# Canonical Claim Boundary

This file defines the public technical-evidence boundary for Mian Zhang / Ouroboros Project. It is intentionally stricter than a narrative landing page.

## Public Claims

### P02-C1

WisdomBench measures longitudinal learning from repeated failure, not only single-round task competence.

This does not prove human-like wisdom, general agent reliability, or deployment safety.

### PCA-C1

High-risk AI action should not receive action credit before warrant and receipt closure exist.

This does not prove real-time trading returns, customer readiness, production readiness, or general safety.

### P24-C1

Adaptive systems need relational observability: relations, constraints, control debt, evidence half-life, and actionability gates.

This does not prove a universal theorem for all adaptive systems. Any broader extrapolation needs independent public traces.

### P20-C1

Physical AI should enter recovery or abstention when evidence degrades, rather than acting directly.

This does not prove detector SOTA, offensive autonomy, or certified real-robot deployment.

### F1-C1

Trading is used only as a high-risk testbed for proof-carrying action discipline, not as a public real-time profit claim.

This does not prove live trading edge, alpha dominance, customer readiness, or protected execution quality.

### P22-CURRENT

The current public P22 version is `v0.2 / 10.5281/zenodo.20582195`; the Zenodo publication record contains six public files.

This does not prove that a support record has been published, and it does not prove that TeX is included in that publication record. The previous version `10.5281/zenodo.20156365` is only part of the version chain.

### P41-P45-CURRENT

P41-P45 are the latest public records from 2026-06-06 to 2026-06-07, but each keeps its evidence tier: local constructed evidence, shadow-only runtime evidence, theory plus constructed witnesses, R-shadow candidate included not completed, and C-controlled positive plus public no-credit receipts.

This does not prove real robot deployment, third-party reproduction, live trading, broker execution, task-relevant public model validation, a hard 0.05 margin, or unconditional GEE optimality.

## Public Evidence

Public evidence may include papers, public datasets, public registries, demo routes, redacted no-go summaries, schema notes, claim-to-evidence tables, artifact scope, and counterexample routes.

## Non-Public Material

The public technical-evidence mirror does not disclose credentials, customer data, real trading records, protected logs, protected financial execution details, protected orchestration, commercial schedulers, non-public materials, or protected operating records.

HF README, llms, and index materials also do not carry main-site storytelling, long-form Chinese essays, investor narrative, sales material, or private communications.

## Valid Challenges

Valid challenges include formula counterexamples, data leakage, provenance errors, stronger baselines, reproduction failures, over-broad claim boundaries, credit leaks, authority leaks, and mismatches between claim scope and evidence scope.

## Out-Of-Scope Requests

Requests for private customer data, credentials, protected execution traces, real trading logs, sensitive instructions, non-public material, sensitive identity clues, or real-system attack steps are outside the public boundary.

## Repair Rule

If a public counterexample holds, the related route should be repaired, narrowed, downgraded, or marked as blocked. Repair work does not automatically receive action credit.