Open material
Use cases, research notes, agent passports and high-level architecture can be public without exposing private runs or customer context.
Freigabestufen
What can be shown, and where.
Different questions need different proof. A public page can explain the architecture. A real proof room needs the actual run chain, artifact boundary and human review state.
Scoped evidence
Sanitized source snapshots, task traces, generated artifacts, gate verdicts and receipts are shown inside a request-scoped room.
Alpha and Valtis
AMR and property-operations proof tracks are handled as domain rooms, with their own limits, data boundaries and review state.
Limits stay in view
Blocked paths, downgraded claims and failed hypotheses remain part of the proof stack. They are not edited into success stories.
Compare before claiming
Where a benchmark exists, AgentBrain is evaluated against a baseline. If the baseline wins, that result remains visible.
Review before release
Proof is released only when the room can separate demonstrated behavior from context, inference and unresolved unknowns.
Quelle bis Entscheidung
A proof room follows one chain.
The goal is not to impress with a dashboard. The goal is to let a reviewer walk the path from source to decision and see exactly where the evidence stops.
source
Snapshot
Documents, datasets, tickets or operational records captured with scope and timestamp.
agent
Identity
Which agent acted, which mandate applied and which memory substrate was in use.
run
Task trace
The actual request, retrieval path, intermediate reasoning surface and constraints.
artifact
Output
The produced report, decision brief, recommendation or operational action.
receipt
Gate
Human review state, permission boundary, receipt and later recall/reuse evidence.
source-bounded
human-gated
sanitized
claim-limited
receipt-backed
Zugriffsgrenze
Proof is matched to the question.
Current proof rooms
Alpha / AMR
Evidence-fabric and decision-intelligence proof for AMR analysis, with claim boundaries around medical, funding and execution conclusions.
Valtis / Daniel
Property-operations proof showing mandate memory, decision continuity, human gating and operational receipt discipline.
Architecture
Agent passport, memory validity, recall discipline, proof-cockpit structure and protected review surfaces.
Public boundary
No raw dump
Private logs, customer data and unresolved internal traces are not published as open web content.
No overclaim
A working run does not automatically prove autonomy, production readiness or domain authority.
No fake certainty
Unknown, inferred and candidate findings remain separate from verified results.
Proof policy: public pages explain the method; detailed proof is reviewed and scoped before release. That keeps evidence useful without turning sensitive traces into marketing theater.
Bring the claim you want tested. AgentBrain Research Coordination will return the relevant proof room, reference case or boundary note by email.