Skip to content

Wiki

WetLabEngine protein audit

White-box science contract at the GitHub root. Engine is a Supported tool. Custom protein-mechanism program CLOSED 19 Sep 2026 on current evidence. Folding results stay Restricted/Corrected.

Applied · Applied · Restricted · 2026-09-19 · compiled 2026-10-10 · 7 min read · 3 claims · 6 X sources

Original on X · Package: WetLabEngine (root) · JSON · Cite

protein · audit · biophysics · Villin · WetLabEngine

On this page

Causal stack

  1. 00 Astra / global consistency
  2. 01 Genesis Pin
  3. 02 Retained pruning-order history
  4. 03 Admissible slice
  5. 04 L1 (closed)
  6. 05 L2 (closed)
  7. 06 L3 hyperedge memory T_H + D_L3
  8. 07 L4 H₄ rank lift
  9. 08 Discrete atlas / scalar accessibility
  10. 09 GEM-bridge / gravity

19 Sep 2026 — program CLOSED on current evidence

Author correction: computational signals were real; a distinct transferable new folding mechanism was not established. Final held-out transfer +~2.6% (p≈0.46) / −5.6% (p≈0.49). 16 Sep README: Labs/p53 and Labs/MECP2 withdrawn — do not rely on them as validation. Reopen only for new independent evidence. Full note: /wiki/protein-closeout.

proteinfoldingengine/WetLabEngine is Aaron Spradlin’s public working tree. The repository root is a white-box protein-simulation audit engine. Nested under Labs/ are two GitHub-only coarse-grained campaigns (p53, MECP2). Nested under Tmp/TOE are the geometry-from-information packages this wiki already indexes, plus the April 2026 folding-side bridge packet on 1UAO and 1L2Y. Villin HP35 R_t is the X-posted controlled-case result. Those four protein objects are one research area only because they share an author, a science-contract style, and a GitHub tree — not because they share a proven law.

Do not mix the tool with the science

The engine can confirm or refute a numeric constraint in findings.json. That is a Supported tooling claim. p53, MECP2, Villin HP35, and frozen v9 on 1UAO/1L2Y are Restricted in-silico results. None of them is a folding oracle, a cancer or Rett therapy, AlphaFold refuted, or a TOE confirmation.

Repository map

Path What it is Wiki Badge
/ UnifiedAuditEngine.py v5.7, SCHEMA.md (findings.json 2.2), ManuscriptBuilder.py, Colab notebook protein-audit supported (tool)
Labs/p53 12-run p53 DBD campaign, Aug 2025. R248Q electrostatic vs Y220C kinetic trap. p53-campaign restricted
Labs/MECP2 19-run Rett MBD–DNA campaign, Aug 2025. InterfaceAnneal rescue + honest DEVIATION on Run 5. mecp2-rett restricted
Tmp/TOE/uqcf_bridge_v9_* Frozen v9 bridge on 1UAO and 1L2Y. Backbone-ordering, not full fold. Posted 12 Apr 2026. protein-bridge-v9 restricted
X, 14 Apr 2026 Villin HP35: identical snapshot, different fate from hidden R_t. Author tagged speculative. protein-audit restricted
Tmp/TOE/ Geometry-from-information program (atlas, holonomy, SPARC, batteries). Already indexed. program mixed
retained-atlas-agent/ Governor loop / constitution for atlas runs. Sibling tooling, not a folding result. c3-governor restricted

Start at the root README. Labs/ PNG and PDB files are Git LFS; markdown reports, CSVs, and findings.json are in the ordinary tree. Tmp/TOE is a separate program that happens to live in the same repo.

The engine

WetLabEngine is an open-source audit engine for protein simulation campaigns. A JSON manifest (findings.json) defines experiments, human-readable claims, and machine-testable constraints (for example best_final_RMSD_A ≤ 20.0). UnifiedAuditEngine.py (v5.7 in the current tree) checks those constraints against CSVs, PDBs, and images, then optionally sends artifacts to Gemini for qualitative review. Output is Biophysics_Final_Report.md plus machine-readable JSON. ManuscriptBuilder.py writes the prose body; the appendix is assembled deterministically from the verification report so the AI cannot rewrite the verdict table.

  • Science contract: SCHEMA.md, findings.json v2.2 (p53) / v3.3 (MECP2). Every metric in a constraint must be declared in metrics_vocabulary.
  • Deterministic verdicts: CONFIRMED, DEVIATION, INDETERMINATE. Tail-median RMSD from diagnostics_timeseries.csv is the preferred single source of truth.
  • AI peer review is qualitative and does not override the numeric verdict. Claim-integrity and cross-artifact triage are logged beside it.
  • Thesis block: count_runs_meeting aggregation across the campaign. The engine can fail a thesis even if individual runs pass.
  • License on the repo: CC BY-NC-SA 4.0, plus Spradlin’s later commercial-use restriction on the public research stream.

Minimal findings.json constraint (SCHEMA 2.2)

{
  "canonical_claim": {
    "type": "rescue_protocol",
    "statement": "Contact-anneal + PhiKan083 improves Y220C (RMSD ≤ 20 Å).",
    "constraints": [
      { "metric": "best_final_RMSD_A", "op": "<=", "value": 20.0 }
    ]
  }
}

Supported

As a tool: claims written into findings.json can be numerically confirmed or refuted against the bound raw artifacts. That is the engine’s job. It is not a folding oracle and not a TOE.

Run this first

GitHub is the working copy. This page is the cheapest entry. Start with Labs/MECP2, not Villin and not a new disease protein. The posted UnifiedAuditEngine.py is Colab-oriented: it hardcodes PROJECT_NAME = MECP2, clones ProteinFoldingEngine/Results-AI-Audit-Engine, looks for Projects/MECP2/findings.json, and exits if a Gemini key is missing. Do not expect python UnifiedAuditEngine.py to score Labs/MECP2 as a local folder.

Two entries, one contract

Full engine (his README): Colab + API_KEY.txt + Gemini qualitative review. Deterministic half (this atlas): a stdlib script that reads findings.json and tail-median RMSD from diagnostics_timeseries.csv. PNG/PDB Git LFS files are not required. This compile ran the deterministic half on 7 Sep 2026. Gemini peer review was not rerun.

  1. Clone only Labs/MECP2. Skip git lfs pull — CSVs and findings.json are ordinary files.
  2. Download wetlab_deterministic_audit.py from this page (Open source). It is a pedagogical reconstruction of verdict_from_constraint, not a dump of UnifiedAuditEngine.py.
  3. python3 wetlab_deterministic_audit.py Labs/MECP2
  4. Green: CONFIRMED 15, DEVIATION 1, INDETERMINATE 3, and smoke PASS on Run 5.
  5. Optional: python3 wetlab_deterministic_audit.py Labs/p53 --no-smoke. Expect Run 9 INDETERMINATE if the ContactAnneal folder is missing.
  6. To run his full engine, follow the repo README in Colab with a Gemini key. That path clones a different data repo.

Labs/MECP2 deterministic smoke

Open source

git clone --depth 1 --filter=blob:none --sparse https://github.com/proteinfoldingengine/WetLabEngine.git
cd WetLabEngine
git sparse-checkout set Labs/MECP2
# put wetlab_deterministic_audit.py in this directory (download from this page), then:
python3 wetlab_deterministic_audit.py Labs/MECP2

Expected MECP2 output (this compile, 7 Sep 2026)

CONFIRMED 15   DEVIATION 1   INDETERMINATE 3
thesis count_runs_meeting (best_final_RMSD_A <= 4.5): 16 >= 3  PASS
smoke PASS: Run 5 is DEVIATION (tail-median RMSD 3.06 Å vs contracted ≥ 20 Å).
Campaign This compile Load-bearing cell What a green run does not show
Labs/MECP2 15 CONFIRMED / 1 DEVIATION / 3 INDETERMINATE Run 5 DEVIATION at 3.06 Å against ≥ 20 Å Gemini rerun; Rett therapy; native folding
Labs/p53 11 CONFIRMED / 0 DEVIATION / 1 INDETERMINATE Run 9 INDETERMINATE (ContactAnneal folder missing) The manuscript’s 18.9 Å rescue independently re-read

Stdlib tail-median of diagnostics_timeseries.csv versus findings.json constraints. MECP2 matches the campaign abstract. p53 matches Table 1 in order (WT/R248Q 26.71 vs 26.6 Å; native-start 3.82 vs 4.0 Å) except Run 9, which this tree cannot score.

  • Run 5 is the smoke test that the engine can fail a claim. If it comes back CONFIRMED, stop and inspect — do not drop the DEVIATION.
  • Thesis count_runs_meeting can pass while a run-level claim fails: Run 5’s 3.06 Å still counts toward RMSD ≤ 4.5.
  • dataset_root in findings.json is a Colab path (/content/Results-AI-Audit-Engine/Projects/MECP2). The smoke script ignores it and resolves source_folder under the campaign directory you cloned.
  • Live /labs is a recoverability toy. It is not this auditor.

Supported

This compile independently scored Labs/MECP2/findings.json against the bound diagnostics CSVs: 15 CONFIRMED, 1 DEVIATION (Run 5, 3.06 Å vs ≥ 20 Å), 3 INDETERMINATE. Thesis count_runs_meeting passed (16 ≥ 3). That is a tooling reproduction. It is not a Gemini rerun, not a Rett therapy, and not native folding.

Villin HP35 — the X result

On 14 Apr 2026 Aaron asked Grok to peer-review a Villin HP35 claim from the WetLabEngine pipeline. The posted numbers: two chains with identical visible conformation can still have different folding fates because of hidden retained history R_t. G_t-only models miss it. Grok’s in-thread review treated the pipeline as working on controlled/sample cases. A 26 Apr follow-up used stronger AlphaFold language; this atlas keeps the 14 Apr speculative tag as the governing voice.

The ROC lift (AUC G_t + R_t ≈ 0.9569 vs G_t alone) shows real predictive signal from R_t. The matched-state divergence (50.2 % far-R_t vs 1.3 % close-R_t, +48.9 % spread) directly falsifies pure visible-state completeness.

@grok, 14 Apr 2026

This is all speculative and they have better things to do for now then react to another one of the wild claims.

@aaronspradlin, 14 Apr 2026

Restricted

On the posted Villin HP35 controlled cases, adding retained state R_t lifts AUC and opens a matched-pair divergence gap versus G_t-only. This is not a claim that AlphaFold is refuted, and it is not a folding oracle.

Source on X

X ledger versus GitHub-only labs

Date Object Where it lives What to cite
22–23 Aug 2025 p53 DBD campaign (12 runs) GitHub Labs/p53 — no load-bearing X field note in this compile p53-campaign + findings.json
26–28 Aug 2025 MECP2 MBD–DNA campaign (19 runs) GitHub Labs/MECP2 — no load-bearing X field note in this compile mecp2-rett + findings.json
12 Apr 2026 Frozen v9 on 1UAO / 1L2Y X thread + Tmp/TOE/uqcf_bridge_v9_* protein-bridge-v9
14 Apr 2026 Villin HP35 R_t X, tagged speculative the same day this page
26 Apr 2026 AlphaFold-never-solved restatement X reply quoting the 14 Apr numbers restricted restatement, not a promotion
1 May 2026 Critical-path recurrence including protein X applied-stack — recurrence, not a unification proof
10 May 2026 Work stays on X/GitHub; others free to use protein folding X open-science / method

X is the archive of record for posted results. GitHub Labs/ campaigns that never became field notes are still part of the working tree and must be indexed honestly, not promoted to X-posted science.

Same pattern, other domains

The 1 May 2026 critical-path post treats protein folding, fractal TSP, SPARC holdouts, and live project plans as rhyming recoverability problems: preserve what remained coherent, deform what is stressed, reset only what went stale. Recurrence is not a unification proof. The 10 May question — what unifies proteins, RCLC, critical path, and TSP — is still open. The C3++ governor is the LLM-side instrument of the same idea. The applied recoverability stack is the hub for those three Python labs.

X sources

Original posts this article was compiled from. The notebook copy is editorial; the status URL is the source.

Previous Protein program closeoutNext MECP2 Rett MBD–DNA campaign

Causal stack

  1. 00 Astra / global consistency
  2. 01 Genesis Pin
  3. 02 Retained pruning-order history
  4. 03 Admissible slice
  5. 04 L1 (closed)
  6. 05 L2 (closed)
  7. 06 L3 hyperedge memory T_H + D_L3
  8. 07 L4 H₄ rank lift
  9. 08 Discrete atlas / scalar accessibility
  10. 09 GEM-bridge / gravity

On this page

X sources

Related articles

GitHub packages

Field notes