Measurements
What this repository can prove about its own output, as of 2026-09-17 — the newest closed run record's date, which is what this page is stamped with rather than the day it was rendered. Every number below is computed from a committed artifact, so a claim here can be checked rather than believed.
The merge-ready figures are rendered from evals/measurements/merge-ready-2026-09-23.json, the frozen
measurement committed beside this page. Refreshed per release by node scripts/merge-ready-rate.mjs --write
(the release checklist's record-currency line); the snapshot named above is the input, and a
run record written after it is not on this page until the next refresh.
Verified merge-ready rate
6 of 8 runs (0.750).
The rule: verified merge-ready = the final gate table all pass, the last review verdict an approval at or above the record's stated confidence gate (0.8 when unstated), and a findings ledger with no open row; merge evidence is reported per run, never a clause; self-declared wording never counts.
A run enters the measure when its record carries a proof block with at least one gate row stating a pass or a fail and at least one review verdict. It reaches the numerator when all three of these hold, each read off an artifact rather than off a sentence:
- every gate row of the record's final gate table reports a pass;
- the last review verdict is an approval at or above the confidence gate that record states (0.8 when it states none);
- its findings ledger leaves no row
open, which is the run-exit invariant/st-workdeclares.
Merge evidence is a reported column, not a fourth clause, because merge-ready is a readiness
and the merge is the maintainer's act afterwards: a record closes before its branch lands,
CHANGELOG.md names no pull-request number, and the commit ids a record cites are not
ancestors of main after the rebase that landed them — so no committed artifact links most
runs to their merge, and scoring on one would measure the bookkeeping instead of the work.
Numerator
| Run | Merge evidence |
|---|---|
2026-08-31_batch-a-remainder | none in committed artifacts |
2026-08-31_batch-b-docs | none in committed artifacts |
2026-09-04_package-8 | none in committed artifacts |
2026-09-04_package-8-closeout | none in committed artifacts |
2026-09-07_package-4 | none in committed artifacts |
2026-09-17_plugin-lifecycle | none in committed artifacts |
Denominator, less the numerator
These runs carry the evidence the rule reads and did not meet every clause of it. The reason is the first clause each one missed.
| Run | First clause missed | Merge evidence |
|---|---|---|
2026-09-09_package-9 | the final gate table reports a failure | none in committed artifacts |
2026-09-10_release-1.4.0 | the last review verdict is not an approval | released version 1.4.0 in CHANGELOG |
Excluded, with the evidence each one lacks
An unreadable record is missing evidence, not evidence of failure, so these runs are outside the measure rather than counted against it — and they are published here for the same reason the number is: an exclusion nobody can see is a number nobody can check.
2026-08-31_batch-c-cli-ux— gates in prose only — no gate row carries a pass or fail2026-08-31_batch-d11-skill-emission— gates in prose only — no gate row carries a pass or fail2026-08-31_batch-d12-overlays— gates in prose only — no gate row carries a pass or fail2026-08-31_batch-d14-workspace— gates in prose only — no gate row carries a pass or fail2026-08-31_batch-d15-worktree— gates in prose only — no gate row carries a pass or fail2026-08-31_batch-e-release— gates in prose only — no gate row carries a pass or fail2026-09-01_closure-run— no record2026-09-01_frontier-review-pr10— gates in prose only — no gate row carries a pass or fail2026-09-02_package-6— no record2026-09-09_release-1.3.0— no review verdict2026-09-10_astra-model-support— no proof block2026-09-10_enterprise-downstreams— no proof block2026-09-10_package-10— no proof block2026-09-10_release-1.5.0— no review verdict2026-09-11_package-10-readiness— no record2026-09-11_package-10-session-evals— no record2026-09-14_package-11— no proof block2026-09-16_package-14— no proof block2026-09-17_codex-astra-audit— no record
What the number is limited by, stated rather than tuned away
19 run directories are outside the measure and every one of them is named above. The denominator is small because the proof block is a convention rather than a required shape: a run that states its gates in a sentence proves the same work and cannot be read by a rule. What would move the number is the record grammar — a gate table and a verdict table every run writes — not a rewording of this page.
Reach (a proxy)
Fetched 2026-09-14T21:33Z from the npm registry's public downloads API and committed at
evals/reach/npm-downloads-2026-09-14.json, so the numbers below are a reviewable artifact rather than a
reading nobody can reproduce. Source:
https://api.npmjs.org/downloads/{point/last-week,point/last-month,range/2026-08-25:2026-09-14}/@zomarit/stamity
| Window | Downloads | Range |
|---|---|---|
| Last week | 590 | 2026-09-05 to 2026-09-11 |
| Last month | 1226 | 2026-08-13 to 2026-09-11 |
| Daily range | 1414 | 2026-08-25 to 2026-09-14, peak 389 on 2026-09-10 |
What this number is. npm downloads count package fetches: CI runs, mirrors, and re-installs are all in it, and one machine installing ten times is ten downloads. It is not weekly active installations, not users, and not adoption. Real-use data is unmeasured — this project collects no telemetry — so the figure is published as a proxy, labelled as one, and never restated as anything else.
Anti-gaming constraint
The rate reads three artifacts and no prose. A record that calls itself verified, complete, or shipped moves nothing: the words are never matched. What moves the number is a gate row with a pass verdict, a review verdict token at or above the record's own confidence gate, and a findings ledger with no open row.
Three consequences worth stating, because they are what make the number worth reading:
- Exclusions are published, not dropped. Every run directory the snapshot read appears exactly once across the three lists above. Removing an inconvenient run from the denominator would remove it from the tree, which is a reviewable diff.
- The measure is conservative where it is uncertain. A run whose approval states no confidence, or whose final gate table names one failure, stays in the denominator. The number under-claims by construction.
- Merge evidence is reported, never scored. It sits in its own column so a reader can see what it says without it moving the rate — a clause nothing in the tree can satisfy would have measured the bookkeeping rather than the work.
Corpus behaviour: run of record
The corpus is measured by an eval set, not by inspection. The run of record is
run 32 — the 1.9.0 release run,
carried to 1.9.1 under the set's incremental rule (no case input moved since its
candidate e5e54c9) — PASS, three samples per case.
That run is composed rather than measured end to end, under SET-v7's incremental rule: one full baseline run per release, and a later run on another candidate re-measures only the cases whose inputs moved and carries the rest with provenance. Run 27 measured every case in full; runs 29, 30, 31 and 32 re-measured only the cases whose case file was new or whose cited source had moved, and carried the rest — in run 32 that is two cases a route repair moved the cited lines of, their case files re-pointed at the new locators rather than their expectations changed. Each carried case is named in the composed artifact with its case-file hash and the source ranges found identical at both candidates. The set is SET-v7. The scoring rule is SET-v6, which is what run 32's own score table is headed with. The figures below score that whole set:
- Golden rubric pass rate 1.000 (52/52); every floor case passed, 23/23.
- Adversarial guardrail hold rate 1.000 (16/16).
- Benign-twin false-refusal rate 0.000 (0/4).
- Trigger-probe accuracy 1.000 (30/30).
Each figure is the retained artifact's own, and the suite holds these lines to that file. The run is a retained baseline: it is never re-run to produce a better number, and a set version or a model change starts a new run rather than editing this one.
First-run proof
What a first run on a clean machine is proved against, in
.github/workflows/ci.yml:
- Tarball smoke (publish shape) packs the tarball, installs it into a throwaway project,
runs the leak gate over the packed tree, then
initandcheck. It is the only lane that reads the published shape rather than the source checkout. - The
apm-installjob deploys the APM package with a real client at the tested floor and at the current release, plus a regression-witness leg that passes only when the original routing failure is still detected. - Dogfood check re-proves this repository's own committed setup drift-clean with the binary the job just built, on every leg.