RH-W-19: Reproducibility and Adversarial Certificate Audit

Placing the W-18 unified backend into a replayable attack model — establishing a zoo of 17 classes of false certificates (Z01–Z17). 16 classes of attacks that should be rejected were all correctly rejected. The 17th class, "simultaneous collusion of verifier and artifact," is deliberately marked as EXPECTED_SURVIVAL: a self-certifying system itself cannot rule out this kind of attack and must introduce an external root of trust. Accompanied by a concrete empirical proof — a 14th-order Hilbert matrix is proven positive-definite using exact rational LDL^T, yet the double-precision eigenvalue solver in this environment returns a negative minimum eigenvalue.

RH_CLAIM = False — The claim status self-reported by the verifier within the package, reproduced here as is. "RH-W-19 establishes a 'certificate of certificate failure'. It does not elevate any RH truth value claims, but only improves the identifiability of research artifacts against errors, tampering, and overclaims."

Connections

Regarding its relationship with other packages, I try to use the words from its own documents, not my interpretation.

Belongs toResearch Starting Point v2.2 · Adversarial Reproducible Audit

Previous PackageRH-W-18: Unified Certificate Backend and Single Verification Entry

Next PackageRH-W-20: Batch 01 Integration and AI Autonomous Math Research Platform Case 0001 Release

Batch 0119 / 20

"If an attacker simultaneously modifies the certificate, verifier, manifest, artifact index, and all hashes, and makes the verifier print only the expected token, there is no root inside the package not controlled by the attacker that can identify it. This is not a patchable minor bug in this backend, but a fundamental limitation of self-certifying systems." — Excerpt from the "External Trust Boundary" section of Document 01 in this package.

Erroneous Certificate Zoo · Z01–Z17

Sourced from adversarial_zoo_catalog.csv within the package, reproduced item by item from the original data, not a summary rewritten by me after the fact.

GAP Ledger

Sourced from RH-W-19_subgaps_v0.1.csv within the package; the status and descriptions are recorded by the package itself, not judgments added by me after the fact.

Reproduce

Unified backend verification → Run full adversarial zoo → Individually verify this round's records:

python rhcert.py verify
python redteam_zoo.py
python rhcert.py verify --record RH-W-19

01 · Reproducibility and Adversarial Certificate Audit

Main Document — Six new protections added in backend v0.2, main empirical demonstrations (floating-point false negatives, tail term omissions, inward intervals), external trust boundaries, and the four-stage verification chain.

02 · Error Certificate Zoo and Detection Matrix

Four-layer error classification (package execution layer / mathematical semantic layer / interval numerical layer / external trust layer), detection principle comparison table, and five core prohibitions.

03 · External Root of Trust and Dual-Implementation Route

The boundaries of self-certification, the four-tier trust upgrade recommended for Batch 02 (signed release manifest / reproducible build / independent verifier / external mirror), and the boundaries against overclaiming.

04 · GAP Updates and Batch 01 Progress

Completed/pending items, Batch 01 progress 19/20, next node W-20 (packaging Batch 01, with no commitment to prove or disprove RH).

Verification Output

From ADVERSARIAL_AUDIT.txt (itemized results for 17 classes) and VALIDATION.txt within the package, reproduced line by line.

Loading...

Loading...

Files

This package similarly archives all original documents from W-04 through W-17 (records/), which are not listed individually but are all included in the complete zip file; for the specific contents of each package, please see their corresponding detail pages.

Download Full Package (1.27 MB)

sha256 07a2312a8a78a6783438165d5adc23c86bc81ad5c7947b72385f550374ffc893