Skip to content

中文

RCR-20260831 | Runtime Continuity Evidence Package

Status: public review and reproducibility package. All identifiers are synthetic. It contains no real task ledger, credentials, signing keys, or product source.

This package supports three articles:

  1. Retry eligibility after an audit failure
  2. Cross-session authority, execution principal, and attribution
  3. Admissible evidence at Runtime decision time

Two recheck layers

Public captured data and Reader. Requires Node.js only, with no network or product dependency. It checks article counts, four time cutoffs, and four negative controls.

Product behavior replay. Requires authorized access to the fixed CodeFlowMu source and installed dependencies. The probe uses real services but only a local synthetic executor. It does not test real HTTP, Git, issues, payments, or provider-account switching.

Download the following files into one directory and run:

text
node check.mjs
node check-manifest.mjs

Expected exit code is 0. The public checker confirms three cross-process recovery controls, eleven authority-consumption cases, eight skill-binding cases, eight Reader cases, and the CatchBench PRE count of 1,187 configurations. These sets must not be added together as a reliability score, detection accuracy, or online benefit.

Claim → provenance → fixture → check

EvidenceSupported claimPublic artifactWhat it proves and does not prove
E-A0P0–P3 differ by audit-fault locationhistorical probesP3 preserves a first effect while being written failed; not incident frequency
E-A1New-process recovery depends on dedupe and digestrestartno dedupe 1→2, stable key 1→1, changed digest rejects; not all production executors
E-B1Session change alone does not invalidate matching approvalauthorizationeleven conditions and two paths that cannot consume twice; not trusted external identity entry points
E-B2Session binding distinguishes verified, invalid, and sessionless claimssession binding2 verified, 5 invalid_claim, 1 not_applicable; public package has no HMAC key
E-C0CatchBench PRE totals and source distributionraw outputfixed PRE output and 1,187 configurations; not a fresh model call or CodeFlowMu score
E-C1Proposition, ownership, cutoff, and integrity change admissibilitytimelineReader executes 4 normal and 4 negative controls; not a product auditor or statistical accuracy

Read contract and limits

The public Reader fixes a proposition, admits artifacts with the correct owner at or before the cutoff whose content matches the manifest, and then reads the latest allowed effect snapshot. It evaluates admissible evidence rather than every log. unknown is not false; not_observed is not “will never happen.”

The byte-identical product probe, reproduce-product.mts, is included but requires source fixed at f0f42f01c8f6d55bfe3d32e108f607841a2900d9 and an isolated research output directory.

The same researcher ran these experiments. There is no independent QA, blind evaluation, production incident frequency, real provider-account escalation result, or online early-warning claim. The manifest detects silent change relative to this package; it is not independent attestation.

Last updated: