Is the verdict an opinion from a model?

No. It is computed from the evidence a run collected, by rules that give the same answer twice on the same inputs. That is why there are three outcomes and no score: a number invites you to read confidence into it, and there is no model judgement here to be confident about.

Does it work on a codebase that is not Python?

Partly. The mission contract, the executed checks, the evidence from Git, the verdict and the approvals are language-agnostic. CodeGraph is not: it reads Python only, and any other repository produces an empty graph.

Can our whole team use this?

Not as a team product. Reality Graph is a single-operator, single-machine tool with no accounts, no seats and no shared state, so there are no roles, no team approvals and no history a colleague can open. Several people can run it, but each of them runs their own.

Do I have to use all of this to get anything out of it?

No, but these are stages of one run rather than modules you switch on. A mission that demands little produces a thin verdict; one that demands more produces a verdict with more behind it. What you vary is how much a run has to establish, not which parts of the product exist.

What is the smallest useful way to start?

One project and one task. The loop does not need a rollout: your task becomes a checked mission, the approved checks run, and you get a verdict with the evidence behind it. Nothing depends on your team adopting it first, because there are no accounts and no shared state to set up. What it will not do is judge a code base it has never run on. The record starts when you start.

Features

Nine features, one question: what actually happened in that run?

The Reality Graph features cover one loop end to end: state what the change should be, bound the context it gets, run the checks yourself, read the evidence out of Git, and end with a single verdict. Reality Graph never writes the code, which is why it works beside the coding tool you already use.

Request access See how it works

Last updated: 2026-08-11

The capabilities

Each one answers a question a review already asks

Nothing here is a category name. Every card is a question somebody raises at merge time, and the page behind it says exactly how far the answer reaches.

A contract that says what done means

Describe the change in one sentence. Reality Graph proposes the goal, the paths and the checks; you correct and confirm them before anything runs.

Empty acceptance criteria fail validation
Eleven declared intents, never guessed
A pre-run recommendation with no authority to block

Context bounded by that contract

Your coding tool receives the files the contract declares, in a fixed order, with credentials removed before the prompt is assembled.

No retrieval and no similarity search
.env refused in every case
Every token count labelled exact or estimated

A local index you can check

One node per module, class, function and method, anchored to file and line, with a content hash that tells you whether the map still matches.

Python only, other languages index to nothing
Five standard-library imports, no network
Fresh, stale or incomplete, by content hash

Checks Reality Graph runs itself

Approved commands run as an argv list with no shell and a bounded timeout, and the observed exit code is recorded rather than reported.

Executed or attested, with no third value
A missing result blocks instead of passing
No sandbox, and the page says so

One verdict, computed the same way every time

Verified, verified with limitations or blocked, from one deterministic function fed by changed files read out of Git.

Green tests can still end blocked
One call site, so CLI and dashboard agree
Receipts carry a digest that flags later edits

Every run, still open

One local workspace on 127.0.0.1 holding every run, review, check and receipt, rendering exactly what the CLI prints.

Eight views, one canonical projection
The browser never computes a verdict
A port per worktree, and no request off the machine

Every prompt becomes a ticket

Your backlog and your agent's work are the same record: a command runs a ticket, and a prompt without one gets one.

The mission lives inside the ticket
No status moves without evidence
Append-only, so failed attempts stay legible

Four tools your agent gets, ten it does not

A local read-only MCP server over stdio. The dangerous operations are not missing from the contract, they are in it and refused by name.

No port, no socket, no network call
approve_own_patch is declared and denied
Deterministic ranking, no embeddings

Approvals that do not follow the branch

A protected path that changed forces a block nothing clears, and an approval is bound to the run, HEAD, scope and checks it was given for.

Any of those four moving marks it stale
A local attestation, not an identity
Detection after the fact, not prevention

By design

The list an evaluator reads first

This is not modesty. It is the fastest way to find out whether Reality Graph fits, and putting it on the page means nobody has to spend a week discovering it.

You can rely on

A contract validated before the run, with acceptance criteria that cannot be left empty.
Context assembled only from what that contract declares, with credentials scrubbed first.
Checks executed by Reality Graph, with the exit code it observed.
Changed files and diffs read from Git, not from the model's prose.
One deterministic verdict, and a local record that outlives the session.

Not true today

Team features. No accounts, no seats, no roles, no shared history.
CI or pull-request integration. No Git hooks, no GitHub App, no pipeline import.
Test-strength analysis. No assertion counting, no coverage, no mutation testing.
Indexing anything but Python, and no incremental rebuild.
Any certification, and no benchmark or token-saving figure.

Validated on Linux with Python 3.12 and the official Codex CLI. Windows and macOS are unverified in the current release matrix.

How the six pieces fit into one run is walked through on the product page, and what leaves your machine is traced on the data-path page.

Questions people actually ask

Is the verdict an opinion from a model?: No. It is computed from the evidence a run collected, by rules that give the same answer twice on the same inputs. That is why there are three outcomes and no score: a number invites you to read confidence into it, and there is no model judgement here to be confident about.
Does it work on a codebase that is not Python?: Partly. The mission contract, the executed checks, the evidence from Git, the verdict and the approvals are language-agnostic. CodeGraph is not: it reads Python only, and any other repository produces an empty graph.
Can our whole team use this?: Not as a team product. Reality Graph is a single-operator, single-machine tool with no accounts, no seats and no shared state, so there are no roles, no team approvals and no history a colleague can open. Several people can run it, but each of them runs their own.
Do I have to use all of this to get anything out of it?: No, but these are stages of one run rather than modules you switch on. A mission that demands little produces a thin verdict; one that demands more produces a verdict with more behind it. What you vary is how much a run has to establish, not which parts of the product exist.
What is the smallest useful way to start?: One project and one task. The loop does not need a rollout: your task becomes a checked mission, the approved checks run, and you get a verdict with the evidence behind it. Nothing depends on your team adopting it first, because there are no accounts and no shared state to set up. What it will not do is judge a code base it has never run on. The record starts when you start.

Pick the feature closest to the argument you keep having

Every page states what its feature does, and where it stops, in the same breath.

Request access See how it works