Part 2 of three on Blindspot. Part 1: reading the video · Part 3: cost, queues, and delivery.
I expected evidence retrieval to be the main quality problem. The first measurements pointed one stage earlier.
On a single Reel, a nano extraction model found 1 of 7 material claims. A stronger model found all seven, and flagged the central framing problem on top: the Reel presented a contested claim still under inquiry as conclusively established.
No amount of excellent research recovers a claim that never entered the evidence request.
Rule 1: a Reel is not a spoken article
The naive design treats a Reel as an article read aloud. Download, transcribe, fact-check the transcript.
But meaning is split across speech, on-screen text, the caption, charts visible for one second, the choice of footage, the publication date, and editing that changes which comparison the viewer sees.
A clip can be authentic while its caption supplies a false location. A number can be correct while the denominator is missing. One Reel can hold seven factual claims, two opinions, and a causal allegation no cited source establishes.
So extraction is one multimodal structured call over transcript, caption, post metadata, and keyframes together, returning strict JSON: checkable claims, on-screen OCR, named entities, events worth retrieving on, and observable framing indicators.
Parsed with JSON.parse, never regex. Claims and framing stay in separate fields. “The payment was ₹10 crore” is checkable. “This proves negligence” is a different kind of claim, extracted only if the Reel actually asserts it as fact.
Rule 2: decide what not to spend money on
Most Instagram posts don’t need a factual investigation. Music, aesthetic edits, jokes, vlogs, and pure political opinion should not be forced into a fact-check verdict.
A cheap triage stage runs after transcription and picks one of three routes:
A no checkable claim ──► deterministic "nothing to check" report
B opinion / framing ──► observable framing only, skip retrieval
C factual claims ──► extraction, retrieval, synthesis
Routes A and B are safety decisions before they are cost decisions. A system that always returns a decisive factual verdict will manufacture one for input that doesn’t support it.
Rule 3: retrieval never gets a verdict
Evidence is another provider seam: built-in web search, an external facts corpus of ~14,000 documents, or both. Blindspot batches every extracted signal into one request, and the service returns passages with provenance: source, URL, publication date, and which claim or keyframe they address.
It is explicitly not allowed to decide whether the Reel is true.
Blindspot Facts service
--------- -------------
read media, extract claims ───► retrieve relevant passages
send keyframes ───► match known images
synthesise report ◄─── return facts with provenance
assign verdicts never assigns a verdict
The point is debuggability. When a conclusion is wrong I can tell whether extraction missed the claim, retrieval returned weak evidence, or synthesis overreached. One end-to-end answer hides all three behind fluent prose.
Rule 4: tell the model what actually happened
An early report claimed Instagram could not be accessed. Plausible-sounding limitation. Completely false. The pipeline had downloaded the video, extracted audio, read keyframes, and produced a transcript.
What had actually happened was that web research found no corroboration. The model turned “search returned nothing” into “the media could not be fetched”, and filed non-evidence notes as sources.
Synthesis input now carries mediaAcquired and keyframeCount, and the prompt states plainly what the pipeline completed. Missing external evidence and missing media are different sentences. Evidence entries must contain URLs; a null search result is represented as no external source found.
Give the model operational state. Don’t expect it to infer operational state from incomplete context.
Verdicts are categories, not a score
No truth percentage. A single number implies precision the evidence doesn’t have and erases disagreement between claims inside one Reel. Outcomes are supported, mixed_evidence, misleading_context, unsupported, satire, not_yet_verifiable.
Every factual conclusion points at cited evidence, and confidence is shown separately from the verdict. A supported claim resting on indirect reporting is not the same object as one backed by a primary document.
The system is allowed to say it can’t verify something yet. Thin sourcing and developing events are normal input, not exceptions.
Next: what it costs, and the parts that break in production.