Every number the README leans on, and whether a reader can reproduce it here.
The repository ran two deployments, both of somebody's real infrastructure. The
identifiers are scrubbed on purpose (scripts/assert_no_estate_identifiers.py),
and the ledgers those numbers were measured from are not in the tree — they are
runtime data, and dataset.jsonl is gitignored. So the honest split is not
"verified / unverified" but two classes:
- Reproducible here — a command in this repository produces the number, and
scripts/reproduce_claims.pyruns it. - Reported — the number was measured on the retired deployments and the source data is not in this repository. The row names the artefact that WOULD back it, so a reader knows exactly what to ask for and what to distrust.
Run python3 scripts/reproduce_claims.py for the machine-readable version of
this table. It fails on a reproducible claim whose command has stopped
agreeing; it never fails on a reported one, because there is nothing here to
run.
| claim | stated in | artefact that backs it | class |
|---|---|---|---|
| hookrelay stays at or under 6,700 source lines | hookrelay/README.md |
scripts/assert_weight.py |
Reproducible |
| hookjudge stays at or under 3,700 source lines | hookjudge/README.md |
scripts/assert_weight.py |
Reproducible |
| every route, knob and containment row is written up or named as deliberately not | docs/containment.md, */docs/reference.md |
scripts/assert_docs.py |
Reproducible |
| the three boards share one design language | the service pages | scripts/assert_design.py |
Reproducible |
| a replayed signed request cannot buy a second verdict | docs/containment.md |
hookrelay/tests/test_signature_replay.py |
Reproducible |
an empty BRIDGE_INBOUND_SECRET refuses to start |
docs/containment.md |
deploy/lark-bridge/tests/test_fail_closed.py |
Reproducible |
| 795 production alerts; 28 of 29 rules answered identically | README.md, hookjudge/eval/README.md |
the retired deployment's webhook_events table — hookjudge/eval/dataset.jsonl, gitignored |
Reported |
| the fold sends 55% fewer cards than the rule as shipped | docs/deployments.md |
a replay against the retired judge ledger (hookjudge/scripts/replay_ledger.py), ledger not in tree |
Reported |
| the first real write happened 2026-09-30 | README.md, docs/deployments.md#the-first-real-write |
the work deployment's run records and audit log | Reported |
kubectl was absent from the work image for ten days |
docs/containment.md, .agents/notes/ |
the release job history and the two nodes' posture records | Reported |
| one HTTP POST from inside a probe reached the chat MCP port | deploy/sidecars/gate.py docstring |
the measurement on the work deployment, 2026-09-30 | Reported |
A project whose most persuasive numbers all come from a private deployment has
two ways to be honest about it: say nothing, or say which ones a reader cannot
check. This page is the second. The alternative — presenting every number as if
it were in the tree — is the failure docs/containment.md names when it says a
boundary trusted for something it never claimed is the way a unit test gets read
as proof about a live model.
If you are evaluating this for real, the Reproducible rows are what you can hold it to today. For the Reported ones, the ask is concrete: the ledger and the run records exist, off this host, and a maintainer can export them.