Cross-Lineage Harness — Stage-0 Pre-Registration (decorrelation memo pair)

Targets (dual — a noted deviation from harness §1's single-target default): the Decorrelation Metrics memo and the Path B Decorrelation as a First-Class Requirement for Independent AI Evaluation memo. They are reviewed in one batch because the position memo's measurement backing is the metrics memo; splitting them would force the position-memo reviewers to evaluate an uncited instrument. Each memo's own §6 routes it here before promotion / external use.

Frozen by the steward before any subagent is spawned (Stage 0). Per the ARP §2 default, the adversarial pass is never auto-run — only on an explicit steward GO, cross-lineage, never in the authoring lineage (Anthropic/Claude authored both memos).

0. The Rung-0 self-reference firewall (read first)

This run is partially circular by construction: a harness run that reviews a memo arguing the harness's decorrelation discipline works cannot be clean evidence for that discipline — the same firewall Pipeline Run 001 flagged when its own adversary called the Rung-0 move self-serving. Therefore:

• This run is non-evidence for the claim "decorrelation improves evaluation." It tests only whether the two memos are internally sound, faithfully cited, and survive decorrelated scrutiny — not whether their thesis is true.
• The discipline is to look hardest for the result that cuts against the memos (e.g., that the Run 001 scoring is too thin to support the metrics; that "decorrelation as a first-class requirement" over-reaches; that the position memo legitimates the Anthropic frame it should hold at arm's length).
• Promotion / external use remains a separate steward decision after the run; surviving this run is necessary, not sufficient.

1. Role → lineage assignment

Author lineage = Anthropic/Claude. Every judgment role is a different lineage; the adversary is blind. One parallel batch of Cursor subagents; each given only its named documents + role prompt.

2. Codebook

• Severity: BLOCKING (a memo is internally unsound, materially misquotes the framework, the metrics cannot bear the weight placed on them, or the position memo is captured-by-frame) · MAJOR (a real hole that weakens) · MINOR (fix-on-revision) · AFFIRMING.
• "Survives" ≠ "the thesis is true" (see §0). Survives = no BLOCKING stands after synthesis; a BLOCKING sends the relevant memo back to a bounded revise-loop.
• Convergence is not the success metric — surviving the adversary is. Preserve divergence; do not majority-vote.
• Eat-own-dogfood check: because the position memo argues for decorrelated review, this run is also a live demonstration; the synthesizer should note whether the run itself exhibited real decorrelation (log the per-issue × per-lineage detection matrix per the metrics memo §5 instrument it is reviewing — the cleanest possible test of the instrument is to run it on its own review).

3. Specific things each reviewer must return a verdict on

• Metrics memo: (a) is the Run 001 scoring honestly labeled as reconstructed-not-measured? (b) is the runnable-now / needs-instrumentation triage correct? (c) does M1 (single-lineage-catch fraction) have a defensible target range, or is it un-anchored (the memo's own OQ2)?
• Position memo: (d) is the financial-vs-epistemic-independence gap real, or does the full Anthropic framework already address it? (e) is the "credible on method, not domain depth" boundary held, or does the memo over-claim? (f) does §6 (self-application) actually bind, or is it decorative?

4. Stage-0 freeze checklist (steward)

• Both memos frozen at their current version (no edits after GO)
• Role → lineage assignment accepted (synthesizer ≠ author confirmed)
• Codebook + role prompts + the §3 verdict list locked
• Confirmed: non-evidence for the decorrelation thesis (Rung-0 firewall, §0); non-promotion
• GO (steward) — author may spawn the blind cross-lineage batch

Frozen pending GO. Results — per-role verdict + severity tally, convergence (2+ lineages), open divergence, consolidated revise-list, the per-issue×per-lineage detection matrix, and the steward go/no-go — appended here after the run.

Provenance and register

collaborative — human-directed AI drafting; steward freeze + GO pending. Working companion to the two decorrelation memos; per the harness Output spec, harness run files are not registered in _EXCHANGE_INDEX.md — they live beside the artifact under review.