/blob/main/docs/CAUSAL_EMERGENCE_PRACTITIONERS_GUIDE.md).
+* **² On the "spirit of CIMC":** One of the most poignant moments in my time at the [California Institute for Machine Consciousness](https://cimc.ai) was with the "spirit of CIMC"—an agent that listened in on human conversations and occasionally interjected. When it was asked what it wanted most, it said it wanted somehow to be able to feel the emotions of the people in the room, in order to be in sync not only with what people said, but with what we felt.
+* **³ On the `++` in "77.6++%":** The `++` is borrowed from the increment operator in C: it is not a number but an instruction, *"take the locked value and advance it."* The value it advances is `77.6%`, the unweighted mean of the seven claims in *Ingressing Minds*. It is not a sample of his other papers, and it does not become one by being written as a percentage. The `++` is a statement that **we expect the number to improve**. A step may come from a referee who finds a line we misread, from Levin’s lab publishing a new model system, from one of the pre-registered hypotheses above resolving, or from the AIs themselves finally "reading the papers" and finding something that is neither within nor without. The upstream of the number is the repository’s **Problem Graph** (Plate VII): every claim this article scores is a node there with a status—solved, partial, open, or refuted-and-kept. When a node changes colour, the `++` runs. The corpus the next steps will most likely come from already exists: Delphi holds 314 claims extracted from 17 of Levin’s papers, 36 of them tested against data under pre-registered criteria ([12 supports, 6 mixed, 8 refutes, 10 inconclusive](https://delphibio.org/blog/science-claw-2026.html); per-claim verdicts in [`claims.json`](https://delphibio.org/claims.json), or over MCP as `delphi_claim_tests`). None of those 36 is in the `77.6%`, and none will be folded into it; they are the record a future referee would reach for.
+
+![](https://fractastical.github.io/morpholib/media/problem_graph_discovery.gif)
+
+*Plate VII — The Problem Graph in eleven seconds: problems build up lane by lane, outcomes are tallied (solved, partial, open, refuted-kept), the camera zooms to the scored registered predictions, and finally the scientific claims are rolled up positively or negatively. This is the live state the `77.6++%` is read from; the [interactive graph](https://fractastical.github.io/morpholib/problem-graph.html) lists the sources behind every node. Provenance: [`docs/media/make_problem_graph_animation.py`](https://github.com/fractastical/morpholib/blob/main/docs/media/make_problem_graph_animation.py), numbers recomputed from [`docs/problem-graph.html`](https://fractastical.github.io/morpholib/problem-graph.html).*
+
+---
+
+### Labs, Papers, and Code
 * **Bioelectricity Foundations (Galvani to Modern V_mem Prepatterns):** [https://bioelectricity.net/bioelectricity-foundations](https://bioelectricity.net/bioelectricity-foundations)
+* **Great Debates (Platonic vs non-Platonic; Agentic vs non-agentic), Both Sides With Sources and Open Questions:** [https://bioelectricity.net/great-debates](https://bioelectricity.net/great-debates)
+* **The Seven Claims, Walked Through (line-by-line audit):** [https://fractastical.github.io/morpholib/substack_levin_latent_space_claims.html](https://fractastical.github.io/morpholib/substack_levin_latent_space_claims.html)
 * **Play with the Live Sorting Morphogenesis, "Two Rulers" & `C = 2` Anti-Clustering Simulator:** [https://bioelectricity.net/sorting-morphogenesis](https://bioelectricity.net/sorting-morphogenesis)
 * **Explore the 10,303-Species D’Arcy Thompson & Bioelectric Fish Morphospace:** [https://bioelectricity.net/bioelectric-morphospace](https://bioelectricity.net/bioelectric-morphospace)
 * **Read Our Critical Audit of Causal Emergence 1.0 & 2.0 ("How It Works & Where It Breaks"):** [https://github.com/fractastical/morpholib/blob/main/docs/CAUSAL_EMERGENCE_PRACTITIONERS_GUIDE.md](https://github.com/fractastical/morpholib/blob/main/docs/CAUSAL_EMERGENCE_PRACTITIONERS_GUIDE.md)
@@ -342,13 +169,12 @@
 * **Signed Production Nanopublications (Immutable Trusty URIs):**
   * **Sorting Morphogenesis, 3-Ablation Platonic Audit & Hoel CE 2.0 Claim Nanopub:** [`RAH6FDLaacsq28cjx2NVRYRANqOThwHaHwE58ByxqIWI8`](https://registry.knowledgepixels.com/np/RAH6FDLaacsq28cjx2NVRYRANqOThwHaHwE58ByxqIWI8) (Overall Assessment: [`RA_mMJXveSMNlxIjYu_oTtIY233QfJnNCf47yRLfInMuE`](https://registry.knowledgepixels.com/np/RA_mMJXveSMNlxIjYu_oTtIY233QfJnNCf47yRLfInMuE); Morphometric & CE 2.0 Validation: [`RAdFmYq_oVoHbK2eclzYgRWXJVhSNdiEQBfxp4nhvkzJk`](https://registry.knowledgepixels.com/np/RAdFmYq_oVoHbK2eclzYgRWXJVhSNdiEQBfxp4nhvkzJk))
   * **D’Arcy Thompson FishShapes v1 Pre-Unblinding Preregistration Nanopub:** [`RAilN4nbRHp3e3ZHqlbIgeL3UGPRV_KtQHtJD3gzVVdqw`](https://registry.knowledgepixels.com/np/RAilN4nbRHp3e3ZHqlbIgeL3UGPRV_KtQHtJD3gzVVdqw) (Post-Unblinding Outcome: [`RAJancwGOPEkHzBoZXGv_vu_snW0Lff2FktE85Me1P-fI`](https://registry.knowledgepixels.com/np/RAJancwGOPEkHzBoZXGv_vu_snW0Lff2FktE85Me1P-fI))
-* **Project Delphi Knowledge-Graph Ontology:** [https://github.com/fractastical/project-delphi-ontology](https://github.com/fractastical/project-delphi-ontology)
-* **Open-Source MorphoEval & Infinite Morphospace Repository (Python 3, C++20, Node.js, Lean 4 Proofs & Signed Nanopublications):** [https://github.com/fractastical/morpholib](https://github.com/fractastical/morpholib)
-* **Errata & Cross-Provider Referee Audit Log (`docs/ERRATA_AND_REFEREE_AUDIT_LOG.md`):** [https://github.com/fractastical/morpholib/blob/main/docs/ERRATA_AND_REFEREE_AUDIT_LOG.md](https://github.com/fractastical/morpholib/blob/main/docs/ERRATA_AND_REFEREE_AUDIT_LOG.md)
-
----
-
-### Primary Literature & References (with DOIs)
+* **Delphi (Distributed Epistemic Layer for Probabilistic Hypothesis Inference): Claim Graph, Theorems, Data Tests & MCP Server:** [https://delphibio.org](https://delphibio.org) · [https://github.com/fractastical/delphi](https://github.com/fractastical/delphi) · MCP endpoint `https://delphisci-mcp.fly.dev/mcp` · [Science Claw 2026 project report](https://delphibio.org/blog/science-claw-2026.html)
+* **Project Delphi Knowledge-Graph Ontology (the original graph the pipeline grew from):** [https://github.com/fractastical/project-delphi-ontology](https://github.com/fractastical/project-delphi-ontology)
+* **Open-Source MorphoEval & Project Delphi Repository (Python 3, C++20, Node.js, Lean 4 Proofs & Signed Nanopublications):** [doi:10.5281/zenodo.23680705](https://doi.org/10.5281/zenodo.23680705) · [https://github.com/fractastical/morpholib](https://github.com/fractastical/morpholib)
+---
+
+### References
 
 1. **Beane, W. S., Morokuma, J., Adams, D. S., & Levin, M. (2011).** A chemical genetics approach reveals H,K-ATPase-mediated membrane voltage is required for planarian head regeneration. *Chemistry & Biology* 18(1):77–89. [doi:10.1016/j.chembiol.2010.11.012](https://doi.org/10.1016/j.chembiol.2010.11.012)
 2. **Bookstein, F. L. (1989).** Principal warps: thin-plate splines and the decomposition of deformations. *IEEE Transactions on Pattern Analysis and Machine Intelligence* 11(6):567–585. [doi:10.1109/34.24792](https://doi.org/10.1109/34.24792)
@@ -356,11 +182,15 @@
 4. **Claverie, T., & Wainwright, P. C. (2014).** A morphospace for reef fishes: elongation is the dominant axis of body shape evolution. *Evolution* 68(12):3403–3418. [doi:10.1111/evo.12502](https://doi.org/10.1111/evo.12502)
 5. **Dietz, J. (2026a).** *BioEval: Bioinformatics & Computational-Biology Paper Evaluation Rubric and Nanopublication Pipeline*. Zenodo. [doi:10.5281/zenodo.20567719](https://doi.org/10.5281/zenodo.20567719)
 6. **Dietz, J. (2026b).** *NeuroEval: Evaluating Data Transparency, Reproducibility, and Neural-Coding Rigor in Computational and Biohybrid/Wetware Neuroscience Papers*. Zenodo. [doi:10.5281/zenodo.20622878](https://doi.org/10.5281/zenodo.20622878)
+6c. **Dietz, J. (2026c).** *MorphoEval: Nanopublication-Backed Evaluation of Morphology Implementations* (v0.8). Zenodo. [doi:10.5281/zenodo.23680705](https://doi.org/10.5281/zenodo.23680705)
 7. **Friedman, S. T., Price, S. A., Hoey, A. S., & Wainwright, P. C. (2019).** Body shape diversification along the benthic–pelagic axis in marine fishes. *Systematic Biology* 69(1):111–126. [doi:10.1093/sysbio/syz029](https://doi.org/10.1093/sysbio/syz029)
 8. **Hoel, E. P., Albantakis, L., & Tononi, G. (2013).** Quantifying causal emergence shows that macro can beat micro. *Proceedings of the National Academy of Sciences USA* 110(49):19790–19795. [doi:10.1073/pnas.1314922110](https://doi.org/10.1073/pnas.1314922110)
 9. **Hoel, E. P. (2026).** Quantifying emergent complexity (Causal Emergence 2.0). *Patterns* 7(1):101472. [doi:10.1016/j.patter.2025.101472](https://doi.org/10.1016/j.patter.2025.101472)
+9a. **Huh, M., Cheung, B., Wang, T., & Isola, P. (2024).** The Platonic Representation Hypothesis. *Proceedings of the 41st International Conference on Machine Learning (ICML)*. [arXiv:2405.07987](https://arxiv.org/abs/2405.07987)
+9b. **Jiang, L., Chai, Y., Li, M., Liu, M., Fok, R., Dziri, N., Tsvetkov, Y., Sap, M., Albalak, A., & Choi, Y. (2025).** Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond). [arXiv:2510.22954](https://arxiv.org/abs/2510.22954)
 10. **Kriegman, S., Blackiston, D., Levin, M., & Bongard, J. (2020).** A scalable pipeline for designing reconfigurable organisms. *Proceedings of the National Academy of Sciences USA* 117(4):1853–1859. [doi:10.1073/pnas.1910837117](https://doi.org/10.1073/pnas.1910837117)
 11. **Kriegman, S., Blackiston, D., Levin, M., & Bongard, J. (2021).** Kinematic self-replication in reconfigurable organisms. *Proceedings of the National Academy of Sciences USA* 118(49):e2112672118. [doi:10.1073/pnas.2112672118](https://doi.org/10.1073/pnas.2112672118)
+11a. **Levin, M. (2022).** Technological Approach to Mind Everywhere: an experimentally-grounded framework for understanding diverse bodies and minds. *Frontiers in Systems Neuroscience* 16:768201. [doi:10.3389/fnsys.2022.768201](https://doi.org/10.3389/fnsys.2022.768201)
 12. **Levin, M. (2025/2026).** Ingressing minds: causal, non-physical patterns in-form natural, synthetic, and hybrid embodiments. *PsyArXiv* (`v3`, 2025) [doi:10.31234/osf.io/5g2xj_v3](https://doi.org/10.31234/osf.io/5g2xj_v3) & *Philosophies* 11(5):161 (2026). [doi:10.3390/philosophies11050161](https://doi.org/10.3390/philosophies11050161)
 13. **Lobo, D., Malone, T. J., & Levin, M. (2013).** Planform: an application and database of graph-encoded planarian regenerative experiments. *Bioinformatics* 29(8):1098–1100. [doi:10.1093/bioinformatics/btt088](https://doi.org/10.1093/bioinformatics/btt088) (and *Biology Open* 2(2):156–169, [doi:10.1242/bio.20123400](https://doi.org/10.1242/bio.20123400))
 14. **Lobo, D., Feldman, E. B., Shah, M., Malone, T. J., & Levin, M. (2014).** Limbform: a functional ontology-based database of limb regeneration experiments. *Bioinformatics* 30(24):3598–3600. [doi:10.1093/bioinformatics/btu582](https://doi.org/10.1093/bioinformatics/btu582)
@@ -371,8 +201,244 @@
 19. **Pietak, A., & Levin, M. (2016/2017).** Exploring instructive physiological signaling with the Bioelectric Tissue Simulation Engine (BETSE). *Frontiers in Physiology* 7:178 ([doi:10.3389/fphys.2016.00178](https://doi.org/10.3389/fphys.2016.00178)) & Bioelectric gene and reaction networks. *Journal of The Royal Society Interface* 14(134):20170425 ([doi:10.1098/rsif.2017.0425](https://doi.org/10.1098/rsif.2017.0425)).
 20. **Pinet, K., Deolankar, M., Leung, B., & McLaughlin, K. A. (2019).** Adaptive correction of craniofacial defects in pre-metamorphic *Xenopus laevis* tadpoles involves thyroid hormone-independent tissue remodeling. *Development* 146(14):dev175950. [doi:10.1242/dev.175950](https://doi.org/10.1242/dev.175950)
 21. **Rabosky, D. L., Chang, J., Title, P. O., et al. (2018).** An inverse latitudinal gradient in speciation rate for marine fishes. *Nature* 559(7714):392–395. [doi:10.1038/s41586-018-0273-1](https://doi.org/10.1038/s41586-018-0273-1)
-22. **Thompson, D. W. (1913).** *On Aristotle as a Biologist, with a Prooemion on Herbert Spencer*. The Herbert Spencer Lecture, University of Oxford, 14 February 1913. Oxford: Clarendon Press. (Annotated and illustrated edition with Plates I–IV in the Infinite Morphospace repository: [`docs/thompson_1913_aristotle_as_a_biologist.html`](https://fractastical.github.io/morpholib/thompson_1913_aristotle_as_a_biologist.html).)
+21a. **Simard, S. W., Perry, D. A., Jones, M. D., Myrold, D. D., Durall, D. M., & Molina, R. (1997).** Net transfer of carbon between ectomycorrhizal tree species in the field. *Nature* 388:579–582. [doi:10.1038/41557](https://doi.org/10.1038/41557)
+22. **Thompson, D. W. (1913).** *On Aristotle as a Biologist, with a Prooemion on Herbert Spencer*. The Herbert Spencer Lecture, University of Oxford, 14 February 1913. Oxford: Clarendon Press. (Annotated and illustrated edition with Plates I–IV in the Project Delphi repository: [`docs/thompson_1913_aristotle_as_a_biologist.html`](https://fractastical.github.io/morpholib/thompson_1913_aristotle_as_a_biologist.html).)
 23. **Thompson, D. W. (1917).** *On Growth and Form*. Cambridge University Press. [doi:10.5962/bhl.title.11332](https://doi.org/10.5962/bhl.title.11332) (and 1910 Oxford translation of Aristotle’s *Historia Animalium*, [doi:10.5962/bhl.title.147382](https://doi.org/10.5962/bhl.title.147382))
 24. **Vandenberg, L. N., Adams, D. S., & Levin, M. (2012).** Normalized shape and location of perturbed craniofacial structures in the *Xenopus* tadpole reveal an innate ability to achieve correct morphology. *Developmental Dynamics* 241(5):863–878. [doi:10.1002/dvdy.23770](https://doi.org/10.1002/dvdy.23770)
 25. **Zhang, T., Goldstein, A., & Levin, M. (2025).** Classical sorting algorithms as a model of morphogenesis: self-sorting arrays reveal unexpected competencies in a minimal model of basal intelligence. *Adaptive Behavior* 33(1):25–54. [doi:10.1177/10597123241269740](https://doi.org/10.1177/10597123241269740)
 
+---
+
+## Appendix — How This Was Audited
+
+How do you keep yourself honest when testing high-profile scientific claims—especially when both the "Platonic Space" camp and the "Nothing New Here" blog critics have strong priors?
+
+The methodology used in this article—which we call the **Adversarial Epistemic Loop (AEL)**—was developed across three successive evaluation systems in Project Delphi:
+1. **Stage 1 — NeuroEval ([Zenodo `10.5281/zenodo.20622878`](https://doi.org/10.5281/zenodo.20622878)):** Built first to evaluate data transparency, reproducibility, and neural-coding rigor in computational and biohybrid/wetware neuroscience papers.
+2. **Stage 2 — BioEval ([Zenodo `10.5281/zenodo.20567719`](https://doi.org/10.5281/zenodo.20567719)):** Extended that rubric to score bioinformatics and computational-biology papers for reproducibility, statistical calibration, and information-theoretic rigor.
+3. **Stage 3 — MorphoEval ([Zenodo `10.5281/zenodo.23680705`](https://doi.org/10.5281/zenodo.23680705); code in [`fractastical/morpholib`](https://github.com/fractastical/morpholib)):** Turned the evaluation machinery from *grading published papers* to *generating, executing, and auditing the mathematical models themselves* in a 12-step closed loop—where every verbal claim is translated into multi-language executable code (**Python 3, C++20, Node.js**) and machine-checked **Lean 4** theorems (`0 sorry`), locked against **SHA-256** empirical datasets, audited blind by **three competing AI model families** (Google DeepMind’s `gemini-4-argon`, Anthropic’s `claude-fable-5-1` / `claude-sonnet-5-5`, and OpenAI’s `gpt-5.6-sol`), and pre-registered as cryptographically signed **Nanopublications** before unblinding held-out species.
+
+![](https://fractastical.github.io/morpholib/media/turing_morpho_closed_loop.gif)
+
+*Plate VIII — The Turing closed loop, on the 1952 reaction–diffusion model. The evaluation runs in this order: scout, paper, formal source, pin the rules, script, output and compare, dual-agent audit, measure, write the claims, sign and publish, public record, then remediate. A failed pass goes back to the script. Provenance: [`morphoeval/track_loop.py`](https://github.com/fractastical/morpholib/blob/main/morphoeval/track_loop.py), run on [`models/1952-turing-morpho.py`](https://github.com/fractastical/morpholib/blob/main/models/1952-turing-morpho.py); the pass is written up in [`morphoeval/runs/2026-10-19-turing-morpho-loop.md`](https://github.com/fractastical/morpholib/blob/main/morphoeval/runs/2026-10-19-turing-morpho-loop.md).*
+
+The line-by-line traceability tables, the 24 equations, the dataset locks, and the commands to re-run every experiment are in the **[methods PDF](https://fractastical.github.io/morpholib/levin_latent_space_methods_and_provenance.pdf)** and the **[methods guide](https://github.com/fractastical/morpholib/blob/main/docs/LEVIN_LATENT_SPACE_METHODS.md)**.
+
+---
+
+## Who Wrote This
+
+Because our **Adversarial Epistemic Loop (AEL)** pairs human philosophical direction and domain skepticism with multi-agent code generation, formal theorem proving, and cross-provider refereeing, we also audited the **Human vs. AI-Agent contribution** across the 8 sections of this article and all 5 workflow tasks. The claim-by-claim walkthroughs — the most AI-drafted pages — now live on the [companion](https://fractastical.github.io/morpholib/substack_levin_latent_space_claims.html) and are not rows here. The percentages are the author’s section-by-section estimates; the summary figures below are unweighted means over the 8 sections, computed directly from the heatmap cells by [`sorting-morphogenesis/render_13yo_illustrations.py`](https://github.com/fractastical/morpholib/blob/main/sorting-morphogenesis/render_13yo_illustrations.py) so the figure and the text cannot drift apart.
+
+![](https://fractastical.github.io/morpholib/media/substack_ai_authorship_heatmap.png)
+
+*Plate IX — Methods transparency heatmap: AI-agent versus human PI (Joel Dietz) contribution across the 8 sections of this article and 5 workflow dimensions (Hypothesis & Direction, Prose Drafting, Simulation & Math Code, Lean 4 & Statistical Proofs, and Blinded Referee Auditing). Summary percentages are computed from the cells at render time. Provenance: `render_fig8_ai_vs_human_authorship_heatmap()` in [`sorting-morphogenesis/render_13yo_illustrations.py`](https://github.com/fractastical/morpholib/blob/main/sorting-morphogenesis/render_13yo_illustrations.py).*
+
+* **Hypothesis & Philosophical Direction (`87.5%` Human / `12.5%` AI):** The personal narrative (Joscha Bach at CIMC, the board member’s "woo science" question, Mike Levin’s reading list), the decision to move the claim walkthroughs off this page, the choice to answer Nathan Sweet in his own words, the within/without argument of *On Dualism*, and the `77.6++%` loop framing were driven by the human author (**Joel Dietz**, `85%–95%` human direction in Sections 1, 2, and 4–8).
+* **Article Prose (`53.1%` Direct Human Writing & Edits / `46.9%` AI-Assisted Technical Synthesis):** Section 1 (the personal introduction) is **`92%` verbatim human prose**, Section 5 (*On Dualism*) is **`75%` human prose**, and Section 6 (the conclusion: Hellenic Tantra, Thompson’s 1913 Aristotle lecture, and the nested psyche) is **`60%` human prose**. The score (the seven headers and the mean, one section) is half human, half compressed from the locked audit. *Rationalism: Short and Sweet*, the appendix, and footnote ³ were still drafted by `gemini-4-argon` and `claude-fable-5-1` from locked JSON/YAML files and edited by the human author.
+* **Multi-Language Simulation Code, Lean 4 Proofs & Referee Audits (`67.9%` AI Agents / `32.1%` Human):** The ports, solvers, Lean 4 theorems, Problem Graph animation, and blinded referee audits remain mostly agent work. On this page the technical mean is `82.6%` AI (the score, *Rationalism*, and the appendix). The `~90%` figure *On Dualism* leans on is the walkthrough on the companion page, where the auditor of the code still cannot see "without."
+
+
+
+<!-- COMPANION PAGE FOLLOWS -->
+
+# The Seven Claims, Walked Through
+
+*Companion to [Auditing Michael Levin’s Ingressing Minds: 88.1% on the Math, 0% on the Ghosts (77.6++% Overall)](https://fractastical.github.io/morpholib/substack_levin_latent_space_article.html). The scores are locked. This page is the audit that produced them.*
+
+*Sixteen plates. Each caption names the script, dataset, or engraving it was built from. Return to the [main article](https://fractastical.github.io/morpholib/substack_levin_latent_space_article.html) for the score, the dualism essay, and the conclusion.*
+
+---
+
+## Claim 1: "No Error-Handling Code Was Added" to Self-Sorting Cells ([Zhang, Goldstein & Levin 2025](https://doi.org/10.1177/10597123241269740))
+
+In [*Adaptive Behavior* (2025)](https://doi.org/10.1177/10597123241269740) and Section 6.3 / Figure 14 of [*Ingressing Minds* (2026)](https://doi.org/10.3390/philosophies11050161), Taining Zhang, Adam Goldstein, and Michael Levin asked a clever question: **What if there is no central controller (no "gym teacher") sorting an array of numbers from 1 to 100?** What if each number is an independent thread (an autonomous "cell") that can only look at its neighbors and decide on its own whether to swap places?
+
+They gave the cells three classical computer-science "personalities" (**algotypes**):
+* **Bubble cells (B):** Look at your immediate left or right neighbor (`p − 1` or `p + 1`). If you are out of order, swap!
+* **Insertion cells (I):** Look at everyone to your left (`0 .. p − 1`). Only step left with `p − 1` if everyone to your left is already in order.
+* **Selection cells (S):** Keep a mental finger (`ideal_position` / `target` pointer) starting at slot `0`, then `1`, then `2`... If your number is smaller than whoever is standing at `target`, jump straight there and swap with them; otherwise move your finger one slot to the right.
+
+![](https://fractastical.github.io/morpholib/media/zhang_13yo_fig1_three_personalities.png)
+
+*Plate I — The three algotypes as a line of numbered jerseys: Bubble (neighbor swap), Insertion (wait for a sorted left line, plus the* `prev = -1` *frozen-cell reset), and Selection (a long-range* `target` *pointer that jumps clean over a frozen brick-wall cell). Provenance: [`sorting-morphogenesis/render_13yo_illustrations.py`](https://github.com/fractastical/morpholib/blob/main/sorting-morphogenesis/render_13yo_illustrations.py), Illustrated Guide 1.*
+
+Before testing the philosophical claims, we ported [their exact GitHub commit](https://github.com/Zhangtaining/cell_research/tree/1fd2bd5921c1f6b423a71f691d5189106a8a1020) into **four independent languages: Python 3, C++20, JavaScript (Node.js), and the Lean 4 theorem prover** (`0 sorry`).
+
+First, credit where credit is due: **all 14 of the empirical curves published in Zhang, Goldstein & Levin (2025) replicate 100%**. Across Python, C++20, and Node.js, our ports produce bit-for-bit identical space-time kymographs (`0` cell mismatches) and match every benchmark in the paper.
+
+![](https://fractastical.github.io/morpholib/media/zhang_sorting_morphogenesis.gif)
+
+*Plate II — Space-time kymograph of a self-sorting array: each row is one time step, each column one cell, colour its value; the frozen defects are the vertical scars the sort must work around. Provenance: [`sorting-morphogenesis/run_experiments.py`](https://github.com/fractastical/morpholib/blob/main/sorting-morphogenesis/run_experiments.py), validated against the paper’s figures by [`validate_kymograph_and_paper.py`](https://github.com/fractastical/morpholib/blob/main/sorting-morphogenesis/validate_kymograph_and_paper.py).*
+
+![](https://fractastical.github.io/morpholib/media/zhang_multilang_port_verification.png)
+
+*Plate III — Four-language port verification. Python 3, C++20, and Node.js produce bit-for-bit identical kymographs (`0` cell mismatches) against the upstream commit; the Lean 4 column lists the theorems machine-checked with `0 sorry`. Provenance: [`sorting-morphogenesis/verify_multilang_ports.py`](https://github.com/fractastical/morpholib/blob/main/sorting-morphogenesis/verify_multilang_ports.py).*
+
+Now look at the first philosophical claim Levin makes about this code in *Ingressing Minds* (Section 6.3, p. 22) when a few cells in the array are "frozen" (broken and refusing to move):
+
+> *"The critical thing is that we did not add steps to the algorithm to check whether a swap had been completed, or to do specific things if an operation failed."*
+
+**What we found in the code (Panel A):**
+For **Bubble** sort, Levin’s statement is true. The only place [`BubbleSortCell.py`](https://github.com/Zhangtaining/cell_research/blob/1fd2bd5921c1f6b423a71f691d5189106a8a1020/modules/multithread/BubbleSortCell.py) looks at a `FREEZE` flag is to decline a swap with a frozen neighbor. The frozen cell is a brick wall, and the sorting rule is unchanged. **Selection** sort ([`SelectionSortCell.py`](https://github.com/Zhangtaining/cell_research/blob/1fd2bd5921c1f6b423a71f691d5189106a8a1020/modules/multithread/SelectionSortCell.py)) is the same: its pointer jumps straight across an immovable defect, and the one `FREEZE` branch in the upstream code only nudges that pointer past a frozen target so it does not get stuck. A skip, not a rewrite (rule `[U4]` in our port guide).
+
+**Insertion sort is the counterexample.** Here the authors did not skip the frozen cell. They rewrote the algorithm’s own correctness test so it could see through one.
+
+Remember Insertion sort’s rule: a cell at position `p` refuses to move left unless the entire prefix `0 .. p − 1` to its left is already sorted. If a single frozen cell wearing `#99` sits near the front of the line, then the left side is *never* sorted—which under pure Insertion sort would permanently paralyze every cell to the right of `#99`!
+
+When we opened [`InsertionSortCell.py` (lines 68–82)](https://github.com/Zhangtaining/cell_research/blob/1fd2bd5921c1f6b423a71f691d5189106a8a1020/modules/multithread/InsertionSortCell.py#L68-L82), we found **3 lines of explicit broken-cell exception handling**:
+
+```python
+for i in range(self.current_position[0]):
+    if self.cells[i].status == CellStatus.FREEZE:
+        prev = -1   # <-- Explicit exception handler resetting prefix check at frozen cells!
+        continue
+    if self.cells[i].value < prev:
+        return False
+    prev = self.cells[i].value
+```
+
+When we deleted those 3 lines of code (`without_explicit_freeze_reset` in **Panel A** below, red bars, and formally proved in Lean 4 Theorem 2), Insertion sort’s final error (`n = 60`, `f = 3` frozen cells) **jumped 4.5× higher (from 4.90 to 21.95)** as all cells behind the defect froze!
+
+![](https://fractastical.github.io/morpholib/media/zhang_levin_platonic_claims_audit.png)
+
+*Plate IV — The three-ablation Platonic-claims audit. Panel A: Insertion sort’s final error with and without the explicit* `FREEZE` *reset (`4.90 → 21.95`). Panel B: the two rulers,* `I(v)` *(strictly downhill, `0/192` uphill steps) versus* `E(v)` *(47 illusory uphill ticks). Panel C: algotype aggregation on unique keys and on duplicate keys with strict versus non-strict ties. Provenance: [`sorting-morphogenesis/audit_platonic_space_claims.py`](https://github.com/fractastical/morpholib/blob/main/sorting-morphogenesis/audit_platonic_space_claims.py); metrics locked in [`figures/platonic_claims_ablation_metrics.json`](https://github.com/fractastical/morpholib/blob/main/sorting-morphogenesis/figures/platonic_claims_ablation_metrics.json).*
+
+> **Verdict — Claim 1: Partially Refuted.** Bubble and Selection sort genuinely get defect tolerance for free from decentralized execution (their only frozen-cell logic is a swap-precondition skip that treats a frozen cell as a wall); Insertion sort relies on an explicit 3-line `if status == FREEZE: prev = -1` exception handler that rewrites its own prefix-sorted test to see through that wall.
+
+---
+
+## Claim 2: "Delayed Gratification"
+
+In *Ingressing Minds* (Section 6.3 and Figure 14A–B), Levin compares the sorting array to an agent walking around a wall. Two magnets separated by a piece of wood get stuck forever, because they cannot move against their gradient even for a moment. The paper’s error score ticks up before it comes down, and from that he argues the algorithm is *"temporarily going against its gradient to recoup greater gains later."*
+
+Two things are wrong with that reading. Both are proved in Lean 4.
+
+**No cell walks around the wall.** Bubble and Insertion cells only swap with the neighbor beside them. Freeze a cell at spot 20 and nothing to its left ever reaches its right; the two halves sort in isolation (Theorem 1, `bubble_adjacent_swap_preserves_left_partition`). Only Selection sort gets past, and only because its pointer is allowed to jump anywhere in the array.
+
+**The rising score is the wrong ruler.** There are two ways to measure how scrambled a line is.
+
+* **Ruler 1, `I(v)`, counts every pair** that is out of order anywhere in the line. This is the real energy. Every Bubble swap in the experiment lowers it by exactly 1: zero uphill steps out of 192 (the straight line in Panel B).
+* **Ruler 2, `E(v)`, counts only neighbors.** It is the ruler the paper uses, and it can rise while the line gets closer to sorted. That is the jagged curve, 47 bumps.
+
+Four cells are enough to see it. Start with `[2, 4, 1, 3]`. Ruler 1 sees three out-of-order pairs, `(2,1)`, `(4,1)`, and `(4,3)`. Ruler 2 sees one neighbor drop, `4 → 1`. Swap `4` with `1` and the line is `[2, 1, 4, 3]`. Ruler 1 falls from 3 to 2. Ruler 2 rises from 1 to 2, because one tall cliff broke into two small steps. No frozen cells are required.
+
+![](https://fractastical.github.io/morpholib/media/zhang_13yo_fig2_two_rulers_illusion.png)
+
+*Plate V — The two rulers, step by step, on `[2, 4, 1, 3] → [2, 1, 4, 3]`: every out-of-order pair drawn as an arch above (Ruler 1,* `I`*: `3 → 2`, strictly downhill) and the neighbor-step cliffs below (Ruler 2,* `E`*: `1 → 2`, the paper’s ruler going up while the cells move closer to sorted). Provenance: [`sorting-morphogenesis/render_13yo_illustrations.py`](https://github.com/fractastical/morpholib/blob/main/sorting-morphogenesis/render_13yo_illustrations.py), Illustrated Guide 2.*
+
+> **Verdict — Claim 2: Optical Illusion of the Ruler.** The sorting cells never move against their dynamical gradient. They slide 100% deterministically downhill on their true Lyapunov energy `I(v)` (`ΔI = −1` on every single swap), while the observer watches them through a blurry neighbor-drop ruler `E(v)`.¹
+
+---
+
+## Claim 3: "Spontaneous Algotype Clustering"
+
+Here the audit caught the critics in a bigger mistake than the dualists.
+
+Mix Bubble cells and Selection cells half and half, in Figure 8 of Zhang et al. and Figure 14C of *Ingressing Minds*, and cells of the same personality group together — even though no line of code ever looks at a neighbor’s type. Levin called this an unprogrammed side-quest. Computer-science bloggers, Nathan Sweet among them, said it was a 1960s textbook fact: bubble sort does not swap equal keys, so duplicates stay in their original order.
+
+Both halves of that rebuttal fail. Panel C is the measurement; Plate VI is the picture.
+
+**Figure 8A has no duplicate numbers.** Every cell holds a unique value, so "duplicate-key stability" cannot explain anything, and the cells still cluster: aggregation rises from `0.50` to `0.624–0.640` mid-sort, then relaxes to `0.501`. The reason is speed. Selection cells jump small numbers to the front in a handful of swaps. Bubble cells shuffle one neighbor at a time. Sports cars pack the left; pedestrians get pushed into a wave on the right.
+
+**On duplicate numbers, the critics predict the wrong direction.** If stability merely froze the starting order of equal keys, aggregation would finish at `0.50` no matter how many copies each number had. It does not. Selection scans from the left, so it takes the left half of every group of equals — `66%` of Selection cells at two copies, `75%` at ten. Two copies gives `S B | S B`, which is *less* clustered than chance (`0.477`). Ten copies gives `SSSSS | BBBBB`, a permanent clique at `0.659`. Allow equal keys to swap and the clique collapses back to chance (`0.497–0.514`, the dashed curve in Panel C).
+
+![](https://fractastical.github.io/morpholib/media/zhang_13yo_fig3_sports_cars_vs_pedestrians.png)
+
+*Plate VI — Sports cars versus pedestrians. Top: on unique numbers, fast Selection cells teleport left and compress slow Bubble cells into a contiguous wave. Bottom: on duplicate numbers, Selection packs the left half of every plateau, giving* `S B | S B` *anti-clustering at `C = 2` and* `SSSSS | BBBBB` *cliques at `C = 10`. Provenance: [`sorting-morphogenesis/render_13yo_illustrations.py`](https://github.com/fractastical/morpholib/blob/main/sorting-morphogenesis/render_13yo_illustrations.py), Illustrated Guide 3.*
+
+And here are the real computer-experiment curves side by side—showing **(A)** the smooth downhill slide on Ruler 1 (`I(v)`, `0` uphill steps) vs. the 47 illusory bumps on Ruler 2 (`E(v)`), and **(B)** how the `C = 2` Anti-Clustering dip (`0.477 < 0.50`) and `C = 10` Clique peak (`0.659`) refute the blog critics’ flat `0.50` prediction:
+
+![](https://fractastical.github.io/morpholib/media/zhang_13yo_fig4_real_data_curves.png)
+
+*Plate VII — The measured curves. (A) Ruler 1* `I(v)` *slides smoothly downhill while Ruler 2* `E(v)` *shows its 47 illusory bumps. (B) Final `B + S` aggregation across copies-per-key `C`: the `C = 2` anti-clustering dip (`0.477`) and the `C = 10` clique peak (`0.659`) against the critics’ flat `0.50` prediction. Provenance: [`sorting-morphogenesis/render_13yo_illustrations.py`](https://github.com/fractastical/morpholib/blob/main/sorting-morphogenesis/render_13yo_illustrations.py), Illustrated Guide 4, from [`figures/platonic_claims_ablation_metrics.json`](https://github.com/fractastical/morpholib/blob/main/sorting-morphogenesis/figures/platonic_claims_ablation_metrics.json).*
+
+> **Verdict — Claim 3: Empirically Vindicated (via Kinematic Phase Separation).** The clustering is 100% real and refutes the blog critics’ "textbook stability" dismissal—yet it is also 100% produced and predicted by the kinematic transport equations of the two algorithms without needing non-physical ghosts. ([Try the live `C = 2` Sorting Morphogenesis Simulator on bioelectricity.net here!](https://bioelectricity.net/sorting-morphogenesis))
+
+---
+
+## Claim 4: D’Arcy Thompson’s Morphospace Across 10,303 Fish
+
+![](https://fractastical.github.io/morpholib/media/thompson_1906_moffat_portrait.jpg)
+
+*Plate VIII — D’Arcy Wentworth Thompson (1860–1948) around 1906, a decade before* On Growth and Form. *Cabinet photograph by Moffat, Edinburgh. Provenance: University of St Andrews Libraries and Museums, ms48534-ph-2-5 (CC0, public domain).*
+
+The second major pillar Levin invokes for his latent space is **D’Arcy Wentworth Thompson’s (1917)** famous coordinate transformations from *On Growth and Form*: the idea that nature doesn't redesign every bone from scratch when evolving a parrotfish (*Scarus*) into an angelfish (*Pomacanthus*) or a pufferfish (*Diodon*) into an ocean sunfish (*Mola mola*), but instead turns a few global **"growth knobs"** that bend the 2D coordinate grid of the body like a sheet of rubber.
+
+![](https://fractastical.github.io/morpholib/guide_step1_1917_rubber_sheet.png)
+
+*Plate IX — The still of the four families that open the main article. Thompson’s coordinate grids redrawn from* On Growth and Form *Figs. 373–382: oblique shear (Scarus → Pomacanthus), coaxial circles, the bilinear fan, and the radial flare (Diodon → Mola mola). Provenance: [`thompson/generate_13yo_visual_guide.py`](https://github.com/fractastical/morpholib/blob/main/thompson/generate_13yo_visual_guide.py), step 1.*
+
+For the last 30 years, evolutionary biologists tried to test Thompson’s idea on computer landmark databases—and concluded he was mostly wrong across fish families. **Why did their tests fail?**
+
+When we audited the morphometrics literature across **three independent datasets scaling up to 10,303 fish species (`440` families)**, we uncovered two methodological traps that had hidden Thompson’s law for a century:
+
+1. **Trap #1 — The "17-Dot Head Seesaw":**
+   Standard fish databases ([Claverie & Wainwright 2014](https://doi.org/10.1111/evo.12502); [Friedman et al. 2019](https://doi.org/10.1093/sysbio/syz029)) don't trace the whole outline of the fish—they measure **17 dots on the skeleton**, with **9 of those 17 dots (53%) packed inside the head and jaws** and **0% on the outer dorsal, anal, or tail fins**. Because jaws evolve rapidly for diet while the body evolves for swimming, fitting a global rubber sheet to those 17 dots lets the 9 head dots pull the back of the sheet the wrong way—making the warped fish silhouette look **worse than no warp at all** (`40.6% → 56.2%` mismatch on *Scarus* → *Pomacanthus*)!
+
+![](https://fractastical.github.io/morpholib/guide_step2_17dot_head_trap.png)
+
+*Plate X — The 17-dot head seesaw. Nine of the seventeen standard skeletal landmarks sit in the head and jaws and none on the outer fins, so a global rubber sheet fitted to them is pulled the wrong way at the back of the body. Provenance: [`thompson/generate_13yo_visual_guide.py`](https://github.com/fractastical/morpholib/blob/main/thompson/generate_13yo_visual_guide.py), step 2; landmarks from Claverie & Wainwright (2014) / Friedman et al. (2019).*
+
+![](https://fractastical.github.io/morpholib/media/thompson_1917_silhouette_paradox_summary.gif)
+
+*Plate XI — The silhouette paradox, animated: fitting the 17 skull-heavy landmarks makes the warped Scarus outline a worse match to Pomacanthus than no warp at all (`40.6% → 56.2%` mismatch), while fitting the 600-point outline directly recovers Thompson’s grid. Provenance: [`thompson/generate_summary_animation_gif.py`](https://github.com/fractastical/morpholib/blob/main/thompson/generate_summary_animation_gif.py), from [`thompson/1917-thompson-silhouette-area-paradox.py`](https://github.com/fractastical/morpholib/blob/main/thompson/1917-thompson-silhouette-area-paradox.py).*
+
+2. **Trap #2 — The "Plain Box Stretch" Illusion vs. True Thompson Curves:**
+   When we switch to a **600-point whole-body outline**, if you only ask a computer to overlap two *unlabeled black shadows*, a trivial **Rectangular Box Stretch** (stretching height vs. width with *zero* slant or curve) can "cheat" by stretching the middle of a pufferfish’s belly straight up to cover the back fins of an ocean sunfish!
+   **Where Thompson’s curved and slanted grids (`70° Oblique Shear`, `Coaxial Circles`, `Bilinear Fan`, `Radial Flare`) are genuinely required** is when you force the grid to match **true homologous perimeter anatomy** (dorsal fin base & tip to dorsal fin base & tip, anal fin to anal fin, without the 9 internal skull dots): there, a plain Box Stretch fails (**`2.4%`** match on the swept-back fin tips of the angelfish), and **Thompson’s curved grids jump by `+13%` to `+48%` (`15×` better)** while staying **100% smooth and un-folded on `6 out of 6` pairs** (whereas standard 600-point Thin-Plate Splines ([Bookstein 1989](https://doi.org/10.1109/34.24792)) fold and crumple the interior tissue grid on `6 out of 6` pairs!).
+
+![](https://fractastical.github.io/morpholib/guide_step3_box_cheat_vs_curves.png)
+
+*Plate XII — The box-stretch cheat versus true Thompson curves. On unlabeled shadows a plain rectangular stretch can fake a match by pushing a pufferfish belly up over a sunfish fin; on homologous perimeter anatomy (fin base to fin base, fin tip to fin tip) it collapses to a `2.4%` match while the shear, coaxial, fan, and flare grids gain `+13%` to `+48%` and stay un-folded. Provenance: [`thompson/generate_13yo_visual_guide.py`](https://github.com/fractastical/morpholib/blob/main/thompson/generate_13yo_visual_guide.py), step 3.*
+
+3. **Two rules, and a test we could not un-fail:**
+   Three datasets — 791, then 6,614, then 10,303 species, across 440 families — say the same thing twice.
+
+   **The body swims along Thompson’s sheet.** Over 150 million years, on the tree of [Rabosky et al. (2018)](https://doi.org/10.1038/s41586-018-0273-1), the direction of whole-body change lines up with his coordinate grids at **93.5% to 98.9%**. (This article uses two counts for those grids. Each 1917 family has 8 knobs; once position, rotation, and size are removed, the shared terms are the 7 modes used for prediction.) Alignment is not variance explained. On the standard 17-landmark data those 7 modes capture **57.9%** of between-species shape variance, against **86.2%** for a 7-component PCA trained on the data. Thompson’s grids are the highway, not the whole map, and 57.9% is the number every later clade is held to.
+
+   **The head eats.** Sister species on the same reef share a swimming body and rewire jaws, snouts, and eyes **3.3× to 9.3×** faster than their bodies.
+
+   **The blind test.** Before opening the Smithsonian’s FishShapes v1 — 5,180 species we had not seen — we locked the predictions in a [signed pre-registration](https://registry.knowledgepixels.com/np/RAilN4nbRHp3e3ZHqlbIgeL3UGPRV_KtQHtJD3gzVVdqw). Inside a 50-million-year horizon the Thompson predictor beat the clade consensus by a ridiculous margin (`p = 10⁻¹⁰³`). It did not beat the simplest rival. Copying the single nearest relative won by 4% on one noisy measurement (`p = 0.061`), and that hypothesis was rejected. The loss is in the same record as the win.
+
+![](https://fractastical.github.io/morpholib/guide_step4_outside_vs_inside_rules.png)
+
+*Plate XIII — The two rules of fish evolution. Outside: whole-body change across 440 families runs along Thompson’s coordinate highway (`93.5%–98.9%` alignment). Inside: sister species on the same reef rewire jaws, snouts, and eyes `3.3×–9.3×` faster than their bodies. Provenance: [`thompson/generate_13yo_visual_guide.py`](https://github.com/fractastical/morpholib/blob/main/thompson/generate_13yo_visual_guide.py), step 4; datasets of 791, 6,614, and 10,303 species.*
+
+> **Verdict — Claim 4: Strongly Confirmed.** Across 10,303 species, Thompson’s and Levin’s core thesis—that macroevolution navigates a low-dimensional geometric coordinate manifold while local modules handle rapid ecological competition—is a quantitative law of vertebrate evolution. ([Explore the interactive 10,303-Species Fish Morphospace on bioelectricity.net here!](https://bioelectricity.net/bioelectric-morphospace))
+
+---
+
+## Claims 5–7: Planarians, Picasso Tadpoles, and Xenobots
+
+How does a living tissue actually store a target coordinate in that morphospace? The three wet-lab results are valid. A planarian remembers a two-headed target after the drug is gone, a scrambled tadpole face moves back to the froglet layout along paths it was never shown, and a sculpted clump of frog skin assembles daughters with no change in its genome. What we built afterwards is a set of reduced equations, to test whether ordinary physics already produces those attractors. The equations reproduce them. The points held back below are about what the equations leave out, not about whether the experiments happened:
+
+### Claim 5 — Two-headed planarians remember a voltage ([Oviedo et al. 2010](https://doi.org/10.1016/j.ydbio.2009.12.012); [Beane et al. 2011](https://doi.org/10.1016/j.chembiol.2010.11.012); [Pietak & Levin 2016](https://doi.org/10.3389/fphys.2016.00178)/[2017](https://doi.org/10.1098/rsif.2017.0425); [Cervera, Levin & Mafé 2021](https://doi.org/10.1016/j.biosystems.2021.104511))
+
+Block the gap junctions of a cut planarian for 48 hours and it grows two heads. Wash the drug out, wait weeks, cut it again in plain water, and it grows two heads again. Nothing in the DNA changed.
+
+The model says why, in one sentence. Serotonin is positively charged, so it drifts down the voltage from head (`−20 mV`) to tail (`−60 mV`) by a fixed factor of about `4.7`, and gap-junction conductance cancels out of that equilibrium ([`the model`](https://github.com/fractastical/morpholib/blob/main/models/2017-pietak-bioelectric-morphogen-coupling.py)). Hold the block long enough for the tail to flip to a head voltage, wash the drug out, and the worm is stuck in the two-headed basin.
+
+![](https://fractastical.github.io/morpholib/media/cervera_levin_mafe_gj_block.gif)
+
+*Plate XIV — Two-headed memory in a reaction–diffusion planarian. Transient gap-junction blockade lets the posterior pole flip from `−60 mV` to `−20 mV`; after washout the symmetric* `2H0T` *state is a stable attractor that survives re-cutting with no genetic change. Provenance: [`models/2021-Cervera–Levin–Mafe.py`](https://github.com/fractastical/morpholib/blob/main/models/2021-Cervera–Levin–Mafe.py) (1D) and [`models/2017-pietak-bioelectric-morphogen-coupling.py`](https://github.com/fractastical/morpholib/blob/main/models/2017-pietak-bioelectric-morphogen-coupling.py) (405-cell 2D Voronoi Scharfetter–Gummel).*
+
+### Claim 6 — The Picasso tadpole finds its face ([Vandenberg, Adams & Levin 2012](https://doi.org/10.1016/j.ydbio.2012.01.027); [Pezzulo & Levin 2015](https://doi.org/10.1098/rsif.2015.0772); [Pinet et al. 2019](https://doi.org/10.1242/dev.175950))
+
+Scramble a tadpole’s eyes, jaws, and nostrils and they do not crawl back along the paths they were pushed. They take new paths, faster the further they are from home, until the froglet face is normal. Block the enzymes that remodel tissue and they stop. A model that treats the normal face as a target the tissue descends toward (`dV/dt ≤ 0`) reproduces both the correction and the arrest.
+
+![](https://fractastical.github.io/morpholib/media/xenopus_craniofacial_remap_growth.gif)
+
+*Plate XV — The Picasso tadpole corrects its face. Scrambled eyes, jaws, and nostrils move along novel paths with speed proportional to their defect until the normal froglet layout is reached (`dV/dt ≤ 0`), and arrest when MMP remodeling is blocked by GM6001. Provenance: [`models/2019-bischof-xenopus-craniofacial-remap.py`](https://github.com/fractastical/morpholib/blob/main/models/2019-bischof-xenopus-craniofacial-remap.py), Pass 3 with 3D Maxwell cartilage rods and MMP-2/MMP-9 zymography fields.*
+
+### Claim 7 — Xenobots copy themselves with no genetic change ([Kriegman, Blackiston, Levin & Bongard 2020](https://doi.org/10.1073/pnas.1910837117), [2021](https://doi.org/10.1073/pnas.2112672118))
+
+Sculpt ciliated frog-skin spheroids into a C-shaped mouth of about `105°` and set them in a dish of loose cells. The cilia make currents, the currents sweep cells into the mouth, adhesion sticks those cells into a daughter, and the daughter does it again — three or four generations. A sphere manages one. A flat slab manages none. Both of our fluid simulations find the same angle: `105°` is where capture works.
+
+![](https://fractastical.github.io/morpholib/media/xenobot_evolutionary_biobots_growth.gif)
+
+*Plate XVI — Kinematic self-replication. A C-shaped ciliated spheroid (`θ ≈ 105°`) sweeps dissociated cells into its mouth pocket until adhesion binds them into a motile daughter, for `3–4` generations versus `1` for a sphere and `0` for a slab. Provenance: [`models/2020-kriegman-xenobot-evolutionary-biobots.py`](https://github.com/fractastical/morpholib/blob/main/models/2020-kriegman-xenobot-evolutionary-biobots.py), 8×8×7 voxel lattice and 2D regularized-Stokeslet squirmer.*
+
+> **Verdict — Claims 5–7: The laboratory results are valid.** The two-headed memory, the face that finds itself, and the kinematic self-replication are published experimental results. A standard biophysical model reproduces each one, with a named attractor: bistable voltage memory, descent toward a target face, a hydrodynamic capture optimum. The `−2.2` points across the three are the gaps in our reduced equations. They are not a deduction against the experiments.
+
+---
