Mian Zhang

Claim Boundaries

Strong claims need sharper edges.

This page states what the public evidence field claims, what it does not claim, and what can be challenged. A claim without a boundary is not stronger; it is easier to misread. The public layer is built for inspection; protected material remains outside the public evidence layer.

Open Registries Submit Counterexample Evidence Map
A diagram of claim boundaries, evidence routes, and public challenge surfaces.
Public Claims, evidence, demos
Private Protected operational material
Allowed Attack Public claims and artifacts
Not Allowed Protected systems or personal harassment

Public Release Classes

Open materials are divided by verification purpose.

Class Public content Allowed use Boundary
PaperPreprints, claims, formulas, limitationsRead, cite, critiqueNo protected logs implied DataPublic datasets, cards, sample-level artifactsRe-score or audit public splitsNo customer or protected trading data CodeMinimal public scripts and demo codeRun public reproduction pathsNo protected orchestration DemoBrowser-safe protocol demonstrationsInspect interfaces and examplesNo deployment guarantee ChallengeCounterexample templates and examplesAttack a claim constructivelyNo protected-system probing ManifestRegistries, schema notes, artifact indicesCheck versioned public evidenceNo secret inventory

What We Claim

The core public claim is an evidence protocol, not a demand for belief.

The public program studies how AI systems can move from plausible answers to auditable action. The recurring interface is: claim, falsifier, evidence threshold, warrant, receipt, regret path, and clean-learning boundary. 

claim -> falsifier -> evidence threshold -> warrant -> receipt -> regret -> repair

This is a research and engineering claim about how high-risk AI action should be bounded, audited, and repaired. It is not a certificate that every current system instance is deployment-ready.

No. 1

No AGI claim

The project does not claim general intelligence, consciousness, or human-like wisdom. It studies measurable learning, evidence, restraint, and repair.

No. 2

No SOTA leaderboard claim

Public experiments are not presented as universal SOTA rankings. They are evidence panels for protocols, failure modes, and claim boundaries.

No. 3

No live-money claim

Finance-facing public materials do not claim live trading profitability or stable alpha unless a future clean denominator and post-cost evidence support it.

No. 4

No deployment certification

Robot, safety, and agent protocols are not deployment approvals. Real deployment requires independent review, real environment validation, and domain governance.

No. 5

No autonomous weapon claim

The public work is framed around safety, restraint, recovery, evidence integrity, and protective high-risk decision systems.

No. 6

No professional replacement claim

The protocols are designed to improve auditability and decision hygiene, not to replace qualified human responsibility in regulated domains.

Public / Private Boundary

Maximum openness around claims; strict protection around execution.

Public Claim boundaries, formulas, protocols, small runnable demos, public benchmark artifacts, negative results, DOI links, schemas, and counterexample routes.
Private Financial execution details, customer data, protected orchestration, commercial schedulers, meeting or client notes, credentials, unpublished implementation details, live logs, and non-public material.
Redacted Any trace that could reveal protected customer context, protected finance execution, credentials, operational paths, or non-public commercial strategy.
Private 1

Financial execution details

Execution rules, live decision thresholds, order routing, account-specific traces, and market-operation internals stay protected.

Private 2

Customer data

Client identities, tenant logs, sensitive instructions, support records, and commercial pilot details are not public artifacts.

Private 3

Private agent orchestration

Internal agent routing, tool schedules, memory policies, and operational playbooks are not exposed as open-source material.

Private 4

Commercial schedulers

Business automation, deployment calendars, pricing logic, sales routing, and protected prioritization remain confidential.

Private 5

Venue-specific review material

Non-public review material is handled outside the public evidence mirror until it is ready for public release.

Private 6

Keys and credentials

API keys, tokens, cookies, protected endpoints, access policies, and local credential caches are never public evidence.

How To Attack

Attack the public claim through public evidence.

  • Find a formula or protocol counterexample.
  • Find data leakage, label leakage, or provenance confusion.
  • Provide a stronger baseline under the same claim boundary.
  • Show that a reproduction command fails.
  • Show that a claim is broader than its evidence.
  • Show that a no-go decision should have been an action, or that an action should have been a no-go.
Open Counterexample Challenge Inspect Registries

Contribute Without Needing Trust

The useful contribution is one that survives public checking.

A contributor does not need protected access, personal trust, or agreement with the project. The clean route is to name the public claim, point to the public artifact, show the failure or improvement, and keep the repair bounded.

  • Use a DOI, GitHub commit, HF dataset version, or public page URL as the fixed target.
  • State whether the issue is a formula counterexample, data leak, stronger baseline, reproduction failure, or boundary overreach.
  • Provide the smallest command, example, or table needed for another reader to see the problem.
  • Do not request protected financial internals, customer records, protected agents, commercial schedules, non-public material, or credentials.
  • If a public claim survives, cite it; if it fails, submit the bounded repair path.
trust-free contribution = fixed public target + failure mode + reproduction path + bounded repair
Open Source Routes Challenge Rules Registries