feat: real-worm behavioural-thermotaxis validation (T7, second modality â honest PARTIAL) - #269
Merged
Merged
Conversation
A second, non-gating real-worm behavioural-validation arm folded into Phase 6a before the T9a synthesis: extend the Logbook-035 chemotaxis validation to thermotaxis (a MUST behaviour) so the biologically-validated-substrate claim spans modalities. Reuses the 035 capture/metrics/grading/harness with a homeostatic-setpoint covariate (thermal error toward the comfort band). MLP primary + connectome companion, n>=8, behaviour-level reference (Ryu & Samuel 2002; thermal klinotaxis). Direct doc task under phase6-tracking, no new OpenSpec change. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
âĶ(capture + reference) Author the OpenSpec change (course-corrected from the initial "no change" call: this extends the archived realworm-behavioural-validation capability with a new capture modality + thermal reference â real capability code â so it warrants a reviewed change). Self-reviewed via nematode-review-spec; fixed the citation (Luo et al. 2014 primary), the D5 specificity-control claim (thermotaxis_mode: derivative IS the analogue of 035's food-derivative control), and the tracker note (now points at this change, sign-only references). §1 capture modality (byte-identical when food): - SensingConfig.capture_behaviour_modality: food | thermotaxis (default food). - Agent _behaviour_capture_fields helper: food returns today's values; thermotaxis returns the homeostatic setpoint drive -|T - Tc|, its one-step derivative, and the live toward-comfort direction â so the existing bias-curve metrics apply unchanged. Tests: food byte-identical; thermotaxis records the setpoint drive + live thermal gradient. §2 thermal reference + modality-aware loading + harness flag: - data/thermotaxis/behavioural_bias_signatures.json (four sign-only statistics; Luo 2014 / Ryu & Samuel 2002 / Clark 2007) + _default_bias_signatures_thermotaxis fallback + load_bias_signatures(modality=...); explicit missing path still raises. - Harness --modality food|thermotaxis grades against the modality reference set + records the modality in the summary. Tests for the thermotaxis reference set. Remaining (under the change): §3 thermotaxis-dominant config â §4 smoke (PAUSE) â §5 panels â §6 logbook 036. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Audit of the thermotaxis eval (verifying no bugs before the panel) exposed a latent bug in the committed curving-rate floor: `suggest_min_path_len` uses 0.25x the median stride, which collapses toward 0 when the worm PARKS (a converged thermotaxis worm dwelling at its comfort target â median stride 0.008mm, 74% of steps <0.1mm). The collapsed floor let the dtheta/path_len creep artifact through and inflated the weathervane wildly (a smoke read +12.8, ~99% artifact; the true slope is ~+0.2-0.5). Fall back to the median *moving* stride (strides above 0.1x the max) only when the raw median has collapsed below that fraction of the max â so a continuously-moving worm (e.g. all of 035's foraging arms) keeps the raw-median floor byte-identical (verified: the three 035 panel JSONs regenerate with unchanged values), while a parked worm floors at the true stride scale. 035 wasn't affected in-flight because its foraging worm moves continuously (median stride ~1mm); thermotaxis exposed the fragility. Regression test added. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
âĶIAL) Second real-worm validation modality (non-gating). On the faithful linear-gradient Tc-seeking assay (MLP, n=4): the thermotaxis WEATHERVANE is reproduced weakly but robustly (theta-free slope +0.020, 95% CI [+0.010,+0.031], all seeds positive, ~5x weaker than chemotaxis) and a derivative-sensing specificity control confirms it is sensor-driven (collapses to +0.000). Thermal KLINOKINESIS is ABSENT in both sensing modes â the substrate reproduces the spatial-steering half of thermotaxis but not the biased-random-walk, because the continuous-Gaussian head has a state-independent std (can't stochastically random-walk; sharpens 035 too) and the RL worm migrates-and-parks rather than isothermal-tracks (temperature isn't consumable -> static endpoint). A genuine behavioural difference motivating Phase 7; chemotaxis (035) remains the strong validation of record. Logbook 036 + supporting artefacts (2 per-arm JSONs + 2 figures), README row, T7.validation.thermotaxis ticked, the two validation configs, the capture_behaviour_modality + --modality docs in AGENTS.md, and the change design/tasks updated to the honest scope (radial-spot geometry rejected as assay-shopping). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
âĶtring; JSON eof Full pre-commit run -a caught two over-length lines (the capture_behaviour_modality comment + a thermotaxis reference-test docstring) and normalised the two 036 supporting JSONs' end-of-file. No behavioural change. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Change complete: thermotaxis capture modality + thermal reference + harness --modality merged into the realworm-behavioural-validation capability (+4 requirements). All §1-§7 task boxes ticked. Result: Logbook 036 (weathervane weakly reproduced + sensor-driven; klinokinesis absent â a behavioural difference). Archived to 2026-07-06-add-realworm-thermotaxis-validation. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Branch-review findings (fail-loud instead of silent-wrong): - load_bias_signatures(modality=...) now raises ValueError on an unknown modality instead of silently falling back to the FOOD references (which would grade thermotaxis captures against Pierce-Shimomura ranges + wrong citations). - run_simulation warns when capture_behaviour_modality=thermotaxis but thermotaxis is disabled (the capture would silently fall back to the food drive, recording a non-thermotactic trajectory). Tests: the toward-comfort direction FLIP when T > Tc (the previously-untested `else tgrad[1] + pi` branch â base above Tc); the unknown-modality raise. Also verified in review (no change needed): capture math exact; food-modality byte-identical; thermotaxis JSON matches the hardcoded fallback (no drift); configs parse; no planning-label refs in the new code. The floor's fully-parked edge (no stride > 0.1x max) is latent + unreachable in practice (a never-moving worm yields no measurable curving) â noted, not fixed. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
|
No actionable comments were generated in the recent review. ð âđïļ Recent review infoâïļ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Run ID: ð Files selected for processing (4)
ð§ Files skipped from review as they are similar to previous changes (4)
ð WalkthroughWalkthroughAdds thermotaxis behavioural capture alongside chemotaxis, with modality-aware thermotaxis references, grading, scenario configs, OpenSpec/spec updates, and logbook/tracker documentation. It also hardens path-length estimation for parked trajectories and adds tests for the new modality paths. ChangesReal-worm behavioural-thermotaxis validation
Estimated code review effort: 3 (Moderate) | ~25 minutes Possibly related PRs
ðĨ Pre-merge checks | â 5â Passed checks (5 passed)
âĻ Finishing Touchesð Generate docstrings
𧊠Generate unit tests (beta)
Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 1
ð§đ Nitpick comments (3)
packages/quantum-nematode/quantumnematode/validation/behavioural_curves.py (1)
242-261: ð Maintainability & Code Quality | ðĩ Trivial | ⥠Quick winAdd Parameters/Returns sections to the docstring.
The rewritten docstring is detailed but omits the NumPy-style
Parameters/Returnssections (kin,fraction, and thefloatreturn aren't documented in that format), unlikeload_bias_signatureselsewhere in this PR which follows full NumPy style.ð Proposed docstring structure
def suggest_min_path_len(kin: Sequence[StepKinematics], fraction: float = 0.25) -> float: """Return a curving-rate displacement floor: ``fraction`` x the typical *moving* stride. Scale-free (data-driven): excludes creep/dwell steps whose tiny ``path_len`` would otherwise make ``dtheta / path_len`` explode. The scale is the raw median stride, EXCEPT when that median has collapsed below ``_MOVING_STRIDE_FRACTION`` of the largest stride â i.e. the worm parks (a converged thermotaxis worm dwelling at its comfort target) â in which case it falls back to the median of the *moving* strides so the floor stays at the true stride scale rather than letting the creep artifact through. A continuously-moving worm keeps the raw-median floor unchanged. - 0.0 when there are no strides. + + Parameters + ---------- + kin : Sequence[StepKinematics] + Per-step kinematics to derive the stride-length scale from. + fraction : float + Multiplier applied to the derived stride scale. + + Returns + ------- + float + The suggested minimum path-length floor, or ``0.0`` when there are no usable strides. """ðĪ 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 `@packages/quantum-nematode/quantumnematode/validation/behavioural_curves.py` around lines 242 - 261, The docstring for suggest_min_path_len is missing NumPy-style Parameters and Returns sections. Update the existing docstring to document kin, fraction, and the float return value in the same style used by load_bias_signatures, while keeping the current explanatory text intact.packages/quantum-nematode/quantumnematode/agent/agent.py (1)
983-1002: ð Maintainability & Code Quality | ðĩ Trivial | ⥠Quick winAdd NumPy-style Parameters/Returns sections to the docstring.
The docstring documents the return semantics via prose but omits the standard
Parameters/Returnssections for the 6 parameters and returned tuple.As per coding guidelines, "Write NumPy-style docstrings in Python code."
ð Proposed docstring fix
def _behaviour_capture_fields( # noqa: PLR0913 self, sensing: SensingConfig, temporal: dict, food_grad_dir: float, food_grad_strength: float, sensing_pos: tuple[float, float], temperature: float | None, ) -> tuple[float, float, float, float]: """Return the captured drive fields, selected by ``capture_behaviour_modality``. - The tuple is ``(drive, drive-derivative, toward-drive direction, gradient strength)``: + Parameters + ---------- + sensing : SensingConfig + Sensing configuration (reads ``capture_behaviour_modality``). + temporal : dict + Temporal sensing fields, used for the "food" fallback. + food_grad_dir : float + Live food-gradient direction (food modality only). + food_grad_strength : float + Live food-gradient strength (food modality only). + sensing_pos : tuple[float, float] + Position to sample the thermal gradient at (thermotaxis modality only). + temperature : float | None + Current temperature at the agent's position, if thermotaxis is enabled. + Returns + ------- + tuple[float, float, float, float] + ``(drive, drive-derivative, toward-drive direction, gradient strength)``:ðĪ 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 `@packages/quantum-nematode/quantumnematode/agent/agent.py` around lines 983 - 1002, Update the docstring for _behaviour_capture_fields to follow NumPy style by adding explicit Parameters and Returns sections. Document all six inputs on the method signature (self, sensing, temporal, food_grad_dir, food_grad_strength, sensing_pos, temperature) and describe the returned 4-tuple with its element meanings. Keep the existing behavioral description, but move the prose into the structured returns/parameter entries so the docstring matches the projectâs Python docstring guidelines.packages/quantum-nematode/tests/quantumnematode_tests/agent/test_behaviour_capture.py (1)
100-129: ð Maintainability & Code Quality | ðĩ Trivial | ⥠Quick winDuplicate agent/env/brain construction across both new tests.
Both tests repeat the same
Continuous2DEnvironment+ConnectomePPOBrain+QuantumNematodeAgentboilerplate, differing only inThermotaxisParams.base_temperature. Consider extending the existing_agent()helper to accept an optionalthermotaxisparam andcapture_behaviour_modality.Also applies to: 144-173
ðĪ 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 `@packages/quantum-nematode/tests/quantumnematode_tests/agent/test_behaviour_capture.py` around lines 100 - 129, The new tests duplicate the same `Continuous2DEnvironment`, `ConnectomePPOBrain`, and `QuantumNematodeAgent` setup in `test_behaviour_capture.py`. Refactor the existing `_agent()` helper to accept optional `thermotaxis` and `capture_behaviour_modality` parameters, and have both tests call that helper instead of rebuilding the agent/env/brain inline. Keep the differing `ThermotaxisParams.base_temperature` values as the only per-test variation.
ðĪ 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 `@docs/experiments/logbooks/036-realworm-thermotaxis-validation.md`:
- Around line 74-88: The Îļ-free confidence interval in the
weathervane/derivative-control summary is inconsistent with the supporting data.
Update the reported interval in this logbook entry where the thermal weathervane
collapse is described so it matches the `klinotaxis_all` values from
`derivative-control-curves.json` (the interval should reflect the rounded JSON
CI), and keep the surrounding `weathervane slope (Îļ-free)` / collapse wording
unchanged.
---
Nitpick comments:
In `@packages/quantum-nematode/quantumnematode/agent/agent.py`:
- Around line 983-1002: Update the docstring for _behaviour_capture_fields to
follow NumPy style by adding explicit Parameters and Returns sections. Document
all six inputs on the method signature (self, sensing, temporal, food_grad_dir,
food_grad_strength, sensing_pos, temperature) and describe the returned 4-tuple
with its element meanings. Keep the existing behavioral description, but move
the prose into the structured returns/parameter entries so the docstring matches
the projectâs Python docstring guidelines.
In `@packages/quantum-nematode/quantumnematode/validation/behavioural_curves.py`:
- Around line 242-261: The docstring for suggest_min_path_len is missing
NumPy-style Parameters and Returns sections. Update the existing docstring to
document kin, fraction, and the float return value in the same style used by
load_bias_signatures, while keeping the current explanatory text intact.
In
`@packages/quantum-nematode/tests/quantumnematode_tests/agent/test_behaviour_capture.py`:
- Around line 100-129: The new tests duplicate the same
`Continuous2DEnvironment`, `ConnectomePPOBrain`, and `QuantumNematodeAgent`
setup in `test_behaviour_capture.py`. Refactor the existing `_agent()` helper to
accept optional `thermotaxis` and `capture_behaviour_modality` parameters, and
have both tests call that helper instead of rebuilding the agent/env/brain
inline. Keep the differing `ThermotaxisParams.base_temperature` values as the
only per-test variation.
ðŠ Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
âđïļ Review info
âïļ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro
Run ID: f8d4bf16-c244-4dd1-a456-2d2ba9ae728a
â Files ignored due to path filters (2)
docs/experiments/logbooks/supporting/036-realworm-thermotaxis-validation/klinotaxis-turn-rate-vs-dcdt.pngis excluded by!**/*.pngdocs/experiments/logbooks/supporting/036-realworm-thermotaxis-validation/klinotaxis-weathervane.pngis excluded by!**/*.png
ð Files selected for processing (24)
AGENTS.mdconfigs/scenarios/thermal_foraging/mlpppo_small_continuous2d_thermotaxis_seeking_derivative.ymlconfigs/scenarios/thermal_foraging/mlpppo_small_continuous2d_thermotaxis_seeking_klinotaxis.ymldata/thermotaxis/behavioural_bias_signatures.jsondocs/experiments/README.mddocs/experiments/logbooks/036-realworm-thermotaxis-validation.mddocs/experiments/logbooks/supporting/036-realworm-thermotaxis-validation/derivative-control-curves.jsondocs/experiments/logbooks/supporting/036-realworm-thermotaxis-validation/klinotaxis-curves.jsonopenspec/changes/archive/2026-07-06-add-realworm-thermotaxis-validation/.openspec.yamlopenspec/changes/archive/2026-07-06-add-realworm-thermotaxis-validation/design.mdopenspec/changes/archive/2026-07-06-add-realworm-thermotaxis-validation/proposal.mdopenspec/changes/archive/2026-07-06-add-realworm-thermotaxis-validation/specs/realworm-behavioural-validation/spec.mdopenspec/changes/archive/2026-07-06-add-realworm-thermotaxis-validation/tasks.mdopenspec/changes/phase6-tracking/tasks.mdopenspec/specs/realworm-behavioural-validation/spec.mdpackages/quantum-nematode/quantumnematode/agent/agent.pypackages/quantum-nematode/quantumnematode/utils/config_loader.pypackages/quantum-nematode/quantumnematode/validation/behavioural_curves.pypackages/quantum-nematode/quantumnematode/validation/datasets.pypackages/quantum-nematode/tests/quantumnematode_tests/agent/test_behaviour_capture.pypackages/quantum-nematode/tests/quantumnematode_tests/validation/test_behavioural_curves.pypackages/quantum-nematode/tests/quantumnematode_tests/validation/test_bias_signatures.pyscripts/analysis/behavioural_chemotaxis_validation.pyscripts/run_simulation.py
Codecov Reportâ Patch coverage is ðĒ Thoughts on this report? Let us know! |
âĶngs, test dedup - pyright (the failing check): the T>Tc flip test compared get_temperature() (float | None) with `>`; coerce None -> 0.0. Type the _agent helper's capture_behaviour_modality as Literal["food","thermotaxis"] to match SensingConfig. - Logbook 036: the Îļ-free weathervane CIs in the results table + the derivative collapse prose showed 95% bounds under an "80% CIs" table; corrected to the JSON 80% CIs (MLP +0.020 [+0.013,+0.027]; derivative +0.000 [â0.008,+0.009]). The separately-labelled 95% claims in the prose are unchanged. - NumPy-style Parameters/Returns on _behaviour_capture_fields and suggest_min_path_len (matching load_bias_signatures / the package's direction). - Dedup the thermotaxis capture tests: _agent() now takes thermotaxis + capture_behaviour_modality; a _thermotaxis(base) helper leaves base_temperature as the only per-test variation. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
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
A second real-worm behavioural-validation modality (thermotaxis), extending the archived
realworm-behavioural-validationcapability from Logbook 035. Non-gating â Gate 3 G3.d stands on chemotaxis; 035 remains the strong, complete validation of record. This is an honest partial / behavioural-difference result (Logbook 036).Result (MLP, n=4, faithful linear-gradient Tc-seeking assay)
Why â a genuine behavioural difference (motivates Phase 7)
The substrate reproduces the spatial-steering half of thermotaxis but not the biased-random-walk, for two principled reasons, reported rather than engineered around:
Method notes / integrity
â|TâTc|, toward-comfort direction), gated bycapture_behaviour_modality(defaultfoodâ byte-identical). Thermal references are sign-only (Luo et al. 2014 primary; magnitudes not literature-comparable).Testing
uv run pytest -m "not nightly"â 4138 passed, 1 skipped, 2 xfailedpre-commit run -agreen ·openspec validate --strictvalid · change archived (capability spec +4 requirements)Post-review hardening (last commit)
A branch review surfaced two wrong-modality footguns, now fixed to fail loud:
load_bias_signaturesraises on an unknown modality (was silently grading against food refs);run_simulationwarns whencapture_behaviour_modality: thermotaxisruns with thermotaxis disabled. + a test for the T>Tc toward-comfort flip.ðĪ Generated with Claude Code
Summary by CodeRabbit