Software now changes faster than you can understand it.

One instrument. Four moments.

Not four products. One loop that runs while the agent works, when you accept its change, and long after.

A diff tells you what changed. A record proves what was needed.

A simulated example of how a witness record is built: each hunk is removed in turn and the same evidence runs again. Real runs happen in your environment, on your tests. Nothing runs on this page.

if (isExpired(coupon)) throw new CheckoutError("coupon_expired")

export const isExpired = (c) => c.expiresAt < Date.now()

import { Cart, Coupon } from "./types"
import type { Cart, Coupon } from "./types"

"dayjs": "^1.11.13",

export const now = () => dayjs().valueOf()

console.log("[checkout] coupon", coupon.code)

full change applied. 3/3 pass.
this is where most review stops.

Let agents move fast. Not everywhere.

Protect watches the agent's tool calls through native hooks. It blocks a short list of high-confidence dangerous actions and leaves everything else to your agent's own permission system. It is optional: proof and the ledger work with or without it.

$ dw protect enable --policy standard

policy     behavior
observe    record findings, never block
standard   block dangerous actions
strict     same, and confirm installs

$ dw protect log
receipts keep the decision and the rule,
never raw commands, prompts or source.

What we claim, and what we don't.

The open runner is the only thing that produces a verdict. An optional private planner can choose which experiments to try first. It is advisory, never authoritative.

• we claimEvery verdict comes from executing your own evidence on the real Git patch. No model reviews another model's work.
• we claimThe core runs locally, with no model API, no hosted account and no source upload. It is MIT licensed.
• we claimUnknown stays unknown. A flaky or timed-out run is labeled inconclusive, never witnessed.
• we don'tThat DiffWitness proves your software correct. A weak test suite still produces weak evidence.
• we don'tThat the private planner saves experiments in production yet. It must cut experiments by at least 20% with zero core mismatches on our benchmark before release, and real-repository measurements are still pending.
• we don'tThat DiffWitness replaces review. It tells you where to look first.

Accepted risk shouldn't disappear after merge.

Git keeps what happened. It does not keep why you accepted it. The ledger gives every obligation a stable DW- identity. Accepting one records your reason without deleting it, and it closes when a recheck shows the condition is gone.

Example ledger, illustrative data.

Free proves value. Paid removes friction.

The runner has no paywall. Paid plans are planned to add memory over time and team workflows. Early access opens during the alpha.

Pro and Team are still being designed. Features above are planned and pricing is not set.

Documentation for humans, and for the agents they work with.

github.com/PaititeResearch/diffwitness
├── README.md               overview, install
├── docs/
│   ├── 60_SECONDS.md       first run
│   ├── PROOF_PROTOCOL.md   what proof means
│   ├── PROTECT.md          runtime policies
│   ├── GUARD.md            proof policies
│   ├── DEBT_LEDGER.md      obligations, recheck
│   └── THREAT_MODEL.md     limits and trust
├── skills/diffwitness/
│   └── SKILL.md            for coding agents
└── SECURITY.md             reporting issues