fix(hivemind): two silent cross-space corruption paths in ec_merge - #467
Open
dennys246 wants to merge 3 commits into
Open
fix(hivemind): two silent cross-space corruption paths in ec_merge#467dennys246 wants to merge 3 commits into
dennys246 wants to merge 3 commits into
Conversation
Both found by a design-review round on the perception-fabric plan; both are
live on the shipped merge surface, independent of that plan.
1. _cosine used zip(a, b), which TRUNCATES to the shorter vector. ec_merge
gates on modality only (never dimension) and EC node payloads carry no
encoder identity, so a 384-dim node and a 768-dim node of the same
modality tag were compared over the first 384 dims and merged whenever
that partial cosine cleared the 0.44 threshold. Different-length vectors
come from different encoder spaces and are not comparable: return 0.0 so
the pair falls below every threshold and the right-side node inserts as
its own node (non-destructive, matches the zero-norm convention).
2. DEFAULT_FROZEN_CENTROID_MODALITIES was {interoception} while
ECConfig.frozen_centroid_modalities is {interoception, audio} — the
comment already CLAIMED they matched. A default-argument ec_merge
therefore running-mean-updated audio centroids across contributors,
exactly what the local EC forbids for that modality (the EC
centroid-drift lesson). Literal corrected; equality pinned by a test
since the layer deliberately avoids internal-module imports.
Guards: 4 tests (dim-mismatch cosine, cross-dim nodes stay separate,
frozen-set equality with ECConfig, audio centroid stays frozen).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
git add -A picked up an untracked design-plan draft that belongs on its own branch. This PR is a pure src/ + tests fix; the plan ships separately after its review-fold. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…CLAUDE.md's false parity claim
A release-readiness audit independently rediscovered the merge.py bugs and
rated them 1.1-blocking, and found two things this PR missed:
1. similarity/ec.py::_cosine_similarity has the SAME unguarded zip. Reachable
within a single agent across a LOAD boundary, not just cross-contributor: an
ec.json written while sentence-transformers was installed holds 768-dim
LinguisticEncoder nodes, but the same encoder silently falls back to a
384-dim bag-of-words hash when the semantic extra is absent — so the next
session compares 384 against 768, truncates, and pattern-COMPLETES onto an
incomparable node. Now returns 0.0 so it pattern-separates instead.
2. CLAUDE.md asserted the hivemind default is frozenset({interoception})
matching ECConfig.frozen_centroid_modalities. That was already false —
audio was added to ECConfig in PR #383 and never mirrored — so a
behavioral-tier invariant documented a parity that did not hold, on the
modality Exp 45/48 both rest on. Corrected, with the divergence history and
the pinning test named.
Guard: test_cosine_similarity_returns_zero_on_dimension_mismatch.
CLAUDE.md invariant lint passes.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
dennys246
added a commit
that referenced
this pull request
Aug 7, 2026
…ty + safety consequences Operator report: Stewart motors 2 and 3 were broken across essentially all of 1.0+, replaced ~2026-08-05, all six legs now healthy. ROOT CAUSE IS OURS: an earlier Maxim commanded a pose beyond physical capability, the motors glitched, the head snapped to the opposite extreme, and the robot rotated off the table. That is a workspace-limit enforcement failure, connecting directly to the orient-vocabulary audit: MoveTool gaze and turn_around both bypass ReachyMiniController.goto_target, so they plausibly bypass workspace clamping too. Item 8 elevated from correctness to SAFETY, top priority after PR #467. DATA QUALITY: every live-hardware measurement in 1.0+ was taken on a degraded platform — including the 2026-07-16 'TRUE characterization' (0.57 az/rad, R^2=0.9982), which the previous note wrongly assumed was the healthy baseline. Better hypothesis for the contested DoA curve: PROGRESSIVE degradation, with 0.57 (July) and 0.19 (August) both real measurements of a declining platform. Predicts healthy hardware now measures >=0.57, nearer geometric 0.637. Direction findings likely survive; MAGNITUDE findings do not — delivered shift is exactly what degradation corrupts, and that line rested on n=1 sessions. All magnitude claims provisional pending re-measurement. Co-Authored-By: Claude Fable 5 <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.
Found by a design-review round on the cross-modal perception-fabric plan; both bugs are live on the shipped merge surface independent of that plan, which is why they ship separately.
_cosinetruncated viazip—ec_mergegates onmodalityonly, never dimension, and EC node payloads carry no encoder identity. A 384-dim and a 768-dim node with the same modality tag were compared over the first 384 dims and merged if that partial cosine cleared 0.44. Now returns 0.0 on length mismatch, so the node inserts separately rather than contaminating a centroid.DEFAULT_FROZEN_CENTROID_MODALITIESsilently diverged fromECConfig—{interoception}vs{interoception, audio}, while the comment claimed they matched. A default-argumentec_mergerunning-mean-updated audio centroids across contributors, precisely what the local EC forbids for that modality (EC centroid-drift lesson). Corrected + pinned by a test, since this layer deliberately avoids internal-module imports.4 guard tests added. Full hivemind suite green.
🤖 Generated with Claude Code