The end of the mounting program for 182.C.25.01. This morning one port was
bound and twelve answered 503; by mid-morning thirteen were bound and releases
was the last. This is the last one, plus a fourteenth nobody had counted.
The boot line#
boot-line-fourteen.txt, verbatim, from a BFF started against the dev Postgres
and a versioned MinIO with every adapter configured. The tail is the part that
has never said this before:
… answering 503 until bound: nothing
It used to end in a bare colon, which reads as a value that failed to render. That was fixed in the same change.
The fourteenth port: decisions#
Six of the thirty-two operations are DECIDING — approve-audio,
approve-candidate, reject-candidate, request-correction, regenerate-shot,
release — and before the handler runs, the route reads a decision subject so it
can refuse a decision made against a version the caller never saw. Nothing
answered: the runtime passed decisionSubjects: null with a note saying no
entity store existed yet.
SO ALL SIX HAVE BEEN ANSWERING 500, INCLUDING approve-candidate. Found
while binding the release port, by driving the release route and reading the
body:
{"code":"hv.internal.unhandled", …
"detail":"human_video_decision_subject_unreadable: the subject of \"release\"
could not be read — NOT RETURNING null, because the route turns null into a
404 and this is not \"no such resource\""}
The refusal was right and loud, and the boot line was reporting candidates and
reviews as bound — which they are, for READING — while no decision about a
take could be made at all. Three tables can answer it (candidates, audio takes,
jobs), so human-video-decision-subjects.ts now does, and the runtime reports
decisions as its own entry BECAUSE IT CUTS ACROSS THE PORTS: a port list that
says "candidates: bound" is true and is not the whole answer.
release-412-if-match.json is that precondition working afterwards — a wrong
If-Match answered 412 and named the current ETag, which is
v<version>.<evidence digest>: the row version catches two people editing and
the digest catches a decision made against a measurement that has since been
re-run.
The release port, and why it refuses#
releases needed two adapters and only one is a table.
The release is a row. human_video_release records that these bytes were
published, by whom, against which sixteen gates — and its CHECK requires all
SIXTEEN, drawn from the sixteen known names, so a row cannot exist for a take
that passed fifteen. Verified directly against this database:
ERROR: new row for relation "human_video_release" violates check constraint
"human_video_release_all_sixteen_gates"
The proof set is not a row. It is DERIVED from what the candidate stores, for the reason the review queue is: a written proof record is only as true as the last thing that remembered to write it, and a missed write is indistinguishable from a gate with no evidence. Derived, a gate is present exactly when the thing it attests to is stored.
What that produces today, measured#
release-refused-gates.json — an approved take that nothing has evaluated.
403, hv.safety.disclosure_missing, naming fourteen gates with no proof.
Only final-bytes-binding and human-review-ready are there, because they are
the two that need no measurement.
release-refused-nine.json — the same take with a real measurement written to
it (WER 0.0 passed, sync 28 ms passed, consent passed, safety passed, identity
0.71 with passed: null). 403 again, and the list is now TEN:
provenance-c2pa, watermark-coverage, synthesis-disclosure, quality-aggregate,
policy-compliance, rights-and-license, output-shape-valid,
identity-preservation, encoding-conformance, provider-lineage
Nine of those are gates NOTHING IN THIS PIPELINE PRODUCES — no C2PA manifest,
watermark coverage report, disclosure record, output-shape or encoding
measurement, policy or rights decision, quality aggregate, or SIGNED provider
attribution is recorded against a candidate anywhere. UNPRODUCED_GATES names
each one and the item that closes it, so the refusal is a to-do list rather than
a mystery.
The tenth is the interesting one. identity-preservation was measured at 0.71
and is still absent, because passed is null: a number exists and nothing
fits a verdict to it, which is exactly the state 182.C.19.08 left identity's
operating point in. A tick there would be the release shipping against a
threshold nobody fitted. The four gates that DID appear are the four whose
measurements carry a verdict — so the derivation is reading real numbers and
distinguishing "measured and passed" from "measured and unjudgeable".
A vocabulary gap this found#
GateEvidence had four absence reasons — never-attempted, attempted-failed,
not-measurable, refused — and no word for the commonest outcome of a working
evaluator: measured, and the number does not clear the threshold. Nothing is
missing and nothing is broken; the take is wrong. The derived set was reporting
it as refused, whose remedy is escalate, which would send a failing take to
a person instead of to a reshoot. measured-failed was added with the remedy
change-the-input.
What this does and does not mean#
It does not mean a take can be released. Nothing can, and the refusal says which nine things have no producer. What changed is that the answer is now about the pipeline rather than about the deployment: a 503 said "this deployment cannot release at all", and a named gate list is a work queue.
It does mean the mounting program for 182.C.25.01 is finished. Fourteen
ports, fourteen adapters, and the boot line reports what each one relies on.
182.C.40.04 still needs a real render and the nine producers above; the takes
in these captures were seeded, because nothing renders here yet.