# Fabricated-capability audit — 2026-09-12

Evidence for `182.C.40.11`: _leave unsupported modes visibly disabled or
fail-loud; do not close the phase with a plan-only adapter, mocked generation,
supplied metric, or provider marketing claim represented as shipped capability._

Four prohibited things and one required property. Each was checked against the
repository rather than against a module's own doc comment, which is the method
`capability-audit.md` established on 2026-09-09 — where a comment and the code
disagree, the code wins.

## 1. Plan-only adapter — NONE BOUND

`createHumanVideoRuntime(null)` — a deployment with no database at all — binds
exactly one thing and refuses eleven:

```
bound:   preset catalogue
503 until bound: project drafts, idempotency records,
                 upload sessions (no session table), identity packs (no pack table),
                 audio takes (no take table), jobs, candidates,
                 reviews (queue and decisions), releases, erasure, project listing
```

The one thing bound with no database is the preset catalogue, and it has **no
store to be missing**: it serves the product's own vocabulary, the same for
every tenant. Everything else answers `hv.internal.store_unavailable` per
request rather than an empty list, because an empty list from an unconfigured
store is a truthful-looking false statement.

Three ports bind only as **all-or-nothing groups**, and the boot line names the
missing half, because the half-bound states are the dangerous ones:

| port                           | the half-state it refuses                                                |
| ------------------------------ | ------------------------------------------------------------------------ |
| upload sessions + byte sink    | a table with no sink takes bytes and drops them                          |
| identity packs + face detector | a store with no detector accepts a pack nothing examined                 |
| audio takes + probe + aligner  | an aligner with no probe places every end-of-take window against nothing |

No port is bound to something that does not do the work.

## 2. Mocked generation — NOT REACHABLE

`human-video-provider-double.ts` has **no importer outside its own spec** —
checked by grep across `apps/oshun/bff/src`, excluding `.spec.`. Nothing in a
production path can reach it.

`Math.random` appears in exactly two phase-182 modules, both annotated at the
call site and both **retry/poll backoff jitter, never a computed result**:
`human-video-poll-policy.ts:185` and `human-video-catalog-client.ts:466`.

## 3. Supplied metric — NO HANDLER ACCEPTS ONE

No handler on the 32-endpoint surface reads a numeric metric from a request
body. The decision handlers read a `screenDigest` (what the reviewer saw) and
acknowledgement lists — both _decisions_, never measurements.

`submit-audio` refuses a client-supplied alignment outright: the alignment is
produced by a bound aligner, and no aligner means no take, because a sync score
computed against numbers nobody measured would read exactly like a real one.

**One soft spot, named rather than fixed.** `adaptAlignerOutput` writes
`confidence: word.confidence ?? 0`, and zero is the worst possible default for a
confidence — it reads as "measured, and certainly wrong" where the truth is "the
aligner did not say". The contract forces it:
`libs/contracts/src/human-video/alignment.ts:50` requires
`confidence: z.number().min(0).max(1)` with no way to express "not reported". It
is not exercised today — the bound CTC aligner reports a confidence for every
word, and nothing gates on the word confidence (the readers in
`human-video-sync-evidence.ts`, `-take-comparison.ts` and `-sync-scoring.ts` all
read the SYNC evaluator's confidence, a different field). Making it expressible
is a contract change and belongs to whoever changes the contract.

## 4. Provider marketing claim as shipped capability — THE THREE H FINDINGS ARE RESOLVED

`capability-audit.md` (2026-09-09) recorded sixteen findings, three at severity
**H**, defined there as "ships a fabricated or false capability signal". All
three are closed:

| #    | finding                                                                                                                 | state on 2026-09-12                                                                                                                                                                                                        |
| ---- | ----------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| F-01 | `@isis/seedance-provider` substituted a fabricated phoneme track when no lip-sync service was injected                  | **Fails loud.** `lip-sync-validation.test.ts:48` — "fails loud when no lip-sync service is bound, instead of inventing phonemes"                                                                                           |
| F-16 | `@isis/ai-video`'s `ensemble-hdr.ts` fabricated VMAF, tone mapping, HDR validation and EXR export behind optional seams | **Deleted.** The file does not exist; `estimatePerceptualSignals` returned `vmaf: quality * 100`                                                                                                                           |
| F-04 | `apps/psyche/avatar-engine` had four unrunnable Nx targets and zero tests                                               | **Retired from capability claims**, recorded in `apps/psyche/avatar-engine/CAPABILITY-STATUS.md`, whose four facts are re-derived by `tests/test_capability_status.py` so the document cannot drift from what it describes |

## 5. Unsupported modes visibly disabled or fail-loud — HELD

- **API**: every unbound port answers 503 with a code and a sentence, per
  request. Twenty-eight endpoint-level refusals are asserted in
  `human-video-handlers.spec.ts`, including that a read answers 503 rather than
  an empty list and that a submission is refused rather than accepted.
- **Studio**: the two acts that are records decline with the reason attached and
  the refusal names _this draft_, not a missing capability — verified in a real
  browser (`e2e/human-video-studio.spec.ts`, 11 passed). An approval is never
  offered for a take the surface did not read from the service, because there is
  no version to decide against.
- **Admin**: quarantine, cancel and invalidate are off with a stated reason, and
  the panel now shows the reviewer role **the service applied** rather than the
  one the app computed — the two read disjoint scope vocabularies, and the
  screen says so when they disagree.
- **Mobile**: a background upload resume is **refused** rather than sent,
  because `UploadTask` hands the OS a file and not a byte range.

## What this audit does not cover

It is about fabrication, not completeness. Thirty-five items remain unchecked
and their blockers are recorded per item: a consented corpus, a CUDA host, a
V1.2 gate, a TTS decision, a manual screen-reader pass, and the §40 gates that
need a complete phase. None of those is a fabricated capability; every one is an
absence that the surface states.
