docs(transgenerational): M6.13 campaign close-out — STOP verdict + GC fix + logbook 020 + archive - #169
Merged
Merged
Conversation
- add-tei-prior-on-m3/tasks.md: drop the previously-planned § 6.10 (archive) and § 6.11 (push) tasks. Both would self-block the archive precondition (`openspec-archive-change` requires every task ticked) and the push step is user-sign-off gated per `feedback_ask_before_push_and_pr.md`. Inlined a short note clarifying that both operator-side steps happen after § 6.9 is ticked. - phase5-tracking/tasks.md M6.12: tick with a verdict-bearing description. M6.12 (`transgenerational+weights`) was NOT TRIGGERED per the PR-A null-floor pre-registered criterion and is now superseded by M6.13 (the wet-lab-aligned reframe of the same composed-inheritance mechanism). Matches the M4.6.5 / M5.9 / M6.8 precedent where ticks mean "decision recorded and propagated", not "experiment ran". Unblocks the eventual M8.7 phase5-tracking archive without leaving a stale [ ] M6.12 row behind. - Forward-pointer note on M6.12 → M6.13 added so a reader of the M6 thread doesn't read the NOT TRIGGERED row as the question being dropped — it was reframed, not abandoned. `openspec validate add-tei-prior-on-m3 --strict` clean. `openspec validate phase5-tracking --strict` clean. Docs-only — SKIP=tests on pre-commit since no Python changed. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Calibration smoke pass 1 (2026-05-20T11:56:58Z, single seed=42 × pop 6 × 2 gens × weights_only arm only, ~10m 28s wall) measured: - F0 mean survival_rate = 0.407 (pop 6: 0.04 / 0.44 / 0.72 / 0.24 / 0.52 / 0.48) - F1 mean survival_rate = 0.600 (pop 6: 0.44 / 0.48 / 0.64 / 0.80 / 0.56 / 0.68) - Lamarckian lift F0 → F1 = +19.3pp in one generation at K=1000 Tripwire verdict (the two the smoke can measure on weights_only): - T1' F0 envelope: 0.30 ≤ 0.407 ≤ 0.70 → PASS - T3' high-end: F1=0.600 > 0.95×F0=0.387 → PASS (no saturation) - T3' low-end: F1=0.600 ≥ 0.80×F0=0.326 → PASS (above floor) T3' beats-control leg (F1 ≥ 1.2×control_F1) is structurally deferred to pilot time per design.md § D5 — the smoke runs weights_only only. YAML changes: - tei_weights.yml: lawn_schedule F1+ entry annotation 'K_test placeholder' → 'K_test (calibration-locked 2026-05-20)'; header note + transgenerational block comment cite the smoke session ID and the observed F0/F1 numbers. - weights_only.yml + control.yml: header + inline learn_episodes_per_eval comment note that K=1000 was calibration-locked on 2026-05-20. The numeric K=1000 is unchanged across all three arms — this is a comment-only commit that converts "placeholder" to "locked-in" so future readers see why K=1000 is the campaign value. Launcher parity check still passes (tei_weights F1+ K=[1000] == weights_only learn=1000 == control learn=1000). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…composed mode
The M6.13 pilot (1 seed × pop 8 × 4 gens × tei_weights arm) crashed
at F2 dispatch on 2026-05-20 with FileNotFoundError on the F0
substrate path. F0 + F1 had completed cleanly; the gen-000/ directory
was empty when F2 children tried to load F0's substrate for the
decay-cascade application.
Root cause at loop.py:_gc_inheritance_dir (pre-existing on main from
M3's initial Lamarckian GC implementation):
for path in gen_dir.glob("genome-*.pt"):
gid = path.stem.removeprefix("genome-")
if gid not in keep:
path.unlink(missing_ok=True)
Path.glob("genome-*.pt") matches BOTH weight checkpoints (genome-X.pt)
AND substrate files (genome-X.tei.pt). Path.stem strips only ONE
suffix, so for a genome-X.tei.pt file the extracted gid is X.tei
(not X), never matches keep_ids, and the substrate gets deleted.
Why this only surfaced now:
- M3 (lamarckian): only .pt files in the GC'd directory; .tei never
appears.
- M6.9+ (transgenerational): _inheritance_active() returns False, so
this main-loop GC doesn't fire. Substrate cleanup is owned by the
inline GC inside _run_f0_substrate_extraction which knows about
.tei.pt.
- M6.13 (weights+transgenerational, composed): both file types coexist
AND the main-loop GC fires. Bug surfaces immediately at end of F0
on the very first cleanup pass.
The cascade case (F2/F3 children inheriting F0's substrate with
decay_factor^2/decay_factor^3) makes the impact load-bearing: the
substrate must survive every main-loop GC pass for the whole campaign,
not just the F0→F1 transition.
Fix: skip .tei.pt entries in the main-loop GC entirely. Substrate
file lifecycle is owned by the F0 substrate-extraction pipeline (the
inline GC at _run_f0_substrate_extraction is already correctly
skipped under composed mode via _combined_inheritance_active()); the
main-loop GC now matches that contract.
Regression test: test_main_loop_gc_preserves_substrate_tei_pt seeds
a gen-000/ with two .pt files + one .tei.pt, calls
_gc_inheritance_dir(generation=0, keep_ids=['elite']), and asserts
the elite .pt survives, the non-elite .pt is GC'd, and the
substrate .tei.pt survives unconditionally.
Verification:
- 469 evolution tests pass (no regressions on M3 / Baldwin / M6.9+
pure-TEI / M6.13 composed paths).
- 11 m613-smoke tests pass including the new GC regression.
- ruff + pyright clean.
Pilot will be re-dispatched after this commit lands.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…erdict
M6.13 campaign-execution closes. Four pilots at K ∈ {1000, 500, 200,
200-F0-matched} all rule against the substrate-accelerates-Lamarckian-
retraining hypothesis. Under fair-F0 comparison, cross-arm
`tei_weights − weights_only` F1-F3 delta is +0.00pp at K=1000
(substrate inert) and −9.33pp at K=200 (substrate INTERFERES). The
K-sensitivity sweep apparent dose-response (Δ 0.00→4.00→5.33pp as K
decreased) was the F0 confound — tei_weights F0 held at K=2000 across
the sweep while weights_only/control F0 matched the swept K.
Disambiguation pilot inverted the sign under fair F0.
Logbook 020: docs/experiments/logbooks/020-tei-prior-on-lamarckian.md
(renamed from 020-tei-prior-on-m3.md to match the other logbooks'
no-milestone-label naming convention — 013/018/019 don't carry M3/M6
prefixes either). Documents the four pilots' retention tables, the
methodological audit of the broken post-hoc per-gen-eval script, the
mechanism diagnosis (action-distribution-bias substrate vs sensory-
excitability shift mismatch), the adversarial-critique findings, and
a substantive future-directions section for the three substrate-
redesign ideas (B/C/D) flagged for a possible Phase 6 quantum re-
evaluation arc. M6.14 frequency-prior ablation NOT triggered per
pre-registered criterion.
Artefacts stashed under artifacts/logbooks/020-tei-prior-on-lamarckian/
following the logbook 019 convention — only the aggregator outputs
per pilot (pilot_pivot_decision.md / retention_table.csv /
summary.md / decision_gate.csv / cross_arm_verdict.csv) plus a smoke
summary. 21 files, 84 KB total, no .pt weights or full diagnostics.
No absolute home paths; LFS rules cover the CSVs at no overhead.
phase5-tracking M6.13 row ticked [x] with the STOP verdict +
mechanism diagnosis + M6 thread-close framing. docs/roadmap.md gains
a new M6.13 row between M6.9+ and M7.
add-tei-prior-on-m3/tasks.md execution tasks 6.1-6.9 all ticked off
with verdict-bearing descriptions matching the M4.6.5 / M5.9 / M6.8
"checkbox = decision recorded and propagated" precedent. OpenSpec
change validates clean (`openspec validate add-tei-prior-on-m3
--strict`); phase5-tracking validates clean.
Docs-only commit; pre-commit hooks clean (SKIP=tests since no Python
changed).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The OpenSpec change drafted § D3 documenting only the inline-GC suppression at _run_f0_substrate_extraction (loop.py:~863-866). That covered the gen-0 substrate-extraction tail path, but missed the SECOND GC site — the main-loop GC at loop.py:~1649 that runs end-of-each-gen via _gc_inheritance_dir. Under composed mode, F0's gen-000/ holds both the elite's weights .pt AND its substrate .tei.pt. The main-loop GC globbed "genome-*.pt", which matched both files; Path.stem strips only one suffix so the extracted gid for genome-<elite>.tei.pt was "genome-<elite>.tei" (with the trailing .tei still attached), which never matches the elite-id keep-set, and the substrate got deleted. This was a pre-existing bug from M3's initial GC implementation but only surfaced under composed mode because M3-only never produces .tei.pt files and pure-TEI's main-loop GC doesn't fire (because _inheritance_active() is False under kind=transgenerational). The bug was caught + fixed during campaign execution at commit b78cdbb (logbook 020 § audit; tasks.md § 6.4 acknowledgment) with the test_main_loop_gc_preserves_substrate_tei_pt regression test. This commit adds the missing § D3.5 paragraph to design.md so the spec accurately reflects the as-shipped contract. `openspec validate add-tei-prior-on-m3 --strict` clean. Docs-only commit; SKIP=tests since no Python changed. Other drift-audit findings (verdict claims, test counts, K_test contract, M6.14 trigger conditions, per-gen-eval bridge) all OK as-is — the forward-looking spec correctly pre-declared both GO and STOP outcomes and the framework shipped as-spec'd. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Archived to openspec/changes/archive/2026-05-21-add-tei-prior-on-m3/ via `git mv` so history is preserved. All 47 tasks ticked; all 4 artefacts (proposal / design / specs / tasks) complete per `openspec status --json`. Spec sync intentionally SKIPPED per the M6.9+ STOP-archive precedent (`2026-05-19-add-transgenerational-memory-redesign` also archived without main-spec promotion). The delta specs describe an action-logit-bias substrate contract that M6.13's four pilots decisively falsified; promoting them to main `evolution-framework/` and `configuration-system/` specs would commit the repo to a known-null architectural choice. Logbook 020 + this archived change directory remain the canonical reference for the framework scenarios + the STOP verdict + the future-work substrate-redesign directions (Ideas B/C/D for a possible Phase 6 quantum re-evaluation arc). After this commit, the only active OpenSpec change is `phase5-tracking` (the long-running Phase 5 milestone tracker; will archive at M8.7 alongside the M8 synthesis change). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
… stale archive refs Independent pre-push review of the M6.13 close-out branch surfaced three issues worth fixing before remote-state mutation: F1 (should-fix, latent footgun): coevolution.py:2186-2189 has the same Path.stem suffix-stripping anti-pattern that b78cdbb fixed in EvolutionLoop._gc_inheritance_dir. Coevolution currently never writes .tei.pt files (no per-side substrate machinery), but a future composed-coevolution arc would silently re-introduce the bug. Added the same .name.endswith(".tei.pt") guard preemptively with a comment cross-referencing the loop.py fix. F2 (should-fix, project-convention violation): the new GC regression test in test_loop_m613_smoke.py:486-487 introduced fresh "M3 / pure-TEI" milestone labels in the docstring. Commit 2371d45 on the parent branch had just scrubbed exactly these refs from the same file. Replaced with behavioural names "lamarckian-only / pure-transgenerational" per `feedback_no_milestone_refs_in_code`. F3 (note, post-archive cleanup): three docstring/comment refs to `openspec/changes/add-tei-prior-on-m3/` (no archive/ prefix) predated this branch but pointed to a nonexistent path after the archive commit (99600f5) moved the directory to `openspec/changes/archive/2026-05-21-add-tei-prior-on-m3/`. Updated the refs in `evolution/lamarckian_transgenerational_inheritance.py`, `evolution/inheritance.py`, and the `test_aggregate_m613_pilot.py` module docstring to point at the archive location. Verification: - 469 evolution tests pass (no regressions from coevolution.py defensive guard). - 71 touched tests pass under default marker selection. - ruff + ruff-format + pyright clean on all 5 touched files. - No remaining non-archive `openspec/changes/add-tei-prior-on-m3/` refs in the tree. - No M3/PR-A/PR-B milestone labels in test_loop_m613_smoke.py or coevolution.py's new lines (pre-existing M3 mentions at coevolution.py:162 + 2720 predate this branch). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Run ID: 📒 Files selected for processing (2)
✅ Files skipped from review due to trivial changes (1)
📝 WalkthroughWalkthroughThis PR closes the M6.13 experiment (TEI-as-prior-on-Lamarckian reframe) with a STOP outcome, delivering a GC substrate-preservation fix, four pilot runs with decision records, configuration documentation updates, and a comprehensive logbook with mechanistic diagnosis of the null result. ChangesM6.13 Experiment Delivery and GC Fix
🎯 3 (Moderate) | ⏱️ ~20 minutes Possibly Related PRs
Poem
🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches🧪 Generate unit tests (beta)
Comment |
chrisjz
marked this pull request as ready for review
May 21, 2026 10:59
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 2
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In
`@artifacts/logbooks/020-tei-prior-on-lamarckian/pilot_k200/pilot_pivot_decision.md`:
- Around line 5-13: The decision text ("Pilot signal: clean GO at K_test" and
"Pivot: NONE") conflicts with the observed metrics (`tei_weights` per-arm
verdict: STOP, per-arm gate pass: False, Wilcoxon p-threshold met: False,
cross-arm mean delta +5.33pp); update the narrative to match the metrics by
changing the pilot outcome and pivot statement to reflect a STOP/REJECT decision
(or alternately adjust the metrics to show GO/Pass if that was intended), and
ensure the artifact uses the same identifiers (`tei_weights`, `weights_only`,
"Per-arm verdict", "Per-arm gate pass", "Wilcoxon p-threshold met", and "Pivot")
so the reported decision and listed metrics are consistent.
In
`@artifacts/logbooks/020-tei-prior-on-lamarckian/smoke/calibration_smoke_K1000.md`:
- Line 7: The T3' Lamarckian-headroom line is inconsistent: with F1=0.600 and
thresholds 0.387/0.326 the condition "F1 ≤ 0.95×F0 AND F1 ≥ 0.80×F0" cannot be
true, so update the record for T3' Lamarckian-headroom to be self-consistent by
either flipping the PASS to FAIL or adjusting the inequality to match the
intended check; specifically, for the existing numeric values (F1=0.600;
thresholds 0.387/0.326) change the "PASS" to "FAIL" for T3' or rewrite the
condition so that it correctly reflects that F1 is greater than both thresholds.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro
Run ID: c72a6e42-a908-4dc9-b2f4-a652f892c24d
⛔ Files ignored due to path filters (12)
artifacts/logbooks/020-tei-prior-on-lamarckian/pilot_k1000/cross_arm_verdict.csvis excluded by!**/*.csvartifacts/logbooks/020-tei-prior-on-lamarckian/pilot_k1000/decision_gate.csvis excluded by!**/*.csvartifacts/logbooks/020-tei-prior-on-lamarckian/pilot_k1000/retention_table.csvis excluded by!**/*.csvartifacts/logbooks/020-tei-prior-on-lamarckian/pilot_k200/cross_arm_verdict.csvis excluded by!**/*.csvartifacts/logbooks/020-tei-prior-on-lamarckian/pilot_k200/decision_gate.csvis excluded by!**/*.csvartifacts/logbooks/020-tei-prior-on-lamarckian/pilot_k200/retention_table.csvis excluded by!**/*.csvartifacts/logbooks/020-tei-prior-on-lamarckian/pilot_k200_f0matched/cross_arm_verdict.csvis excluded by!**/*.csvartifacts/logbooks/020-tei-prior-on-lamarckian/pilot_k200_f0matched/decision_gate.csvis excluded by!**/*.csvartifacts/logbooks/020-tei-prior-on-lamarckian/pilot_k200_f0matched/retention_table.csvis excluded by!**/*.csvartifacts/logbooks/020-tei-prior-on-lamarckian/pilot_k500/cross_arm_verdict.csvis excluded by!**/*.csvartifacts/logbooks/020-tei-prior-on-lamarckian/pilot_k500/decision_gate.csvis excluded by!**/*.csvartifacts/logbooks/020-tei-prior-on-lamarckian/pilot_k500/retention_table.csvis excluded by!**/*.csv
📒 Files selected for processing (27)
artifacts/logbooks/020-tei-prior-on-lamarckian/pilot_k1000/pilot_pivot_decision.mdartifacts/logbooks/020-tei-prior-on-lamarckian/pilot_k1000/summary.mdartifacts/logbooks/020-tei-prior-on-lamarckian/pilot_k200/pilot_pivot_decision.mdartifacts/logbooks/020-tei-prior-on-lamarckian/pilot_k200/summary.mdartifacts/logbooks/020-tei-prior-on-lamarckian/pilot_k200_f0matched/pilot_pivot_decision.mdartifacts/logbooks/020-tei-prior-on-lamarckian/pilot_k200_f0matched/summary.mdartifacts/logbooks/020-tei-prior-on-lamarckian/pilot_k500/pilot_pivot_decision.mdartifacts/logbooks/020-tei-prior-on-lamarckian/pilot_k500/summary.mdartifacts/logbooks/020-tei-prior-on-lamarckian/smoke/calibration_smoke_K1000.mdconfigs/evolution/tei_prior_m613_control.ymlconfigs/evolution/tei_prior_m613_tei_weights.ymlconfigs/evolution/tei_prior_m613_weights_only.ymldocs/experiments/logbooks/020-tei-prior-on-lamarckian.mddocs/roadmap.mdopenspec/changes/archive/2026-05-21-add-tei-prior-on-m3/.openspec.yamlopenspec/changes/archive/2026-05-21-add-tei-prior-on-m3/design.mdopenspec/changes/archive/2026-05-21-add-tei-prior-on-m3/proposal.mdopenspec/changes/archive/2026-05-21-add-tei-prior-on-m3/specs/configuration-system/spec.mdopenspec/changes/archive/2026-05-21-add-tei-prior-on-m3/specs/evolution-framework/spec.mdopenspec/changes/archive/2026-05-21-add-tei-prior-on-m3/tasks.mdopenspec/changes/phase5-tracking/tasks.mdpackages/quantum-nematode/quantumnematode/evolution/coevolution.pypackages/quantum-nematode/quantumnematode/evolution/inheritance.pypackages/quantum-nematode/quantumnematode/evolution/lamarckian_transgenerational_inheritance.pypackages/quantum-nematode/quantumnematode/evolution/loop.pypackages/quantum-nematode/tests/quantumnematode_tests/campaigns/test_aggregate_m613_pilot.pypackages/quantum-nematode/tests/quantumnematode_tests/evolution/test_loop_m613_smoke.py
Comment on lines
+5
to
+13
| **Pilot signal: clean GO at K_test (D6 row 2).** `tei_weights > weights_only` by ≥ 5pp — substrate prior accelerates Lamarckian retraining at K_test. The substrate-accelerates-retraining hypothesis is supported. | ||
| Pivot: NONE. Proceed to full campaign at K_test only (no K sweep — the test-point selection was made at calibration time). | ||
|
|
||
| Observed metrics: | ||
|
|
||
| - `tei_weights` per-arm verdict: **STOP** | ||
| - Cross-arm mean delta (tei_weights - weights_only): +5.33pp | ||
| - Per-arm gate pass (tei_weights): False | ||
| - Wilcoxon p-threshold met: False |
There was a problem hiding this comment.
Align the pivot narrative with the listed observed metrics.
The "clean GO / Pivot: NONE" statement conflicts with the recorded metrics (STOP, gate False, Wilcoxon False).
Please update the decision text (or the metrics) so this artifact reports a single consistent outcome.
🧰 Tools
🪛 LanguageTool
[grammar] ~5-~5: Ensure spelling is correct
Context: ...2).** tei_weights > weights_only by ≥ 5pp — substrate prior accelerates Lamarckian retraining...
(QB_NEW_EN_ORTHOGRAPHY_ERROR_IDS_1)
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In
`@artifacts/logbooks/020-tei-prior-on-lamarckian/pilot_k200/pilot_pivot_decision.md`
around lines 5 - 13, The decision text ("Pilot signal: clean GO at K_test" and
"Pivot: NONE") conflicts with the observed metrics (`tei_weights` per-arm
verdict: STOP, per-arm gate pass: False, Wilcoxon p-threshold met: False,
cross-arm mean delta +5.33pp); update the narrative to match the metrics by
changing the pilot outcome and pivot statement to reflect a STOP/REJECT decision
(or alternately adjust the metrics to show GO/Pass if that was intended), and
ensure the artifact uses the same identifiers (`tei_weights`, `weights_only`,
"Per-arm verdict", "Per-arm gate pass", "Wilcoxon p-threshold met", and "Pivot")
so the reported decision and listed metrics are consistent.
Two findings from PR #169 inline review: (1) [FIX] smoke/calibration_smoke_K1000.md T3' line was internally inconsistent: it claimed "F1 ≤ 0.95×F0 AND F1 ≥ 0.80×F0: PASS" but with F1=0.600 and 0.95×F0=0.387, the first conjunct is False and the conjunction cannot be True. The PASS verdict itself is correct (Lamarckian grew F1 from 0.407 to 0.600, well past F0+0.05=0.457 — design.md § D5's "clear headroom" PASS clause). The wrong text was a tripwire-table-style summary glommed onto a procedural-rule outcome. Rewrote the line to cite the actual procedural PASS trigger (F1 > F0 + 0.05) and note both bands were cleared in the upward direction. (2) [DECLINE] pilot_k200/pilot_pivot_decision.md narrative ("clean GO ... proceed to full campaign") superficially conflicts with its own observed-metrics list (per-arm verdict STOP, gate False, Wilcoxon False). The file is a verbatim capture of aggregate_m613_pilot.py output, not hand-written narrative — editing it would falsify the forensic record showing how the F0-confound was discovered (K-sweep apparent GO → disambiguation flipped sign). The two-channel structure (pivot-row classification vs full GO-gate metric list) is a deliberate aggregator feature so operators see exactly which signal a pivot decision is keyed to. Mitigation for (2): added a README at the artefacts stash root explaining the two-channel reporting structure, why these files are forensic captures rather than narrative documents, and why the F0-confound disambiguation pilot supersedes the K-sweep pivot-row classifications for M6.13's actual verdict. Future readers see the reconciliation guidance immediately on entering the artefact dir. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Codecov Report❌ Patch coverage is
📢 Thoughts on this report? Let us know! |
6 tasks
chrisjz
changed the title
chrisjz
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Closes the M6.13 (TEI-as-prior-on-Lamarckian) milestone with a decisive STOP verdict. Combined with the M6.9+ STOP on pure-TEI K=0, the M6 transgenerational-memory thread closes coherently — TEI doesn't transfer in either form on this LSTMPPO substrate. Logbook 020 captures the verdict + future-directions for a possible Phase 6 quantum re-evaluation arc.
tei_weights − weights_onlyF1-F3 delta is +0.00pp at K=1000 (substrate inert) and −9.33pp at K=200 (substrate INTERFERES) under fair-F0 comparison. The K-sweep apparent dose-response (+0.00 → +4.00 → +5.33pp) was the F0 confound — disambiguation inverted the sign. Mechanism diagnosis: action-distribution-bias substrate is the wrong abstraction for the wet-lab single-circuit excitability shift.loop.py:_gc_inheritance_dirwas deleting F0 substrate.tei.ptfiles becausePath.glob("genome-*.pt")matched substrate files too, andPath.stemstrips only one suffix so the extracted gid ended in.tei. Pre-existing bug from M3's initial Lamarckian GC implementation; only surfaced under composed mode. Fix + regression test inb78cdbb5. Preemptive sibling guard added tocoevolution.py:_gc_inheritance_diragainst future composed-coevolution arcs.artifacts/logbooks/020-tei-prior-on-lamarckian/.phase5-trackingM6.13 row ticked with STOP verdict;docs/roadmap.mdgains a new M6.13 row.add-tei-prior-on-m3moved toarchive/2026-05-21-add-tei-prior-on-m3/viagit mv(history preserved). Spec sync intentionally skipped per the M6.9+ STOP-archive precedent — the delta specs describe an architectural contract M6.13 falsified; promoting them would commit the repo to a known-null choice.Commits
e7223c1521919c8bb78cdbb504ede5582c4744d699600f5531378d51Test plan
not smoke and not nightlymarker selectionphase5-trackingvalidates ✓openspec/changes/add-tei-prior-on-m3/(non-archive) refs remain in production code or tests/tmp/build_pilot_per_gen_from_eval_diag.pyintentionally NOT committed (logbook 020 explains)Next milestone
Per discussion: skip both the proposed M6.14 substrate-extraction diagnostic AND the M7 NEAT optional milestone. Go directly to M8 (Phase 5 synthesis logbook) as the next milestone — synthesise M2/M3 GO + M4/M5/M6 STOPs with substrate-grounded diagnoses, cross-cite 2024-2026 literature (Resendez Prado, Mougi, Chen, Kaletsky, mammalian-TEI), and recommend Phase 6 priorities. M7 NEAT remains OPTIONAL in the tracker (architecture-symmetry hypothesis deferred to Phase 6 quantum re-evaluation by other means).
🤖 Generated with Claude Code
Summary by CodeRabbit
Bug Fixes
Documentation
Tests