Skip to content

FIX-0-7: faithfulness noise floor from forced-independent OptiCAM passes #400

Description

@mateuszwalo

Track: research · Level: core · Effort: ~6h · Phase: 0 · Depends on: FIX-0-2
Source: T20 findings III.7 (#389)

Why this matters

The base diagnosis and base_frozen were expected to be two independent OptiCAM passes over
identical weights, which would give the noise floor for the faithfulness metrics. Observed
paired |dEBPG| has median 0.0002 and maximum 0.0004, which indicates the second pass reuses
cached maps. T20 therefore reports no floor, and the ICD/AICD dissociation rests on complete
separation and paired consistency instead of on a noise-threshold argument.

Same cache as FIX-0-2, so that one lands first.

Steps

  1. Add a forced-recompute path that bypasses the cache for a diagnostic pass.
  2. Run two genuinely independent OptiCAM passes over identical weights.
  3. Report the floor for EBPG, IoU@0.5 and Pointing Game.
  4. Re-state the III.3 conclusions against the measured floor.

Done when

The findings carry a measured noise floor, and any faithfulness difference is reported
relative to it.


How to take this: comment "taking this" and wait to be assigned. Branch fix-0-7-faithfulness-noise-floor from upstream/main, and put Closes #<this issue number> in your PR. Full workflow: the Cohort Handbook (pinned in Discord).

Plan: plans/FIX-PLAN-T20-2026-08-23.md, section 4.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    P2Audit backlog: next cyclecohortBNNR 2026 cohort tasklevel:coreCore tasktrack:researchCohort research task

    Type

    No type

    Projects

    No projects

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions