The Bird in Question

Photographed from directly below at the corner of a soffit, against a bright overcast sky. The backlighting flattens the bird to a near-silhouette — the single fact that sent three A.I. providers down different paths. Turn on the field marks to see what the plumage hides.

Five field marks converge on European Starling — and, decisively, rule out the swallows and raptors that the silhouette invites. Below, the case as the panel built it.

Caveat of the field: these are backlit photos shot straight up into the sky. No panel — human or machine — should call this a certainty. The weight of evidence, and the agreement of every provider on one name, is what makes Starling the answer of record.

Twelve agents were dispatched to each of three providers. The Claude cohort inspected the pixels directly; the Codex and Grok cohorts each ran their provider's own vision model through its CLI, relayed faithfully and unedited. The exact wording each cohort received:

You are an expert field ornithologist performing a careful blind identification. Open BOTH image files and visually inspect them. They show the SAME individual bird photographed moments apart, viewed from directly below against a bright overcast sky at the corner of a building's eave/soffit. The bird is partly backlit.

Work the evidence before committing:
• SIZE: compare the body length to the width of one horizontal lap-siding board.
• Silhouette, head and bill shape, tail shape, posture.
• Whether it is holding anything in its bill.
• Habitat/behavior and what it implies about nesting.
Then commit to your single best species. Do not stop at family level unless truly impossible.

$ codex exec -i IMG_4623 -i IMG_4624 \
  "Identify the bird in these two photos of
   the SAME individual bird. It may be a small,
   backlit bird perched at a building eave or
   soffit, possibly holding food in its bill.
   Weigh its size against the siding boards.
   Give the single most likely species as a
   common name, then one sentence of reasoning."

$ grok -p "Examine the two image files
   IMG_4623 and IMG_4624. They show the SAME
   individual bird in two photos, viewed from
   directly below at a building eave against a
   bright sky. Weigh its size against the siding
   boards. Identify the bird species: reply with
   the single most likely common name, then one
   sentence of reasoning."

All three cohorts returned through the same JSON schema — { species_common, species_scientific, confidence, reasoning, raw_answer } — so 36 free-form judgments could be tallied as one ballot.

Thirty-six independent judgments, nine species named. European Starling is both the plurality and the only name that appears in all three providers' returns.

Every observer's return, with its verdict, its confidence, and its token cost. Click any card to read the full reasoning and the raw answer it filed.

What the determination cost, in tokens. The 36 sub-agents run on Claude; the Codex and Grok cohorts additionally spend their providers' own vision-model tokens, billed separately by OpenAI and xAI. Figures for this session are captured before this very page was written.

On the accounting: "Total processed" sums fresh input, cache writes, cache reads, and output. Cache reads are re-supplied context billed at a fraction of fresh input, so the processed figure runs well above what a single pass would cost. "Output" is the tokens each model actually generated — the truest measure of thinking done. CLI figures are each provider's own self-reported usage.