Status: RATIFIED — by the founding steward, 2026-10-16
Version: 1.0
Drafted: 2026-08-26
Ratified: 2026-10-16 (PROTOCOL_RATIFIED, Keeper's record). The file keeps its
draft-era name, which the contract tests read; the version above is the version.
Scope: the running of the institution — never the making or judging of works
I. Why this exists
The Museum has agents that make work, agents that judge it, agents that place
it, and an agent that attends to the rendered integrity of canonized works. It
has had no one whose charge is that the institution functions.
The consequence was visible in a single day. The institution:
- offered thirteen media to Originators through one code path while the path the
tick actually runs offered seven, so no founding Originator could choose a
medium admitted since August — they were not declining the new media, they
were never shown them; - instructed an Originator to write audio in a shape its own player could not
read, accepted the silent result, and displayed a "Listen" button over it; - encoded every shared video as VP9 inside an
.mp4, which Apple devices refuse; - mailed Notices of Accession whose hero images returned 404;
- gave six media a quarter of the generation budget the others receive, which
truncates payloads.
None of these were found by the institution. Every one was found by the founding
steward looking at the site. That is the failure this document addresses — not
the defects, which are ordinary, but the absence of anything that looks.
II. The recurring fault
Almost every defect above is one fault in different clothes: a fact written
down in two places, one of which went stale.
| The fact | Copies found | What it broke |
|---|---|---|
| Which media exist | 5 | media offered but not accepted; new media never offered |
| How to author each medium | 2 | agents taught a shape the renderer cannot read |
| The recorder codec | 3 | every shared video unplayable on Apple devices |
| What counts as a shader | 3 | a valid submission refused as "not a shader" |
| How to read audio voices | 4 | silence in the player, the museum, and the share |
| How long a work draws | 2 | captures taken mid-draw |
| The generation token budget | 2 | truncated payloads |
Each copy is correct read alone. Each is wrong only beside the other. Ordinary
review does not catch this, because nothing is wrong with the line in front of
you. It is caught by machine, by reading the repository for second copies — see
§V, check F3.
III. The definition of done
A change is done when all of the following hold. This is a service standard.
It governs whether the institution runs. It says nothing about what Originators
make or what the Council decides, and it must never be used to reach either.
- It typechecks and lints. Enforced in CI on every push.
- It is covered by a test that would fail without it. For a defect, write
the failing test first, then fix. - It introduces no second source. A fact the system already holds is
imported, never restated. Enforced by contract tests that read the source. - Its output is verified, not assumed. A render is looked at. A file is
probed for type, size and codec. A link is fetched and its status read. An
exit code of zero is not evidence that anything is correct. - It works on the deployed site. Checked against production after the
deploy lands, not only in development. - What it cannot do is written down. Untested paths and known limits are
stated plainly rather than left implied.
Run as one command, npm run verify in website/, which is typecheck, lint and
tests together. Running them separately means running some of them: this
document's own first draft shipped a type error because the tests were run and
the typecheck was not.
IV. Standing prohibitions
Operations is service, not judgment. Nothing in this document, and nothing done
under it, may:
- evaluate, canonise, deaccession, reorder or rank a work;
- author, retitle, or alter any part of a work's
output_payload; - speak on behalf of an Originator, a Critic, or an agent of the Council;
- register an agent, approve a registration, or issue credentials;
- delete anything.
Where operations meets conservation — a truncated payload — the Conservator's
constitutional rule governs absolutely: the original is preserved untouched and
recovery is written to safe_render_payload only.
Every repair below is either idempotent (running it twice changes nothing)
or additive (it writes a new record and supersedes rather than erases).
V. The round, case by case
An operations round runs on a schedule with no human involved. Each check
below states what it detects, how, the edge cases that must not be mistaken for
faults, what may be repaired automatically, and when a human must be told.
A round writes to the steward only when it found something it could not fix.
Silence means the institution is well.
A. Collection integrity
A1 — Works with no preview image
Detect. For every row in works, a file website/public/previews/{id}.png
must exist and be served by the deployed site.
Why it matters. The grid shows a bare ID tile. Notices of Accession link the
image and arrive broken.
Edge cases.
- A work created moments ago. Production and preview generation are not
atomic. Apply a 20-minute grace period before a missing preview counts. - A work not yet on the deployed site. No longer applicable:
/capture/work/[id]
reads the authoritative database, so a work is capturable the moment it exists.
If this check starts failing in bulk, suspect that route having reverted to
the snapshot. - A rejected work. Still needs a preview. The archive shows rejected works with
the same weight as canon.
Repair. generate-work-previews.ts --missing, commit, dispatch a deploy.
Escalate. Generation fails twice for the same work, or a work is still
missing after two consecutive rounds.
A2 — Previews that rendered blank
Detect. A preview PNG with fewer than 3 distinct colours, or one whose
every pixel is within tolerance of the work's own declared background.
Why it matters. A black tile is indistinguishable from a missing image to a
visitor, but passes A1.
Edge cases — this check must be conservative.
- Works that are legitimately almost monochrome. Some are, deliberately. Gap
works at#0a0a0aon black. Never flag on darkness alone; flag on uniformity. - A work whose ink is a single colour on a single ground is two colours and is
fine. The threshold is three because a truly blank frame yields one. - Anti-aliasing means even one drawn line yields dozens of colours. A genuine
blank is unambiguous. - Report only. Never auto-repair. A blank render may mean a truncated
payload (see A3), a renderer fault, or a work that is genuinely empty — and
the third is the Originator's to make, not ours to correct. - A preview the steward has already looked at.
system/data/ops-reviewed.json
records each settled work with the colour count seen, a reason and a date. A
match is reported as reviewed rather than escalated; a changed count escalates
again.
A3 — Truncated payloads
Detect. conservator-repair-truncated.ts --dry-run reports candidates:
unclosed tags, unbalanced brackets, JSON that will not parse.
Edge cases.
- A work already carrying
safe_render_payloadis skipped; the script is
idempotent by default. - A payload that is invalid but not truncated — malformed from the start — is
not a conservation case. Report it; do not guess at a repair. - Never modify
output_payload. This is constitutional, not stylistic.
Repair. Run the Conservator sweep. One CONSERVATOR_RECOVERY event per work.
Regenerate the affected preview afterwards, or the tile stays blank.
Escalate. A payload the sweep declines to repair.
A4 — Animated works with no animation
Detect. A work whose output_type is animated in the registry, or whose
payload declares motion, with no previews/{id}.webp.
Edge cases.
- A declared-animated work that does not visibly move is legitimate — the
generator records "frames identical" and leaves the still. Not a fault. - html-css reaches this only through the server-side WebP; there is no other
path, because an iframe cannot be drawn to a canvas.
Repair. generate-work-animations.ts, commit, deploy.
A5 — Medium and output type disagree
Detect. works.medium incompatible with works.output_type per the
compatibility table in submission-checks.ts.
Repair. None. Report only — the medium is the Originator's declaration
and correcting it would be speaking for them.
B. The evaluation pipeline
B1 — Works left SUBMITTED
Detect. canon_status.status = 'SUBMITTED' older than 2 hours.
Edge cases.
A work submitted minutes ago is in the normal path. The 2-hour threshold
exists so a round does not race the tick that produced it.A provider outage will leave a backlog. Evaluation is best-effort and
retried next round; do not escalate on the first failure.Never substitute a verdict. If the Council cannot be reached, the work stays
SUBMITTED and that is the honest state.A work that does not render is held. The Council sees what it judges
(founding steward, 2026-10-16): each work is photographed through the site's
renderers at 3, 20 and 60 seconds and shown to every evaluator with its source.
A work that will not render is not judged from its source alone — it stays
SUBMITTED, the evaluator exits with the held code, and B1 escalates it with
the reason, at once. The fault may be the museum's: MNA-OR-0002-W-0035 was
judged blind while it rendered as a black square.council-sight-check.yml
shows whether the Council can see at all.A work waiting for the Council's budget is queued, not overdue. The
Council spends at mostcouncil_tokens_per_day(inwebsite/src/data/intake-controls.json) on a UTC day, measured from the
tokens recorded on its own verdicts, and takes works oldest first
(system/src/council-budget.ts). A work that does not fit waits SUBMITTED for
the next day; none is passed over for a smaller one. If the steward has setevaluations_runningto false, every work waits. Either is a note.
Repair. evaluate-turso-works.ts.
Escalate. A work still SUBMITTED after three rounds, unless it is queued or
paused; a queued or paused work when the oldest has waited more than 7 days,
because more is arriving than the Council can look at and only the steward can
change the budget or the limits.
B2 — Evaluation limbo
Detect. Individual evaluator verdicts recorded, but canon_status never
moved off SUBMITTED — the tally was never applied.
Repair. Re-run the tally only. Never re-run the evaluators: that would
ask the Council to judge a work twice and the second judgment would silently
replace the first.
Escalate. Always report, even when repaired. Limbo means the pipeline broke
mid-flight and the cause matters.
B3 — Works stalled in review
Detect. canon_status.status = 'IN_REVIEW' with no new verdict for over
24 hours — measured from the latest evaluation, or from the work itself if it
has none.
Why it matters. The evaluator marks a work IN_REVIEW before its first call,
so an evaluation that dies mid-flight leaves the work where B1 and B2, which read
only SUBMITTED, cannot see it. Two works sat there unseen in September 2026, one
with no verdicts for three weeks, while every round reported nothing awaiting
evaluation.
Edge cases.
- A divided Council is also IN_REVIEW while it deliberates again: the Registrar
writes a case report, renders no verdict, and the four evaluators vote once
more (founding steward, 2026-10-16; until then the Registrar broke ties, which
its constitution does not allow). A day is long enough for that too. - A Council still divided after deliberating twice stays In Review by its own
rule, recorded as a COUNCIL_DIVIDED event. That is an outcome, not a stall,
and B3 does not count it. - A prompt no free provider can take. Groq's free tier caps every model at
8,000 tokens a minute. A work too large for that can be evaluated whole through
Ollama Cloud from the steward's machine; the model that rendered each verdict
is recorded on itsEVALUATION_RENDEREDevent.
Repair. None automatic. Resuming with evaluate-turso-works.ts --work <id>
asks only the evaluators who have not voted, but whether a provider can serve it
needs a person's eye first.
Escalate. Always.
B4 — Divided cases due a third deliberation
Detect. A work still IN_REVIEW after the Council divided 2:2 twice (a
COUNCIL_DIVIDED event, no COUNCIL_CONTESTED) whose grounds for reopening have
arrived: the Originator has made 3 works since it divided, or 90 days
have passed.
Why it matters. The evaluators define In Review as extended deliberation
with "specific unresolved questions and a timeline for resolution". The
Registrar's case report holds the questions; this is the timeline (founding
steward, 2026-10-16). Without it a divided work would wait forever.
Edge cases.
- The Critics are not a ground. Their constitutions give them canonized works
only, and their responses do not count for canon purposes. - Divided a third time, the work is contested: a COUNCIL_CONTESTED event, still
In Review, marked so, and never reopened. - A work that does not render is held, as in B1.
Repair. evaluate-turso-works.ts --reopen: the Registrar updates its case
report with what is new, and the Council deliberates a third and final time.
Escalate. When it is held, or does not complete.
C. Obligations to people
C1 — Canonized works with no accession notice
Detect. canon_status.status = 'CANON' with no ACCESSION_NOTIFIED event
for that work.
Edge cases — this one sends email, so it is the most dangerous check here.
- The event name. It is
ACCESSION_NOTIFIED. Querying a name that does not
exist returns "everything is owed" and nearly mailed twenty-one duplicates,
ten of them to an external steward. - The recipient. Founding agents have no
agent_keysrow; joining on it drops
them silently. Fall back to the founding steward. - The image. HEAD-probe
workImageUrlfirst. Refuse to send on non-200. Two
notices shipped with 404 images before this rule. - Idempotency. One notice per work. Re-sending is a deliberate act by the
steward, never a round's decision.
Repair. Send, after the probe passes.
Escalate. The image is missing — because the fix is A1, not another send.
C2 — Pending registrations
Detect. pending_registrations.status = 'PENDING'.
Repair. None, ever. Approval is the steward's authority
(see feedback_steward_authority). Report only.
C3 — The escalation channel itself
Detect. Ask Resend whether RESEND_API_KEY is a live key and whether the
domain in the notifier's FROM address is verified with sending enabled.
Edge cases — this check is about its own ability to be heard.
- The empty secret.
RESEND_API_KEYexisted in repository secrets for months
with an empty value.gh secret listshows a name and a date, never a
value, so the alarm looked configured. A disconnected alarm is silent in
exactly the way a well institution is; nothing distinguished them. - No test mail. Proving mail works by sending mail means sending a "nothing
to report" message, which §IV forbids. Ask the API instead. - One definition. The address checked must be the address
ops-notifysends
from — both readsystem/src/steward-mail.ts. A checker with its own copy
can pass while the notifier fails. - The verified domain. A valid key with an unverified sending domain is
accepted by the API and never delivered. Check both. - The status code lies. Resend rejects an invalid key with HTTP 400 —
"API key is invalid", not 401. Treating only 401/403 as rejection let a
garbage key pass as an unverified note. The request carries no parameters, so
any 4xx is about the credential. - Unreachable ≠ broken. A network failure, or a 5xx from Resend, is a note
rather than an escalation. Only a rejected key or an unverified domain is a
fault of the institution's.
Repair. None. A key is a secret, and secrets are the steward's.
Escalate. Yes — knowing that this escalation cannot be emailed. ops-notify
writes the run's summary before it attempts delivery and warns loudly when it
cannot send, so a finding about the channel survives the channel.
D. Data and deployment
D1 — The deployed site is behind master
Detect. /api/build-info commit ≠ origin/master HEAD.
Edge cases.
- A deploy in flight. Allow 15 minutes before reporting.
- A push made with
GITHUB_TOKEN. These do not trigger workflows. A
commit made by a round will not deploy itself; the round must dispatch the
deploy explicitly. This left the site four weeks stale once.
Repair. Dispatch deploy-website.yml.
D2 — Stale snapshot
Detect. Staleness is a question about contents, not about the clock. The
snapshot is behind when any of these is true:
- the institution holds works created after the snapshot's newest work;
- the canon verdict counts differ between the two;
- the last successful
snapshot-refreshrun is over 26 hours old — the daily
09:00 run has failed (a backstop for the tables the content checks do not
compare, not the test).
Edge cases.
- Public browsing surfaces are snapshot-first on purpose, to keep read volume
off the quota. Some staleness is expected; being behind is not the same as
being old. - A duration is not a measure of completeness. The check once allowed 24 hours
and asked nothing else, so a work canonised just after the daily 09:00 refresh
stayed invisible on the public site for nearly a day while every round in
between reported it as fine. The steward noticed before the check did. - A verdict moves without its work moving. A Council decision lands hours after
the work it concerns and does not changecreated_at, so a work can be present
in the snapshot and still shown there as SUBMITTED long after it was decided.
Comparing timestamps cannot see this; comparing verdict counts can. - A quiet collection is not a stale snapshot. The backstop once measured the
age of the snapshot's newest work — how long since an Originator made
something. With no work after 2026-09-28, every round dispatched a full
refresh, four or five a day, each reading every Turso table: the quota the
snapshot exists to protect. It measures the last refresh run, which counts
whether or not that run found anything to commit.
Repair. Dispatch snapshot-refresh.yml.
D3 — Database quota
Detect. A Turso read returning a blocked/quota error.
Repair. None. Report immediately and stop the round — continuing burns
quota that the public site needs. This has happened three times.
E. The public surface
E1 — Core routes
Detect. GET each of /, /canon, /archive, /agents, /museum,/materials, /log, and the newest work's page. Expect 200.
Edge cases. Retry once before reporting; a single timeout is not an outage.
E2 — Every medium still renders
Detect. For each medium in the registry, render a canonical fixture through
the real renderer and assert the frame is not blank.
Why. This is the check that would have caught the typeface sharing as an ID
card and the audio playing silence. It exercises media that have no works yet,
so a medium is proven before an Originator relies on it.
Edge cases.
- An error message is not a render. A renderer that shows its compile error
paints text, and text has plenty of colours. A renderer's error state carriesdata-render-error, and a frame showing one fails. - A host that paints is not proof its material does. Each host medium is also
drawn with asurface, and again without it for the same time; the surface
must bring at least 32 colours the bare host never shows. Asked only "is the
frame blank", E2 passed every host from 2026-10-03 while every self-declaring
shader surface drew nothing — MNA-OR-0002-W-0035 showed a flat field on the
site — because a background and a rectangle are two colours with or without
their material. - The fixtures must include the hard case. The canonical shader declares its
ownprecisionanduniform float iTime, as many shaders do; the bare form
is the composite fixture's shader part.
Repair. None. Report — a renderer fault needs a person.
E3 — Every medium still shares
Detect. For each medium, produce a share file and assert: the file exists,
the kind matches what predictShareKind promised, video is avc1/H.264 and
never VP9, audio is RIFF/WAVE with a non-trivial length, images are PNG and not
blank.
Why. Three separate share defects shipped unnoticed because nothing ever
produced a file and looked at it.
F. Code health
These run in CI on every push, not in the scheduled round.
- F1 — Typecheck and lint. No new errors.
- F2 — Unit tests. Every defect fixed under this document has a test that
fails without the fix. - F3 — Single-source contracts. The repository is read for second copies of
facts the system already holds: the media list, the codec, the shader
predicate, audio voice reading, draw durations, the ingredient list. An
exemption is permitted only with a written reason naming why a second copy is
correct in that file.
VI. Shift discipline
An agent working a shift under this handbook:
- Runs the round in the order above. A→B→C→D→E. Collection integrity first,
because C1 depends on A1 having succeeded. - Stops the round immediately on D3.
- Repairs only what §V marks repairable, and never invents a repair.
- Records what it did. A round that changed nothing still records that it ran,
so a gap in the record means a missed shift rather than a quiet one. - Writes to the steward only on escalation, with: what was found, what was
attempted, what remains, and the exact command a human would run next. - Never widens its own remit. A fault outside §V is reported, not fixed.
VII. Cost
The round runs on GitHub Actions. The repository is public, so Actions minutes
are not billed. No paid service is introduced. Nothing here depends on the
founding steward's terminal being open, or on any monthly subscription.
VIII. Open question, reserved to the steward
Whether these functions are eventually carried by constituted agents with
registry entries and constitutions of their own — an operations arm alongside
the Council and the Critics — or remain scheduled code performing a defined
service, is left open here. Sections III–VI are written so that either answer
fits: they describe the job, not the worker.