Warrant

When a machine says something was allowed, can you check why — without trusting the machine that allowed it?

Warrant is a small, signed, content-addressed decision record. It names what was decided, the exact policy bytes in force, the reasons and evidence, the actor, and the decisions that came before it. An executable reason can be re-run by the reader instead of trusted as a log entry.

{
  "decision": "reject",
  "subject": {"hash": "d5cf37…", "note": "PR-42"},
  "under": ["cb3a0a…"],
  "because": [
    {"kind": "check", "check": "05d234…", "runtime": "cmd@v1",
     "verdict": "fail", "transcript": "9dc0c3…"},
    {"kind": "prose", "text": "policy clause 1: coverage drops"}
  ],
  "evidence": ["9dc0c3…"],
  "actor": {"id": "agent-b@vendor2"},
  "prior": ["00f79f…"],
  "ts": 1751677200
}

The canonical body hash is the record's identity. Changing the decision, policy reference, reason, evidence or prior edge changes that identity. The signature says which key signed the body; whether that key belongs to the named actor is a separate trust-configuration question. A rejection is a durable record, not an absence.

What a clean verification does not show: that the reason implements the policy. Warrant pins the policy bytes and re-runs the reason, but nothing binds one to the other. A record filed under a DENY_ALL policy, with a re-executed ski@v1 reason that computes something unrelated, and an accept, verifies with 0 errors in all three implementations, even at settlement grade with --require-bound (tests/countermodel_policy_binding.py). Nor that the decision is the one the policy answered: an accept whose reason is the policy's own compilation, replaying to false, also verifies clean. The missing links, policy → computation → decision, are open design work (WRT-020, with a read-only profile that reports them); whether the facts are true is WRT-010.

Warrant is not an agent framework, blockchain or observability system. It is one file format and five filing verbs, designed to be boring.

The format is written up in The Reason Runs Again (10.5281/zenodo.22172098, paper v1.0.0, 2026-08-30 — deposited, not peer reviewed; see papers/).

Try it

Open the browser lab — inspect a signed repository example, re-run its portable reason locally, change its bytes, or open your own store. The browser runs the Rust core compiled to WebAssembly; the report keeps the same schema as the CLI. Its local limits and build identity are shown on the page. See web/README.md for implementation and deployment details.

pipx install warrant-verify   # or: pip install warrant-verify

warrant init
warrant keygen --out me.key
printf 'demo diff\n' > diff.patch
printf 'clause 1: no coverage drop\n' > policy.txt
printf '#!/bin/sh\necho coverage 71%%; exit 1\n' > check.sh && chmod +x check.sh

POL=$(warrant policy add policy.txt)
P=$(warrant propose --subject diff.patch --under "$POL" \
      --reason "utility functions needed" --actor me@host --key me.key)
./check.sh > run1.txt
R=$(warrant reject "$P" --check check.sh --verdict fail --transcript run1.txt \
      --reason "clause 1: coverage drop" --actor me@host --key me.key)
printf '#!/bin/sh\necho coverage 84%%; exit 0\n' > check.sh
./check.sh > run2.txt
A=$(warrant accept "$R" --check check.sh --verdict pass --transcript run2.txt \
      --actor me@host --key me.key)

warrant why "$A"
warrant verify

Every file used above is created above. The final warnings say that the key-to- actor binding is unverified until a trust configuration supplies that authority; a valid signature is not silently promoted into a valid identity claim. Supplying it is two commands — you decide which key speaks for me@host, and which record the chain starts from:

warrant trust add --actor me@host --key me.key     # writes trust.json
warrant trust root "$P"
warrant verify --settlement --trust-config trust.json   # 0 errors, 0 warnings
warrant verify --settlement --trust-config trust.json --require-bound   # same, and an actor's own unbound key would be an ERR (WRT-018)

The transcripts matter. Without them the accept carries no new evidence and no new outcome fingerprint, and settlement grade reports it as a re-litigation that cites nothing new — a filer who writes a different word is not a new reason (THREAT-MODEL SA-1). warrant never runs check.sh itself: a cmd@v1 verdict is the filer's claim, bound to the script and transcript bytes; only ski@v1 reasons are re-executed by the verifier.

The warrant-verify distribution installs four commands:

• warrant — file, inspect and verify decisions;
• warrant-mcp-server — let an MCP client file its own decisions;
• warrant-mcp — seal calls passing through another MCP server;
• warrant-anchor — batch WarrantIDs into a Merkle anchor.

ski@v1 reasons replay offline through one bundled Σ-GLYPH Book I v0.5 evaluator, pinned by digest and checked before import. Reserved ski@v2 is not admitted by any body version and ships no executable candidate bytes.

The longer walkthrough, including negative controls, is docs/try-this-in-fifteen-minutes.md.

Hardware keys (P-256)

A key that lives in phone hardware (Secure Enclave, StrongBox) signs ECDSA P-256, not Ed25519. Warrant accepts both (WRT-019): Python, Go and Rust verify both; the Python CLI also makes a software P-256 key to try it with. A hardware signer files the same entry shape (alg: ecdsa-p256-sha256, compressed key, 64-byte low-S r‖s):

warrant keygen --alg p256 --out phone.key        # never overwrites an existing file

In Claude Code

A plugin witnesses the permission decisions Claude Code makes about your agent's tool calls: auto-mode denials, and calls that were asked about and then ran. It keeps them as signed records in a local store and never takes part in the decision itself.

/plugin install warrant --marketplace s0fractal/warrant

Then /warrant:decisions lists what was refused, what ran after asking, and which asks have no observed outcome. See integrations/claude-code-plugin/ for what a record proves and what it cannot see.

Machine boundary

For CI, MCP or an agent framework, verify an initialized store with --store-mode --json:

warrant --store ./evidence-pack/.warrants verify --store-mode --json | jq -e '.ok'
warrant-go verify --store-mode --json ./evidence-pack/.warrants | jq -e '.ok'

{"report":"warrant.verify-report@v0","grade":"base","ok":true,
 "records":3,"errors":0,"warnings":1,
 "findings":[{"level":"WARN","subject":"<WarrantID>","message":"..."}]}

--store-mode is part of the safe predicate: a missing or uninitialized store fails closed instead of looking like an empty successful verification. Python takes the store through the global --store; Go takes it positionally.

Consumers may rely on these boundaries:

• ok == (errors == 0), and the error/warning counts equal the corresponding findings;
• warrant.verify-report@v0 is closed: seven top-level fields and {level, subject, message} findings;
• Python and Go agree on normative fields and (level, subject) pairs; message is human prose and may differ;
• grade distinguishes base from settlement;
• the report is unsigned and carries no Warrant authority.

warrant why <WarrantID> --json emits the same walk why prints as one warrant.why@v0 object — every record visited, its status (verified, verify_failed, missing) and its whole body — with why's exit status. It is a convenience, not a SPEC surface: verified there means the name recomputes and the record carries a valid signature by its actor, not that the key is trusted; ask verify --settlement for that.

The normative contract is SPEC.md §11, with schemas/verify-report-v0.schema.json alongside. tools/check_release_surface.py fails CI and publishing when the documented CLI surface is absent from the checkout or built wheel.

Re-execute a real specimen

The Air Canada specimen is an authored reconstruction of the decision record the airline did not have; it is not evidence produced by Air Canada.

pipx install warrant-verify==0.9.0
curl -LO https://github.com/s0fractal/warrant/releases/download/v0.8.0/air-canada-pack.zip
echo '74b36f1d5c7777ea9a3ee240e32f992483a3cd2c0dda0c7d065229c49f1a8249  air-canada-pack.zip' | shasum -a 256 -c
unzip air-canada-pack.zip
warrant --store air-canada-pack/.warrants verify
warrant --store air-canada-pack/.warrants check b423b6a82c3451bfbd75563b39e6391093a64db57941d9247a61a6c620bd997f

The last command re-executes the content-addressed, ATP-bounded reason locally; it does not merely read the filed verdict. The pinned asset, source specimen and portable layout are documented in demos/air-canada/, EVIDENCE-PACK.md and PUBLISHING.md.

Integrate it

As a GitHub Actions gate:

- uses: s0fractal/warrant@v0.6.0
  with:
    store: ./evidence-pack
    version: '0.6.0'

Pin 0.6.0 or newer for domain-separated signatures. The action checks the required capability, emits the machine report and fails on verification errors; see action.yml.

From an MCP client:

warrant-mcp-server --store /abs/path/.warrants
claude mcp add warrant -- warrant-mcp-server --store /abs/path/.warrants

The server exposes filing, store verification and reason inspection with fresh ski@v1 replay. It is deliberately distinct from warrant-mcp, the sealing proxy for somebody else's downstream MCP server. See integrations/mcp-server/ and integrations/mcp/.

Contract map

Surface	Authority
Record format, canonicalization, signatures, replay, settlement and trust config	SPEC.md
JSON schemas	schemas/ — derivative; SPEC remains normative
Portable .warrants/ bundle	EVIDENCE-PACK.md
Writing WPL checks	docs/authoring-checks.md
WPL design boundary	docs/policy-language-choice.md
Threat model	THREAT-MODEL.md
External implementation contract	conformance/
Releases and artifacts	PUBLISHING.md and CHANGELOG.md

The Python reference implements the five filing verbs and settlement. Go is an independent verify/settle implementation. Rust (impl-rs/, the warrant-verify crate) is a from-scratch, dependency-free implementation of the whole warrant CLI — filing, settlement grade, ski@v1 re-execution, Ed25519 signing and verification — held to the reference's output byte for byte by tests/rs_parity.py. All three were produced within one author/model lineage: agreement is conformance evidence, not independent custody or adoption.

python3 impl/warrant.py conformance examples
python3 impl/warrant.py selftest
(cd impl-go && go build -o warrant-go .)
./impl-go/warrant-go conformance examples
(cd impl-rs && cargo build --release)
./impl-rs/target/release/warrant-rs conformance examples
python3 tests/differential.py

A fourth JavaScript candidate was created without Warrant implementation source and reached the complete base grade under a self-certified iterative local-model experiment. Its settlement grade is unimplemented. This establishes bounded implementability against that frozen corpus, not external custody, adoption or correctness beyond it; see needs/README.md.

To test another implementation without executing ours:

python3 conformance/run.py --candidate "./your-verifier probe"   # 166 vectors
python3 conformance/run.py --candidate "./your-verifier probe" --self-check

The runner checks the candidate's typed answer. 83 of the 166 vectors are MUST-REJECT, and the pack reports base and settlement separately. Details and digest verification are in conformance/README.md.

Breaking signature boundary

Since release 0.6.0, signatures cover the domain-separated message

"warrant-sig-v1:" || WarrantID_raw

rather than the bare WarrantID. There is no dual-accept window. Pre-0.6.0 signatures therefore do not verify under the current protocol. Where the old signing key is available, migrate only signatures — WarrantIDs and body links do not change:

warrant resign --key mykey.key --dry-run
warrant resign --key mykey.key

Settlement

Settlement is a separate verification grade defined by SPEC §5.1, §7, §9 and §12. It derives key state and active roots from explicit trust configuration, replays settlement-active reasons, and requires new evidence or a new outcome fingerprint to reopen a subject. Prose alone never reopens settlement.

python3 impl/warrant.py verify --settlement --trust-config trust.json
python3 impl/warrant.py settle <settling-wid> candidate-body.json
./impl-go/warrant-go verify --settlement --trust-config trust.json <store>
python3 tests/settlement.py

An unusable requested trust configuration fails closed; it does not fall back to a clean base-grade report. genesis.json is advisory unless explicitly pinned by the verifier.

The only known format consumer is the sibling sigma-glyph repository. It is a working same-author integration, not outside adoption. Published software, protocol adoption, green CI, self-review and independent validation remain different claims.

License: MIT.