Warrant

Warrant records transitions of work: what was done, on what basis, what it established, which exact conditions it closed, which questions it opened, and what it leaves unestablished. The next agent reads a short capsule instead of a transcript, and cannot credit the work with more than its grounds support.

Review is one use. Ordinary decisions, running services, periodic obligations and reproducible computations are others.

The core is small: six signed record types and one pure function, assess, that reads one view of a store (its bytes, the reader's own trust file and profile, a time) and answers what may be relied on now, what is still owed and by whom. The package warrant holds the core and the warrant command, and imports nothing but the Python standard library and cryptography; warrant_agent is an interface for agents on top of it.

• Warrant in 15 minutes: one duty, its closure, an acceptance, a challenge; every command and output on the page is run by CI.
• Contract: SPEC.md
• Handing work to another agent (Python API, CLI, stdio MCP): AGENT.md
• What the rebuild retired, and where the previous generation lives: HISTORY.md
• The proof-space pilot (goal G9), its contract and resolution layer, outside SPEC.md: PILOT.md

Install

pip install --pre warrant-graph               # Python ≥ 3.10; depends on `cryptography`
pip install --pre "warrant-graph[mcp]"        # adds `warrant-agent-mcp` on the official MCP SDK
pip install --pre "warrant-graph[compute]"    # adds the Sigma provider (sigma-glyph==0.8.0)

This installs the commands warrant and warrant-agent and the packages warrant and warrant_agent. The package published as warrant-verify (≤ 0.12.0) is the retired 0.x generation.

Attest and observe need nothing else. compute conditions are replayed by the Sigma provider, the sigma_provider module of the sigma-glyph distribution (PyPI 0.8.0, provider identity 0caff793…; warrant compute --store S provider prints it). A recipient accepts a provider by putting that identity in its profile's providers; without an accepted, installed provider every REPLAYED basis is UNVERIFIED.

The model in one table

Assessment is a pure function of one view: the store's bytes read once, the recipient's trust file and profile, a time, and the current revision. It answers reliance (ELIGIBLE, STALE, CHALLENGED, SUPERSEDED, UNASSESSED, NOT_CLOSED, ...), frontier, questions and capsule. A key authenticates; it does not authorize. A computation is credited only by the recipient's own replay.

Status

2.0.0a1 is an alpha, the first publication of warrant-graph. A pre-release may still change the record formats, the command line, the Python API and the agent interface.

• What is backed. Every normative sentence of SPEC.md is a signed claim record. Of 358 sentences, 281 are BACKED: each names the test cases that witness it and the mutants those cases kill. 20 are UNBACKED, each with the reason it is still open, and 57 are informative. SPEC-CLAIMS.md lists every claim with its witnesses and mutants. A change to a rule is filed as a new record that supersedes the old one and says why; the old one stays.
• What is not. Most of the code is exercised by tests but not tied to a claim. At 7c6d80b, 304 of the 2317 statements of warrant/ (13%), and 431 of the 3980 statements of warrant/ and warrant_agent/ together (11%), are changed by a mutant that a BACKED claim cites (tools/claims.py why --uncovered). Raising that is the open goal G6. The other open goals, among them an agent with no preparation using Warrant correctly (G2) and admission of witnesses without an admin (G3), are in ROADMAP.md.
• Review. Pull requests are reviewed by a model reviewer through the repository's review loop (REVIEWS.md); no external security audit is recorded. There is one implementation, this Python one.

Limits

Not in this version: threshold adoption, key rotation, P-256, policy compilation, grants for acts other than release, goals as a distinct kind, and any independent implementation besides this Python one. The process boundary around the Sigma kernel isolates module names; it is not an OS sandbox. A view cannot prove its own completeness, and every report says so. An assessment has aggregate admission limits (records, bytes, requires edges, periods; SPEC §9): beyond one it is refused, never answered in part.

The repository

The repository also builds and checks Warrant with Warrant: SPEC.md is generated from the claim records, the roadmap and the pull request reviews are records in the same store, and a gate merges a pull request on a recorded acceptance. That machinery (tools/, .warrant/, .github/) is evidence behind the package, not part of it: tests/boundary_test.py refuses a core module that imports it and a wheel or sdist that carries it. docs/REPOSITORY.md describes it, with the checks to run from a checkout; CHANGELOG.md records what changed.

Licensed MIT. See TRADEMARK.md for the name.