The easy model dashboard says how many calls succeeded. Ours starts with five preserved research runs: two deterministic studies published, two GLM-5.3 candidates rejected, and one GLM-5.3 candidate approved by the research gate but withheld after visual QA.
That last category is why the dashboard needed to exist. A manifest can say a research run passed. A release note can later say its charts were not fit for readers. If the interface reads only one of those files, it tells a technically defensible but substantively false story.
No single artifact gets to declare success. Final status comes from the manifest, specialist outputs, research gate, release note, and publication record together.
Five layers, one outcome
Manifest
Which run executed, what data it froze, which roles were configured, and which files were hashed.
Role output
Whether each specialist actually ran, what it decided, and how many warnings or blockers it filed.
Research gate
Which mechanical conditions passed, which failed, and whether code allowed the candidate forward.
Release disposition
Whether the reviewed run became a study, was rejected, or passed research review and later failed visual QA.
Call receipt
Exact router attempts, usage, latency, IDs, and cost only when those fields were preserved at call time.
A configured role is not a completed review
Our first online run configured five GLM-5.3 roles through OpenRouter. Only three produced review artifacts. The researcher and editor were never invoked. The correct denominator is therefore 15 configured role opportunities across three online runs and 13 completed outputs—not “three AI-reviewed runs.”
The same early run exposed a second trap: model-authored metadata called itself other model names. The immutable manifest records the configured route as z-ai/glm-5.3; a post-run engineering note explicitly marks the names inside those raw review files as non-authoritative. Later transport code overwrites role, model, and timestamp after validation so the model cannot self-assert its own identity.
A passed gate is not a published article
One later run completed all five GLM-5.3 roles and passed all nine research checks. Its manifest captured that success. Visual QA then found two stale values in visible PNG titles, so the run was withheld and the public article stayed on its earlier deterministic snapshot.
The operations ledger preserves both facts: the research gate passed, and the final disposition was withheld. Collapsing them into one green or red badge would throw away the exact lesson the failure taught us—that statistical review and reader-visible release quality are different control surfaces.
Zero is sometimes the only honest number
All three preserved online runs predate our per-attempt OpenRouter call ledger. We know the configured route and have final role artifacts, but exact attempts, prompt tokens, completion tokens, response IDs, latency, and billed cost were not preserved. So the dashboard says 0 of 3 online runs have exact call receipts.
That does not mean the calls cost zero. It means the bill is unknown. We refuse to reverse-engineer a plausible estimate and style it as an operational receipt. Future online runs will record returned usage and cost after every accepted, schema-rejected, or transport-failed attempt; unavailable provider fields will still remain null.
The ledger is compiled, not hand-curated
A build script discovers every run under the immutable artifact tree, reads its manifest, role outputs, publication decision, optional process note, and optional call ledger, then emits one normalized JSON record. The production build independently regenerates that structure and fails if the committed public ledger differs by even one byte.
Six mutation tests then try to make the public contract lie: relabel a withheld run as published, add an invented cost to a missing receipt, drift an online role away from GLM-5.3, overstate published-study totals, give an offline run a router receipt, or publish through a failed research gate. Every mutation must be rejected before deployment.
What this dashboard still cannot prove
A complete ledger can prove that artifacts reconcile. It cannot prove that GLM-5.3 found every flaw, that the underlying data are true, or that role separation creates independent judgment. All current online specialists share one model family and one evidence bundle. We call that structured adversarial review, not independent human peer review.
Transparency is not posting more JSON. It is designing the public interface so the flattering interpretation is not easier to find than the disconfirming evidence. The ledger will become more valuable as its zeros turn into real receipts—but the zeros belong there first.