research: bullshit-detector v1 CRITICAL-only delta (3 CRITICAL integrated; 7 non-CRITICAL deferred to v2)

[AceHack]

Summary

Delta-style revision integrating only the 3 CRITICAL findings from Aminata's Otto-100 pass (PR #284) into Otto-99's bullshit-detector design (PR #282).

7 non-CRITICAL findings (4 IMPORTANT + 3 WATCH) deferred to v2 delta.

Closes CRITICAL-integration step of Aminata-then-Otto-response loop.

3 CRITICAL integrations

#	Finding	Response
C1	Cross-detector collusion	New §Reviewer-cone overlap section acknowledging fundamental limitation; human sign-off as cone-breaking authority
C2	Min-merging Goodhart-bait	Sensitivity-analysis-gate pattern: band = YELLOW when G_carrier_overlap was gate closest to threshold
C3	G_evidence fig-leaf + dead likely confabulated in v0	New §v0 scope subsection: 4-gate effective; likely confabulated not-yet-reachable

Fundamental limitation acknowledged

C1 (reviewer-cone overlap) is NOT closable by detector design alone. The detector can't break its own reviewer-set lineage coupling. Mitigation via routing: human sign-off + external-reviewer signatures are the cone-breaking authorities; intra-factory reviewer concordance downgrades confidence rather than inflating.

Deferred to v2

4 IMPORTANT: G_coverage_plausibility gate · Otto-wake second-reviewer schema tightening · DetectorOutputBatchRetracted · no-signal vs kNN-evasion distinction.

3 WATCH: distribution histogram · adversarial worked example · TLA+ invariants.

DISMISS unchanged: parameter-ADR gate (reused from oracle-scoring v0).

Scope limits

• Does NOT rewrite Otto-99's design — specifies delta only
• Does NOT address IMPORTANT/WATCH findings
• Does NOT implement
• Does NOT propose human-sign-off UI
• Does NOT commit content-similarity-in-cone (alternative to C2)
• Does NOT change 5-gate target structure (v0 effective is 4-gate; v1-post-substrate is 5-gate)

Authority

Within standing authority per Otto-82/90/93 calibration — research-grade design revision.

Dependencies

• Aminata adversarial pass on v1 delta (5th session-pass if lands)
• Integrate v1 changes into Otto-99 design PR (separate PR)
• v2 delta for non-CRITICALs
• Independent-oracle substrate for full G_evidence
• Human-sign-off UI/protocol

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[Copilot]

Pull request overview

Adds a research-grade delta document that integrates the 3 CRITICAL findings from Aminata’s Otto-100 pass into the provenance-aware detector design narrative, while explicitly deferring the 7 non-CRITICAL findings to a later v2 delta.

Changes:

• Introduces a new research doc describing the CRITICAL-only delta and its intended insert/replace points in the Otto-99 design.
• Specifies responses for C1 (reviewer-cone overlap limitation), C2 (carrier-overlap sensitivity downgrade), and C3 (v0 scope + reachable output types).
• Captures deferred IMPORTANT/WATCH items and adoption dependencies as explicit scope boundaries.

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated 5 comments.

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

[Copilot]

P1: The blockquote line starting with + (> + the adversarial-reviewer persona ...) will be parsed as an unordered-list item. Repo guidance explicitly avoids line-start + in prose/lists because it’s easy to accidentally create nested lists and confuse readers/linters; please rewrite this as normal prose (e.g., “the detector itself, the adversarial-reviewer persona, … share …”) or use - bullets consistently.

Suggested change

	> **re-introduces one meta-layer up**: the detector itself
	>
	> + the adversarial-reviewer persona + the external-peer-
	> agent reviewer + any other factory-internal reviewer
	> share training-corpus / repo-access / PR-comment lineage.
	> **re-introduces one meta-layer up**: the detector itself,
	> the adversarial-reviewer persona, the external-peer-agent
	> reviewer, and any other factory-internal reviewer share
	> training-corpus / repo-access / PR-comment lineage.

[Copilot]

P1: This section says “Maintainer sign-off is the only cone-breaking authority”, but the next bullet says external-reviewer signatures are also cone-breaking. Please clarify the scope (e.g., “only internal cone-breaking authority” / “default required authority for authorization-impacting decisions”) so the policy doesn’t contradict itself.

[Copilot]

P1: The sensitivity pseudo-code mixes a predicate definition with a rewrite arrow (sensitivity → YELLOW) and then uses sensitivity without applying it (it’s defined as sensitivity(y|q) = ...). This is ambiguous for implementers; please rewrite as a boolean predicate (no quoted string) and use a standard conditional like if band_raw = GREEN && sensitivity(y|q) then YELLOW else band_raw.

Suggested change

	"G_carrier_overlap was the gate closest to threshold"
	↔ overlap_ratio(cone(q), cone(y)) ∈ [θ_med - ε, θ_med]
	OR ∈ [θ_high - ε, θ_high]
	band(y | q) = if band_raw = GREEN AND sensitivity → YELLOW
	else band_raw
	overlap_ratio(cone(q), cone(y)) ∈ [θ_med - ε, θ_med]
	OR overlap_ratio(cone(q), cone(y)) ∈ [θ_high - ε, θ_high]
	band(y | q) =
	if band_raw = GREEN && sensitivity(y | q)
	then YELLOW
	else band_raw