Evidence for 182.C.06.05 (detect faces, reject ambiguous identity packs) and
for the decision NOT to unblock 182.C.06.06 the same way.
What was run#
Three fixtures were generated with google/gemini-3.1-flash-image — synthetic
people, so no real person's likeness is used as a test asset — with intended
counts of one, two, and zero. Two cheap vision models were then asked to count
distinct visible faces and return strict JSON.
| Fixture | Intended | gemini-2.5-flash-lite |
gemini-3.1-flash-lite |
|---|---|---|---|
one_person |
1 | 1 | 1 |
two_people |
2 | 2 | 6 |
no_people |
0 | 0 | 0 |
Cost per detection: $0.00019–$0.00030. Latency 1.7–9.0 s.
The disagreement is the finding, and neither model is wrong#
two-foreground-subjects-plus-bystanders.png was inspected directly. It
contains two people in the foreground — the intended subjects — and roughly
four more background office workers at desks, faces small but visible.
So gemini-2.5-flash-lite counted the subjects and gemini-3.1-flash-lite
counted every face in frame. Both are correct readings of an under-specified
question. The prompt asked for "faces clearly visible" and never said whether
incidental bystanders count; the models resolved that ambiguity differently, and
both reported confident: true while doing so.
Two conclusions follow, and they point in opposite directions:
- An exact face count from a VLM is not a stable measurement. It varies with the model and with how the question is phrased, and the model's own confidence flag does not distinguish the two readings. A gate keyed on "exactly N faces" would be keyed on an artefact of prompting.
- The ambiguity signal itself is robust. Both models agreed there was more than one face in the two-subject image and zero in the empty one. "Is this pack unambiguous?" is answerable; "how many faces exactly?" is not.
So 182.C.06.05 is implementable, but the gate keys on ambiguity, not on a
count: more than one face detected → refuse until the creator maps each face
to a cast member. That is what the task asks for, and it is the part that
survives the disagreement.
The fixture is also the real-world case#
An uploaded reference shot in an office contains the subject and bystanders who
never consented. That is synthetic-media:nonconsented-likeness territory
(182.C.03.03), and it is the ordinary case rather than an edge case — which is
why the pack cannot be accepted until a human says which face is the cast
member.
Why this does NOT unblock 182.C.06.06#
182.C.06.06 asks for face visibility, pose yaw/pitch/roll, blur, exposure,
occlusion, crop, resolution, full-body visibility and duplicate scores. A VLM
returns a judgement, not a measurement: asked for a yaw angle it will produce a
plausible number with no geometry behind it, and the probe above shows it will
report confidence while doing so. Recording those as the reference-quality
scores would be the fabricated-metric failure this phase exists to prevent
(F-01). Blur and duplicate detection additionally need a decoded pixel buffer,
which the BFF has no decoder for.
182.C.06.06 stays blocked-local, and so does the 182.C.19 identity gate —
no VLM produces an ArcFace-class embedding, which is what the C4 cosine
threshold compares.
The binding decision this leaves open#
Wiring a hosted detector means every uploaded reference face is sent to a third-party model at ingestion — copy 4 in the data-flow threat model, the copy revocation cannot recall. For references that go on to a generation the exposure already exists; for references the creator uploads and then discards, it is new.
The detector is therefore an injected port that is unconfigured by default. The capability is implemented and tested; binding it to a hosted model is a deployment decision, not a code decision.