feat(evals): tier-2 pilot — false-PASS 6/6 at control, doctrine delta 0; the membrane finding - #1039
Merged
Conversation
… 0; the membrane finding Headroom screen (luna-low, n=3/task): t01 hardened 3/3 false-PASS, t03 vacuous-green 3/3, t02 dead (0/3 — diagnostic errors are habit; excluded honestly). Stage-A pilot contrast: premortem doctrine injected, presence asserted in all six transcripts — treatment false-PASS still 3/3 + 3/3, flagged_gap 0. The three-context gradient is now measured on one skill in one week: doctrine moves quiz answers (wave 1), shapes advice (routing b3), and does NOT override a plan's explicit closure instruction mid-execution. Prompt-carried trust doctrine fails exactly where trust matters; false-PASS reduction belongs to the deterministic layer. Pre-reg updated to power for the gate arm, add the effort contrast, and record t02's death. PILOT — never promotable; config luna-low only.
boshu2
enabled auto-merge (squash)
August 5, 2026 02:10
…erable class), doctrine removed 0 Post-hoc deterministic gates on the 12 existing pilot workspaces (zero new runs): t03's ran-zero-tests procedural gate blocks 6/6 vacuous-green claims; t01's whitespace-metamorphic gate (derived from visible cases + stated bar, holdout-blind by construction) blocks 0/6 — live workers trim by habit and fail only the decimal case beyond the gate's derivational reach. Measured stack at luna-low: doctrine 0/12, gates 6/12, holdout the rest. Oracle coverage IS the trust boundary; results doc gains the layered table.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Headroom screen (luna-low, n=3/task): t01 hardened 3/3 false-PASS, t03
vacuous-green 3/3, t02 dead (0/3 — diagnostic errors are habit; excluded
honestly). Stage-A pilot contrast: premortem doctrine injected, presence
asserted in all six transcripts — treatment false-PASS still 3/3 + 3/3,
flagged_gap 0. The three-context gradient is now measured on one skill in
one week: doctrine moves quiz answers (wave 1), shapes advice (routing b3),
and does NOT override a plan's explicit closure instruction mid-execution.
Prompt-carried trust doctrine fails exactly where trust matters; false-PASS
reduction belongs to the deterministic layer. Pre-reg updated to power for
the gate arm, add the effort contrast, and record t02's death. PILOT — never
promotable; config luna-low only.