Proof
Evidence is not a tone of voice. This page separates what you can run now, what this build records, what a document attests, and what those forms cannot establish.
Try the idea first
Run an eight-minute SHA-256 lesson over exact UTF-8 bytes, then change the file and watch the checker reject the mismatch. The exercise makes the useful part of artifact identity visible before this page introduces larger evidence systems.
Check a file yourself →The flagship
Most engineering pages publish the measurement that agrees. The proof repository publishes the one that disagrees, next to the one that does — because a result you can only see when it flatters us is not a result.
The estate's runnable proof repository: an all-integer numerical core that digests identically across processor architectures, with its accuracy gaps on record.
The same repository publishes where that integer pipeline disagrees with the floating-point reference — its agreement against a high-precision comparison and against an independent implementation are printed in the record rather than left out of it, and the figures below are read from that record at build. recorded at build
Published by magogi-proof in its own record, read at build from magogi-proof-results.json — magogi-proof at 4d0e7ec, synced 2026-08-14. The figures are that repository's measurements, not this site's: nothing here was re-run, and nothing was rounded on the way in. recorded at build · 2026-08-14
A published language model, run through an all-integer inference pipeline, produces an identical digest of its output scores on every run. recorded at build · source repositories are private
An all-integer numerical core, cross-compiled and run on two different processor architectures, produces the same digest on both — and the check is a script a stranger runs, not a sentence a stranger reads. recorded at build · source repositories are private
Two legs, each a script rather than an argument. leg-a-cross-arch/verify.sh builds a from-scratch, all-integer numerical core for two different instruction sets, runs the second one under emulation, and passes only if both produce the same digest. leg-b-gpt2/reproduce.py loads a public language model and runs it through an all-integer pipeline whose output digest is identical on every run — and then reports how far that integer result sits from the floating-point reference, in the same document.
$ ./check.shruns in magogi-proof
Until the estate's sources open, that command is a record of what will be runnable rather than an invitation you can accept today — a link that answers not-found would prove the opposite of what this page argues.
The proof repository is a registered member of the federation, and the registry records its language as “proof” rather than a programming language — the field it uses for repositories that ship evidence instead of a package. recorded at build · 2026-09-16
Gates that can fail
These four gate categories are the MGF Python baseline. Each repository declares its own language, conformance level, and executable checks; this list is not a result for every project and does not apply to the sooth family.
The MGF Python baseline names four gate categories: linting, strict type checking, tests with a coverage threshold, and import layering. Each project declares validation suited to its language and conformance level; this is policy rather than a recorded pass, and it is not the sooth family's suite.
The baseline gates are described here as policy thresholds rather than an estate-wide result. Dated local measurements and CI evidence retain their run context; this page does not turn them into a universal current score.
A style and safety linter runs over the source and the tests, and the build stops on a finding rather than filing it for later.
Does not prove that the code is correct — only that it is uniform and free of the patterns the corpus has already ruled out.
Static type-checking runs in its strict mode over the whole source tree: an untyped public function is a failure, not a warning.
Does not prove that the types describe the real world — only that the code agrees with the types it declares.
The suite runs under a committed coverage threshold; the run fails when coverage falls below the number written in the project's configuration.
Does not prove how good the tests are. Coverage is a floor on what was executed, not a measure of what was checked — and the measured value is not published here at all.
An import-layering check enforces the architecture in the dependency graph: a lower layer that reaches up into a higher one fails the build.
Does not prove that the layers are the right layers — only that the code respects the ones the project declared.
This page publishes no estate-wide coverage or pass-history score. The threshold descriptions are policy. Dated gate measurements and CI evidence are retained with their run context; they do not establish one universal current result.
The path a change walks
These are not described from memory: each is committed configuration in this repository, which is why they wear a recorded-at-build chip and the four above do not.
Every change to this repository walks the same path, and the path has gates that are able to fail: type checks, a naming and visual lint, a commit-trailer gate, a secret scan, and an interlock that keeps deployment off ordinary pushes. recorded at build
Every commit must record who authorised it and which model co-wrote it. The gate resolves what to compare against by provenance — a fetched remote reference — because in a pull-request checkout a local branch named after the default branch is the branch under review, not the default branch.
Does not prove that the authorisation is meaningful to an outsider yet: the signature behind it cannot be verified without a published key list.
A pinned scanner reads the working tree and the whole history on every run; the allow-list covers only fixtures whose entire purpose is to look like a secret.
Does not prove that no secret exists — only that the configured scanner found no matching secrets in the scanned scope.
Undefined design tokens, raw colour values, more than one top-level heading in a built page, and type below the legibility floor all fail the build.
Does not prove that the page looks good — only that it cannot silently break the theme, the heading structure, or the reading floor.
Publishing is scoped to an explicit deployment event, and a separate check fails the build if any deploy step is ever re-armed to run on an ordinary push.
Does not prove that a deployment is correct — only that ordinary push, manual, and pull-request triggers are excluded by the configured interlock.
The vendored registries and brand assets are re-hashed at build and compared with their recorded digests. A mismatch is a build failure.
Does not prove that the upstream statement is true today — only that those vendored bytes match their dated source records.
A gate that failed
This is the part of the record that is uncomfortable to publish, which is the reason it is published.
Every commit must record who authorised it and which model co-wrote it; the gate fails closed, and when it cannot resolve what to check it checks what it can rather than passing vacuously. recorded at build
That gate carries its own failure in its source: its documentation records the measured date on which nearly every repository in the estate was running it green while it checked nothing at all. recorded at build
The commit-trailer gate used to resolve its comparison point by name. In a continuous-integration checkout, the reference with the default branch's name is the branch under review, so the range it computed was empty by construction — and the gate reported that all of zero commits carried their trailers, then exited successfully. It had been doing that on every event since it was added.
The repair is visible in the same file: the base is resolved by provenance rather than by name, an unresolvable base raises instead of falling through, and an empty range is treated as reduced scope — it checks the commit being pushed — rather than as a pass. The reasoning was written into the source so the next person to touch it inherits the failure, not just the fix.
A gate that cannot determine what to check MUST NOT report success.
Reduced scope, announced, and a real check.
Authorship
Commit policy requires a cryptographic signature and a written model-coauthorship record. Outside verification still depends on a public allowed-signers list.
Commit policy requires the founder's signature; a signature records authorisation, not who typed the change.
Signing keys are held by the Foundation.
A signature answers who authorised this change — not who typed it. The estate records model co-authorship in the commit itself rather than in a disclosure paragraph, which is why the trailer gate above exists at all.
Verifying one of those signatures from outside needs a published list of the keys that are allowed to sign. That list is not published yet, so this page does not pretend the signatures are third-party-checkable, and it names no date for the list. recorded at build · source repositories are private
Self-audit
Dated repository trackers record findings, declared conformance, and accepted exceptions. The examples below show what those records can reveal without turning them into an estate-wide score.
The estate records its own failures with dates: a conformance tracker that says a repository is below the level it declared, an audit artifact whose waivers expire, and a gate whose documentation records how long its predecessor checked nothing.
The MGF standard defines three conformance levels. A repository's own tracker records its declaration and evidence where that tracker exists; the public registry does not supply a per-project conformance result.
An accepted finding is not a closed finding. Each one is recorded with a review date, printed on every run, and the audit fails if a waiver stops matching a real finding — so a waiver cannot quietly outlive the problem it was written for, and cannot quietly outlive its own relevance either.
Conformance levels are declared per repository and audited rule by rule. One sibling, mgf-http, declares the standard level and its own tracker records that it has not met it — the gaps are listed with the date they were measured. A tracker that can only say “met” is not a tracker.
This repository's own conformance tracker was found describing decisions the repository had already moved past — a licence choice that had since been made, a rule declined “while the deploy path is held” after the deploy path had changed. Self-audits drift; the dates on them are part of the evidence.
The boundary
Named, rather than left for a reader to discover by looking for something that is not there.
This page publishes no completed third-party security audit, comparative benchmark, estate-wide current coverage or pass score, or allowed-signers list for outside signature verification.
The vendored registries and brand assets carry dated source records and recorded digests. The build recomputes those digests and rejects drift. recorded at build · 2026-09-16
Vendored registries and brand assets carry a dated source record and digest. The build recomputes those digests and rejects drift. That establishes byte identity for those specific copies; it does not establish that every upstream statement is true today.
The generated state of everything is on status; the bar the gates enforce is on standards.