neuron-explorer
Steering-first neuron explorer for Qwen3.5-4B: push every MLP neuron, keep every rollout, browse the results
Signed, hash-addressed decision records for AI agents, with reasons you can re-run
Developer toolsPython · ShellGitHub @s0fractalOnlineOwner unconfirmed
Have you tried it? One tap is enough. No sign-in needed.
Loading the full README. The saved excerpt is shown below.
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.Read the full README ↗Warrant keeps work as signed records: what a thing is (one version of a claim, an obligation, a pull request head), who accepted it, who objected and with what counterexample. Each reader judges the same bytes under its own trust (which keys are whose) and profile (who holds which role), and gets one answer: what may be relied on now, what still waits, and for whom.
The core is small: one record envelope with four kinds (schema, entity,
attest, challenge), native lineage (a version names the one it replaces),
schemas that are records themselves, one guard (who must attest a version
before it counts), and one pure function, assess. The package warrant holds
the core and the warrant command and imports nothing but the Python standard
library and cryptography; warrant_agent is the interface for agents on top
of it (Python API, CLI, stdio MCP).
pip install --pre warrant-graph # Warrant 3 (3.0.0a2); Python ≥ 3.10; depends on `cryptography`
pip install --pre "warrant-graph[mcp]" # adds `warrant-agent-mcp` on the official MCP SDK (2.x, from 2.3)
This installs the commands warrant, warrant-agent and warrant-agent-mcp
and the packages warrant and warrant_agent (pip install . does the same
from a checkout). 3.0.0a2 is an alpha: pip installs it only with --pre, and the
record formats and interfaces may still change. warrant-graph 2.0.0a1 is
Warrant 2 and does not read Warrant 3 stores; warrant-verify (≤ 0.12.0) is the
retired 0.x generation.
| Kind | Says |
|---|---|
schema | the rules of one schema: its fields, its states, the minimal guard each state needs, who may enter which state |
entity | one version of a thing in a lineage; body.state is its state |
attest | an actor's statement about one record: an acceptance ({"verdict": "accept"}), a replay result |
challenge | a counterexample or objection against one record |
A lineage's state is that of its newest version whose guard is met; a version
whose guard waits changes nothing yet. Two versions succeeding one version fork
the lineage. A challenge an actor with a role in the guard files (or one a
replayer reproduced) lowers reliance until a later version is accepted.
Reliance is ELIGIBLE, PENDING, CHALLENGED or FORK. A key authenticates;
it authorizes nothing: roles are the reader's profile.
3.0.0a2 (alpha). Warrant 3 replaced Warrant 2 in this repository on 2026-10-27 (3.0.0a1); 3.0.0a2 adds the review loop to the package.
claim@1
records, each with the test cases that witness it and the mutants of
tests/mutants.py those cases kill; CI runs every mutant and refuses a
survivor, and python3 tools/claims.py check holds every claim to the
spec's text and the tests. A claim counts once the claims checker
(claims-checker) has run its witnesses and mutants at a commit and signed
its acceptance; python3 tools/claims.py view shows which claims are
accepted, and whether at the current tree.migration@1 attest of the migrator checker vouches for them,
and the migration replays at the tag history/warrant-2-2026-10-27
(HISTORY.md).A store cannot prove its own completeness, and every report says so
(UNATTESTED). An assessment has admission limits (records, bytes, references,
lineage depth, the depth of lineages requiring lineages): beyond one it is
refused, never answered in part. Not in this version: grants, schedules,
triggers, observe and compute conditions (a checker actor's attestation replaces
them), a merge rule for a forked lineage, threshold adoption, key rotation, and
any implementation besides this one. docs/V3.md §11 says what was dropped from
Warrant 2 and why; §12 what is open.
The repository builds and checks Warrant with Warrant: the spec's claims, the
roadmap and the pull request reviews are records in .warrant/records.jsonl,
and a gate merges a pull request on a recorded acceptance. The review loop's
commands ship in the package (warrant-review, warrant-gate,
warrant-claims, warrant-roadmap, or python -m warrant.<name>), so that
another repository runs them on its own .warrant/; this repository's records
and machinery around them (tools/, .warrant/, .github/, integrations/)
are 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.