[PIR][WP6] CI gate: controlled-replacement causal audit (arXiv:2607.26773) required before latent-channel merge

Epic: #837 · Plan: docs/research/perpetual-intelligence-runtime/03-program-plan.md

Goal

Require every PR touching latentmesh-align-consuming code or the quarantine module (WP7) to pass a controlled-replacement causal audit before merge. A standing CI gate, not a one-time paper reproduction.

Evidence

arXiv:2607.26773 — "Do Latent Channels Actually Communicate? A Causal Audit of Latent Multi-Agent LLM Communication." Grade A.

Date correction: arXiv's submission timestamp is 2026-07-29, not Aug 5 as the founding brief stated. No Aug 5 revision exists. Use the correct date in every citation.

The paper's argument is why this gate exists: end-task performance gains alone do not establish that a receiver actually used task-relevant information passed through the latent (KV-cache) channel. A system can improve for reasons unrelated to the mechanism its authors credit. The paper introduces a controlled-replacement audit — substitute the channel contents, observe whether behavior changes as the causal story predicts — tested on Qwen3-4B/8B over GSM8K, ARC-C, and MATH-500.

This is the evidentiary basis for invariant 3 ("every agent communication is attributable").

Acceptance criteria

• A controlled-replacement causal audit implemented per the paper's methodology, runnable as a CI job.
• CI blocks merge on audit failure for any PR touching latentmesh-align-consuming code or the WP7 quarantine module. Advisory-only is not sufficient — a gate nobody must pass is documentation.
• The audit report is witness-logged.
• Audits run against benchmark families equivalent to the paper's (GSM8K / ARC-C / MATH-500, or internal tasks of matched character), with the mapping from their benchmarks to ours documented.
• Pass/fail criterion stated numerically and pre-registered, not decided per-run after seeing results.
• A known-null control: a channel deliberately carrying no task-relevant information must fail the audit. Without this, the gate cannot be shown to have discriminating power.
• Nightly re-runs scheduled via Cloud Scheduler against the day's promoted mutations, feeding WP12's daily cycle.

Dependencies

WP5 (#849) — there must be a live channel to audit. Blocks WP7 (#853).

Start with docs/research/perpetual-intelligence-runtime/03-program-plan.md, then inspect code consuming latentmesh-align and the WP7 quarantine module. Define the controlled-replacement audit, its numerical pre-registered pass/fail rule, witness logging, benchmark mapping, known-null control, merge-blocking CI job, and Cloud Scheduler nightly rerun; done means every listed acceptance criterion is implemented and documented.