Independent Verification Center

Every real criticism, failed experiment, revision, bug, and correction documented across this lab, in one place. Not new content — a front door to what's already real and already published.

PurposeMake the honest record easy to find, instead of scattered across pages someone would only find by already knowing they exist.

StatusActive Development — an index, not a new document

Built fromReal pages already live on this site

Depends on—

Superseded by—

EvidenceEvery link below goes to a real page with real content, checked before this index was published.

Bugs found and fixed

Component A Changelog 9 real fixes, in order, including a documented self-correction mid-investigation. Isolation Under Adversarial Pressure 500 adversarial calls against a naive baseline (277 breaches) and a genuinely hard competitor (20 breaches) — NCFCA: 0 both times. Full-Stack Integration — Five Claims, One Scenario Real cross-component wiring: a tripped circuit breaker correctly blocks an unrelated action with its own valid prerequisites and token. Value Laundering — Found & Fixed A real vulnerability in Claim 1's isolation guarantee, diagnosed and closed. Governance Layer A real operator-precedence bug, with a concrete before/after case proving it mattered. Pipeline Orchestrator 4 real bugs found on first read, fixed, then run end to end for the first time. Core Kernel & Pipeline Infrastructure One real stray-enum bug found in multi_stage_pipeline.py.

What's still only backtested, not field-validated

Full System Audit What's wired in vs. genuinely orphaned — and a real drift risk found by auditing the actual call graph. Backtest-Only Audit Full list of what's only ever been tested against its own construction, ranked by real priority. Open Problems Every unresolved question the proof itself names — several with real test data already attached. Execution Kernel Audit Cycle 10-item hardening checklist, line-checked — 4 fixed, 1 partial, 5 confirmed still open.

Real attacks tested against the architecture

State Contamination Benchmark 10,000 deterministic trials — isolated architecture: 0% contamination. Shared context: 55.9% of trials contaminated. Fully reproducible, download included. Evaluation Package Full adversarial threat model — testing against the strongest competing approach, not the weakest. Estimator Correlation Finding 6 of 15 estimators found effectively measuring one quantity through different names.

Simulation log

Simulation Log Self-contained probes of baseline-tracking design (genesis poisoning, boiling frog, compound attacks, adversarial scenarios, environment noise, combined stress), published with code and a plain reading of what each does and does not show. Not tests of the architecture itself.

Methodology — how this lab stays honest with itself

How This Lab Stays Rigorous 8 real verification rules, each with a real example of the error it caught.