Hebbian weight vs. query-local relevance tradeoff (collapse v2)

[ClaudioDrews]

Summary

The collapse module's Hebbian cross-source corroboration (amplify_gain) was applied as a blind multiplier — facts that appeared in multiple sources received a flat 15% boost per corroborating source, regardless of how relevant they were to the current query. This rewarded globally-important structural facts (e.g., "Honcho abandoned as memory platform") over query-locally-relevant facts (e.g., a domain-specific fact about Qdrant usage) when the query was about Qdrant configuration.

Root cause

In score_all() (icarus/collapse.py:162):

boost = min(corro * amplify_gain, amplify_cap)

amplify_gain (0.15) was applied without interaction with the candidate's base salience. A fact with Corro:1 and base 0.640 received the same +15% boost as a fact with Corro:1 and base 0.900 — but the lower-base fact is less query-relevant and shouldn't receive full amplification.

Fix (committed in #24)

boost = min(corro * amplify_gain * bases[i], amplify_cap)

The base salience — which already encodes query-token overlap — attenuates the Hebbian boost. A globally-important fact with low query relevance receives proportionally less amplification.

Effect on real data (query: "como configurar o Qdrant para memory-os")

Candidate	Base	Corro	Boost (v1)	Boost (v2)	Salience (v2)
Memory OS origin (fabric)	0.521	3	0.450	0.235	0.644
Honcho abandoned (facts)	0.556	1	0.150	0.083	0.603
Qdrant usage fact (facts)	0.546	0	0.000	0.000	0.546

Result: the query-relevant Qdrant fact now survives (was pruned in v1). Honcho still survives but dropped from rank 3 to rank 5.

Tuning

ICARUS_COLLAPSE_AMPLIFY_GAIN (env var, default 0.15) is the knob for this tradeoff:

• Current (0.15): with attenuation, effective boost is ~0.08 per corroboration for average-base (~0.55) candidates, ~0.12 for high-base (~0.80) candidates.
• If corroboration feels undervalued in production: raise to 0.20. This restores effective boosts to roughly pre-attenuation levels (~0.11 per corroboration at average base, ~0.16 at high base).
• If structural facts still crowd out query-relevant ones: lower to 0.10 or raise ICARUS_COLLAPSE_OVERLAP_WEIGHT (query overlap weight, default 0.55).

Monitor via ICARUS_COLLAPSE_DEBUG=1 — logs the full salience-ranked pool with base, corroboration, and final salience for every candidate.

Related

• Collapse module: icarus/collapse.py
• Integration: icarus/hooks.py → _apply_collapse()
• Tests: _test_collapse.py (all passing with v2)

[ClaudioDrews]

Resolved in 4dec57e (PR #24).

The attenuation boost = min(corro * amplify_gain * bases[i], amplify_cap) was implemented, tested against real pipeline data, and merged. Key results from the test (query: "como configurar o Qdrant para memory-os"):

• A query-relevant Qdrant usage fact now survives (was pruned in v1)
• Honcho attenuated from 0.736 → 0.603 (base 0.556 × 1.083 vs old 0.640 × 1.150)
• Context reduction maintained at 30-33%

Future tuning, if needed: ICARUS_COLLAPSE_AMPLIFY_GAIN (default 0.15). Raise to 0.20 if genuine corroboration feels undervalued in production. Open a new issue with concrete data when that happens.