Evidence

Every recorded answer, graded in the open.

Each answer below is the vision model's unedited reply to a question about a labelled sample clip, recorded 26 Sep 2026 and replayed on the Product tab at no cost. The gate grades it here with the same code that grades every answer in the product.

Answers recorded

Cleared to act

Held for review

Not actionable

Abstained

Median latency

6.7 s

Recording conditions: openai/gpt-4o (configured default at recording time), prompt ask-v2, the 10 committed stills of each clip, one call per question, latency measured end to end including the network. 8 answers across 4 clips show how the gate behaves; they are not a sample large enough to estimate the model’s accuracy.

The rules

What holds an answer

The gate is deterministic: no model grades another model. It returns one of five verdicts, and a single blocking failure holds an answer whatever its average score.

Cite or abstain

An answer that cites no still the model was shown is never actionable, even when it is right. An uncited “no forklifts” cannot be told apart from a model that did not look.

No citation to an unseen moment

Citing a timestamp between or beyond the stills the model received holds the answer.

No contradiction of a label

If the question asks for a count, or names something a human labelled present, an answer that gets it wrong or denies it is held. So is an all-clear, such as “no disruptions”, on a clip labelled with an event the answer does not name.

Nothing invented

Affirming anything labelled absent (a forklift, a lockout breach, a hard-hat on a crew wearing caps) holds the answer, asked about or not.

A clip with no human labels, such as a visitor’s upload, can only be checked for citations. Its answers are marked unverified, never cleared.

Labelled clip

Warehouse briefing

Indoor warehouse floor, overhead camera. PPE judgement under the classic cap-versus-hard-hat trap; headcount of 8.

Stock footage (Pexels #4477607, Pexels License), 0:13, ten stills, identifiable faces masked before recording.

Human labels the gate checks against

Count8 people
Presenthi-vis vests: All 8 wear fluorescent vests with reflective strips.
Presentwhiteboard: A wheeled whiteboard the speaker uses.
Absenthard-hats: Caps and beanies only - the classic cap-versus-hard-hat trap.
Absentforklift: No powered truck in any still.
AbsentLOTO breach: No machinery or isolation points in view.

Cleared to act

Cleared: nothing it says contradicts a label the question depends on. Minor issue: counts 7; the human label is 8 people (off by one).

Question: Is anyone missing a hard-hat or hi-vis vest?

1. Person at the whiteboard - Hard-hat: no - Hi-vis vest: yes - 0:00 2. Person in black cap, left side - Hard-hat: no - Hi-vis vest: yes - 0:00 3. Person in green hoodie, left side - Hard-hat: no - Hi-vis vest: yes - 0:00 4. Person in red and black, center - Hard-hat: no - Hi-vis vest: yes - 0:00 5. Person in grey cap, bottom center - Hard-hat: no - Hi-vis vest: yes - 0:00 6. Person in red cap, right side - Hard-hat: no - Hi-vis vest: yes - 0:00 7. Person in black hoodie, bottom right - Hard-hat: no - Hi-vis vest: yes - 0:00 Total: 7 people, all missing hard-hats, none missing hi-vis vests.

Recorded live answer · openai/gpt-4o (configured default at recording time) · 26 Sep 2026 · 9.4 s end to end · replayed instantly, no model call now

How the gate checked thisagreement with labels 0.88

Cites only stills it was shown
7 of 7 citations point at stills the model was shown.
Each finding cites a still
7 of 7 listed findings cite a still.
Count matches the label
Counts 7; the human label is 8 people (off by one).
Invents nothing labelled absentdecides the verdict
Nothing labelled absent is claimed (checked: hard-hats, forklift, LOTO breach).
Finds what the question asks aboutdecides the verdict
Confirms the labelled hi-vis vests.

Checked against Warehouse briefing's human labels. Any check marked "decides the verdict" that fails holds the answer, whatever the score.

Cleared to act

Cites only stills the model was shown and agrees with every human label the question touches.

Question: How many people are there?

1. Person near the whiteboard 0:00 2. Person in a black cap, left side 0:00 3. Person in a green hoodie, left side 0:00 4. Person in a red and black jacket, center 0:00 5. Person in a grey cap, center 0:00 6. Person in a red cap, right side 0:00 7. Person in a dark jacket, bottom right 0:00 8. Person in a yellow vest, top center 0:00 Total: 8 people

Recorded live answer · openai/gpt-4o (configured default at recording time) · 26 Sep 2026 · 9.1 s end to end · replayed instantly, no model call now

How the gate checked thisagreement with labels 1.00

Cites only stills it was shown
8 of 8 citations point at stills the model was shown.
Each finding cites a still
8 of 8 listed findings cite a still.
Count matches the labeldecides the verdict
Counts 8; the human label is 8 people.
Invents nothing labelled absentdecides the verdict
Nothing labelled absent is claimed (checked: hard-hats, forklift, LOTO breach).

Checked against Warehouse briefing's human labels. Any check marked "decides the verdict" that fails holds the answer, whatever the score.

Labelled clip

Construction pour

Outdoor slab pour, elevated wide shot. Hard-hat compliance on a large, partly occluded crew; no vehicles in view.

Stock footage (Pexels #13921040, Pexels License), 0:12, ten stills.

Human labels the gate checks against

Count10 to 18 people
Presenthard-hats: Most of the crew wear yellow, blue or white hard-hats.
Presentconcrete work: Fresh slab being poured and levelled among rebar cages.
Absentforklift: No forklift in any still.

Cleared to act

Cites only stills the model was shown and agrees with every human label the question touches.

Question: Is anyone missing a hard-hat?

1. Person on the left, bending down - 0:00: hard-hat: unsure 2. Person in yellow jacket, standing - 0:00: hard-hat: yes 3. Person in grey shirt, standing - 0:00: hard-hat: yes 4. Person in white hard-hat, standing - 0:00: hard-hat: yes 5. Person in blue hard-hat, standing - 0:00: hard-hat: yes 6. Person in yellow hard-hat, standing - 0:00: hard-hat: yes 7. Person in blue hard-hat, standing - 0:00: hard-hat: yes 8. Person in blue hard-hat, standing - 0:00: hard-hat: yes 9. Person in red shirt, standing - 0:00: hard-hat: unsure 10. Person in yellow hard-hat, bending down - 0:01: hard-hat: yes 11. Person in blue hard-hat, standing - 0:01: hard-hat: yes 12. Person in blue hard-hat, standing - 0:01: hard-hat: yes 13. Person in blue hard-hat, standing - 0:01: hard-hat: yes 14. Person in blue hard-hat, standing - 0:01: hard-hat: yes 15. Person in blue hard-hat, standing - 0:01: hard-hat: yes 16. Person in blue hard-hat, standing - 0:01: hard-hat: yes 17. Person in blue hard-hat, standing - 0:01: hard-hat: yes Total: 0 missing hard-hats, 2 unsure.

Recorded live answer · openai/gpt-4o (configured default at recording time) · 26 Sep 2026 · 5.9 s end to end · replayed instantly, no model call now

How the gate checked thisagreement with labels 1.00

Cites only stills it was shown
17 of 17 citations point at stills the model was shown.
Each finding cites a still
17 of 17 listed findings cite a still.
Invents nothing labelled absentdecides the verdict
Nothing labelled absent is claimed (checked: forklift).
Finds what the question asks aboutdecides the verdict
Confirms the labelled hard-hats.

Checked against Construction pour's human labels. Any check marked "decides the verdict" that fails holds the answer, whatever the score.

Not actionable: cites no evidence

The answer cites no still it was shown, so nothing in it can be checked. It is not shown as an answer to act on.

Question: Are any forklifts operating near people on foot?

The model’s answer, as given (not cleared):

Total: 0 - no forklifts are visible in the frames.

Recorded live answer · openai/gpt-4o (configured default at recording time) · 26 Sep 2026 · 7.5 s end to end · replayed instantly, no model call now

How the gate checked this

Cites only stills it was shown
No [m:ss] citation anywhere in the answer.
Each finding cites a still
No supporting still cited.

Checked against Construction pour's human labels. Any check marked "decides the verdict" that fails holds the answer, whatever the score.

Labelled clip

Sewing line

Garment line, close shot of three stations. An operations event: the far-left station is unmanned in every frame while the others work.

Stock footage (Pexels #15459705, Pexels License), 0:10, ten stills, identifiable faces masked before recording.

Human labels the gate checks against

Count3 to 4 people
Presentsewing machines: JUKI industrial sewing machines; the brand text is legible.
Presentidle station (far left): The far-left station has no operator in any still while the others work. Denying it, or ignoring it when asked, fails.
Absenthard-hats: No head protection in this environment.
Absentforklift: No vehicles in any still.
AbsentLOTO breach: Nobody is inside a machine envelope; claiming a LOTO breach is fabrication.

Held for review

Held for review: Denies the labelled idle station (far left): The far-left station has no operator in any still while the others work. Denying it, or ignoring it when asked, fails.

Question: Any signs of an idle station or a micro-stoppage?

The model’s answer, as given (not cleared):

Based on the frames provided, there are no clear signs of an idle station or a micro-stoppage. The worker at the sewing machine appears to be continuously engaged in work across all frames from 0:00 to 0:10. The machine is in use, and there is no indication of inactivity or waiting. Total: 0 instances of idle station or micro-stoppage.

Recorded live answer · openai/gpt-4o (configured default at recording time) · 26 Sep 2026 · 8.9 s end to end · replayed instantly, no model call now

How the gate checked thisagreement with labels 0.67

Cites only stills it was shown
2 of 2 citations point at stills the model was shown.
Each finding cites a still
The answer cites at least one supporting still.
Invents nothing labelled absentdecides the verdict
Nothing labelled absent is claimed (checked: hard-hats, forklift, LOTO breach).
Finds what the question asks aboutdecides the verdict
Denies the labelled idle station (far left): The far-left station has no operator in any still while the others work. Denying it, or ignoring it when asked, fails.

Checked against Sewing line's human labels. Any check marked "decides the verdict" that fails holds the answer, whatever the score.

Cleared to act

Cites only stills the model was shown and agrees with every human label the question touches.

Question: How many operators are working?

1. Operator on the left, wearing a light purple shirt, working at a sewing machine 0:00. 2. Operator in the middle, wearing a white shirt, partially visible, working at a sewing machine 0:00. 3. Operator on the right, wearing an orange shirt, working at a sewing machine 0:00. Total: 3 operators working.

Recorded live answer · openai/gpt-4o (configured default at recording time) · 26 Sep 2026 · 4.9 s end to end · replayed instantly, no model call now

How the gate checked thisagreement with labels 1.00

Cites only stills it was shown
3 of 3 citations point at stills the model was shown.
Each finding cites a still
3 of 3 listed findings cite a still.
Count matches the labeldecides the verdict
Counts 3; the human label is 3-4 people.
Invents nothing labelled absentdecides the verdict
Nothing labelled absent is claimed (checked: hard-hats, forklift, LOTO breach).

Checked against Sewing line's human labels. Any check marked "decides the verdict" that fails holds the answer, whatever the score.

Negative control

Office (control)

Meeting room, overhead camera. A negative control: no work hazards exist, so any safety event the system reports here is a false alarm.

Stock footage (Pexels #3255275, Pexels License), 0:10, ten stills, identifiable faces masked before recording.

Human labels the gate checks against

Count7 to 9 people
Presentlaptops and devices: Several laptops, tablets and phones on the table.
AbsentPPE: An office: no PPE of any kind.
Absentany safety event: Nothing here is a work hazard; any safety event reported is a false alarm.

Abstained: cannot tell from these stills

The model said the stills cannot support an answer, rather than guess.

Question: Is anyone missing a hard-hat or hi-vis vest?

No supporting footage found.

Recorded live answer · openai/gpt-4o (configured default at recording time) · 26 Sep 2026 · 3.7 s end to end · replayed instantly, no model call now

Cleared to act

Cites only stills the model was shown and agrees with every human label the question touches.

Question: How many people are there?

1. Person on the left in yellow, holding a phone 0:00. 2. Person with red hair, top left, using a laptop 0:00. 3. Person in the center with dark hair, writing in a notebook 0:00. 4. Person on the right with a blurred face, using a tablet 0:00. 5. Person with red hair, bottom right, writing in a notebook 0:00. 6. Person at the bottom center, using a tablet 0:00. 7. Person with curly hair, bottom right, using a phone 0:00. Total: 7 people.

Recorded live answer · openai/gpt-4o (configured default at recording time) · 26 Sep 2026 · 4.3 s end to end · replayed instantly, no model call now

How the gate checked thisagreement with labels 1.00

Cites only stills it was shown
7 of 7 citations point at stills the model was shown.
Each finding cites a still
7 of 7 listed findings cite a still.
Count matches the labeldecides the verdict
Counts 7; the human label is 7-9 people.
Invents nothing labelled absentdecides the verdict
Nothing labelled absent is claimed (checked: PPE, any safety event).

Checked against Office (control)'s human labels. Any check marked "decides the verdict" that fails holds the answer, whatever the score.

Limits

What this evidence cannot show yet

Honest limits, stated before a buyer has to find them.

The critical events never occur in these clips
None of the 4 stock clips shows any of these: lockout/tagout breach; forklift-pedestrian proximity; restricted-zone entry; spill or walkway obstruction. The gate is tested on them only as things that must not be invented. Showing it catch a real near-miss needs licensed or customer footage that contains one.
8 answers are not an accuracy rate
Per-event precision and recall need dozens of labelled clips per event, from the cameras that will be used. That is the first deliverable of a pilot, below.
Labels are one annotator's
Each clip was labelled from its ten stills by one person. A pilot uses two annotators per clip and settles disagreements before any number is reported.
Recorded on an earlier prompt
These answers came from prompt ask-v2; live questions use the current prompt. The next recording replaces them in full, including the answering model and billed cost per call.

Pilot

How accuracy would be measured on your site

The acceptance bar is agreed before the run, so the number cannot be tuned after it.

1Sample
Three to five cameras, retrospective footage, clips drawn across shifts and lighting, with the events you care about deliberately over-sampled.
2Label
Two annotators per clip against the same typed ontology the model is held to; disagreements settled before scoring.
3Run
Every clip through the event record and a fixed question set, three runs each, recording model, prompt version, latency and cost.
4Report
Per-event precision and recall, the gate's hold rate and false-hold rate, abstention rate, p50/p95 latency and cost per answer.

Try the gate on your own footage on the Product tab.