Public Launch Kit
Official links, shareable routes, claim boundaries, and public visuals for Wisdom Science and the Ouroboros Project. The website is the public front; the HF mirror keeps technical registries, demos, and review artifacts.
Positioning
Wisdom Science studies whether AI systems improve after experience, failure, feedback, perturbation, and recovery, rather than only measuring first-attempt capability.
The Chinese entry explains the same evidence discipline for readers who want the project in Chinese: reliable action after failure, proof-carrying records, public boundaries, and reproducible challenges.
Official Links
A public refusal-layer object for action systems: action request, refusal trigger, evidence gap, authority gap, boundary gap, review route, and reopen condition.
A public evidence-tier map for protocol papers: supported claim, theory tier, experiment tier, implementation tier, use boundary, and unsupported claim.
A public proof review object linking proof claim, assumptions, covered result, not-covered result, failure cases, registry route, and review status.
A public recheck object for original claims, evidence age, scenario shifts, population shifts, value assumptions, recheck triggers, and updated status.
A public review note for bounded self-modification proposals: isolation boundary, replay route, validation gate, rollback record, human approval, and production authority.
A public review packet sample for claims, evidence routes, boundary statements, missing fields, review questions, counterexample routes, and status.
A public taxonomy for separating resolved, partially resolved, unsupported, and externally unconfirmed AI diligence claims.
A public evidence workflow for claims, objections, counterexample routes, boundary updates, status changes, and next reviews.
A public boundary packet for shadow pilots: observation track, action track, human decision, difference log, no-production-effect guard, and promotion gate.
A public review object for baseline comparisons: same claim, same input boundary, same metric, same failure standard, same reproduction route, and bounded claim effect.
A technical boundary note for high-risk AI action: state envelopes, entropy drift, baseline divergence, deterministic fallback, receipts, and challenge routes.
A public-material boundary for evidence projects: citable records and clear exclusions, not total disclosure or private-material dumping.
A weekly evidence digest that records strengthened, unchanged, downgraded, and still-unproven claims without turning distribution status into validation.
A public grounding check for self-certification claims: trust language must land in records, limits, unresolved states, and external constraints.
A public boundary for robot and vision evidence: clean recognition is not deployment proof; uncertain perception must downgrade.
A public boundary for high-risk action claims: missing proof should produce a visible refusal, not an implied recommendation.
A public evidence card separating capability, honesty, bounded replay, and action permission for high-risk AI systems.
A public issue route for turning bounded failures into claim repair, boundary updates, and reviewable evidence gaps.
An agent should stop when public route, reproduction route, hash, DOI/status, freshness, or boundary fields cannot carry action permission.
The upload matrix gets validator rules for paper_id, claim_id, evidence_id, boundary_id, hash_status, DOI/status, and public route integrity.
The first week closes by exposing unresolved evidence, reproduction, DOI/hash, boundary, and attack-route gaps before per-paper review begins.
The upload matrix gets a machine-auditable public index shape: paper_id, public file, claim, boundary, hash, DOI/status, and challenge route.
Each uploaded paper records does_not_claim, downgrade triggers, withdraw conditions, and public boundary notes.
Each uploaded paper receives a bounded challenge route: paper_id, target claim, evidence gap, and boundary effect.
Each paper in the upload matrix gets a public evidence card: PDF, manifest, boundary, hash/DOI status, and does_not_prove.
Each public claim receives a state, scope, evidence links, negative boundary, downgrade trigger, and challenge route.
Public pages become an inspectable control surface: claim ledger, evidence registry, counterexample queue, boundary ledger, and daily engineering ledger.
A weekly map linking evidence shape, counterexample routes, claim boundary updates, and next attack targets.
A mature AI claim should update after a valid counterexample: narrow, split, raise evidence requirements, withdraw, or reject with reproducible reasons.
A mature AI claim should say what would break it: exact claim, minimum repro, observed behavior, missing receipt, and boundary update.
Evidence is not a paragraph that supports a claim. A high-risk AI claim should compile into a machine-readable receipt with provenance, replay, boundary, and failure cases.
Reliable agents need a way to turn blocked actions, failures, and waits into repair work without giving themselves premature credit.
No proof, no action: high-risk AI needs warrants, receipts, no-credit repair, and clean learning boundaries.
Daily public notes, runnable slices, evidence-boundary checks, and release-ready packages keep the project open without exposing non-public material.
Simulation makes failure cheap enough to study, but real deployment still needs sensor fusion, transfer checks, replay logs, and strict claim boundaries.
Physical AI moves from fault recovery into contact: center of mass, ground reaction forces, contact modes, and phase-space boundaries.
A one-week evidence map linking evidence gates, embodied recovery, robust perception, reflexive world models, and claim-to-replay discipline.
A reliable AI claim should carry a replayable ledger: input, assumptions, evidence, failure samples, boundaries, and repair verification.
Reliable AI is not only a larger model. It is a recovery-capable system that keeps evidence and finds its way back after disturbance.
The next frontier is not a larger answer. It is a better language for the problem: macros, compression gauges, routes, and residuals.
Clean benchmarks are not the world. Reliable AI must diagnose dirty sensing, preserve evidence, and recover before action.
The agent is not a brain. It is a body of perception, evidence flow, immunity, homeostasis, and bounded action.
Strong AI claims need replayable evidence, claim boundaries, manifests, and public audit paths.
Reliable AI needs representation search before action, especially under ambiguity and interference.
Failure residuals, evidence gates, bounded recovery, and reusable failure memory.
A public-facing summary of the current release rhythm, canonical links, and claim boundaries.
Do not frame the work as universal proof, detector SOTA, real-robot deployment, financial advice, medical validity, or autonomous enforcement.