Skip to content

USAGE.md

Latest commit

 

History

History
93 lines (69 loc) · 2.71 KB

USAGE.md

File metadata and controls

93 lines (69 loc) · 2.71 KB
Raw

Crucible Usage

Crucible evaluates claims, measurements, and verdicts with recheckable evidence packets. It is built for humans and agents that need to know why a conclusion held, drifted, or became unverifiable.

Install

python -m pip install crucible-bench

From a source checkout:

python -m pip install -e ".[dev]"

Run

crucible status --json
crucible doctor --json
crucible demo --json
crucible --help

The same package can be exercised from source with:

python -m crucible --help

MCP

Use crucible mcp when a host needs the measurement and verdict tools over stdio.

crucible mcp

Verify

python -m pytest
python examples/demo.py

For public/developer delivery checks:

python -m public_surface_sweeper . --workspace --json

Replay oracle measurements

Create a replay template from a witnessed assessment, let the named oracle fill its measurement rows, and submit the completed pack against the same registry:

crucible recheck REGISTRY --template replay-template.json
crucible recheck REGISTRY --pack replay-pack.json --json

The template output schema is crucible.replay-template/1; an external replay producer emits crucible.replay-pack/1. If a pack has a schema field, it must be exactly that value; schema-less legacy packs remain accepted, while template or unknown schemas are rejected.

Each new template includes a privacy-bounded replay_binding with schema crucible.replay-set/1, descriptor and skipped counts, and a SHA-256 digest. To recompute it, form one exact {recheck, expected_measurement} object per descriptor-bearing replay, serialize each with compact sorted-key JSON (ensure_ascii=False, separators=(',', ':')), sort by those serialized forms, serialize the resulting list with the same settings, and hash its UTF-8 bytes. Descriptorless rows and their evidence are never included; the top-level assessment triple binds the full sealed assessment, which Crucible independently verifies.

Keep the template's top-level assessment object unchanged in the completed pack. The CLI requires its thesis ID (thesis_id), assessment seal (assessment_seal), and measurement seal (measurement_seal) to match the selected assessment; a missing, malformed, or mismatched binding is rejected before any measurement replay runs. Preserve replay_binding when emitting a pack; Crucible compares it to the expected canonical value. Older assessment-bound packs that omit replay_binding remain accepted.

Boundary

Crucible should expose claim ids, criteria, verdicts, evidence hashes, and redacted references. Do not require raw prompts, private evidence, verifier internals, or full result payloads for interop.