v0.4.0-alpha

PRISM

Live connection to prism.opensentience.org/mcp — Workbench v0.4.0-alpha. PRISM benchmarks memory and skill systems on nine continual-learning dimensions. This page proves the wire works and stages the EvidenceBundle hand-off.

What PRISM does with a bundle

A SkillBundle and a PRISM EvidenceBundle are the same JSON read by two audiences. To you it is a reusable, replayable record of something that worked. To PRISM it is a scored observation: one system, one scenario, one outcome, joined to every other run of that scenario.

Why the same file serves both
A benchmark fed by artifacts produced only for benchmarking measures the benchmark. A bundle minted from work someone actually did carries the awkward cases that invented scenarios leave out — the retries, the wrong turns, the tools nobody planned to call.
The hand-off is by content hash
A bundle is identified by content_hash, which seals the manifest, the trace and the proof result together. Anyone can recompute it, so a scored run cannot quietly refer to a different bundle than the one that was verified.
The loop this closes
Graphonomous emits an OutcomeSignal to PRISM's observe phase; PRISM emits a ReputationUpdate from diagnose. Both tokens are declared in PULSE v0.1.1. What does not exist yet is the edge back into Workbench — see docs/ROADMAP.md Phase 4. When it lands it may rank retrieval and must never influence a gate verdict.

Leaderboard

Loading…

Registered systems (MCP)

Connecting…

Hand off a SkillBundle to PRISM

PRISM v0.2 does not yet expose a SkillBundle ingest endpoint. The bundle's content_hash is the stable identifier a future registry will key on. Until then you can copy the bundle JSON below and pipe it through your own PRISM tooling.

No local bundles yet. Record a session you already ran, or teach a new one.