ARCHIVAL IDENTIFICATION TRIAGE
Last Seen Alive finds clues inside a nameless film fragment, searches historical records, tries to prove itself wrong, and shows an archivist exactly what the evidence can, and cannot, establish.
A five-role Google ADK workflow extracts typed multimodal clues, researches through six Parallel surfaces, re-opens every cited page to confirm the quotation, attempts falsification, then passes structured evidence to a pure deterministic gate with mandatory human approval.
NEW HERE? THREE PAGES, IN THIS ORDER
Then, if you want the rest: run one yourself · call the API · what archivists asked for · how it is built.
THE FAILURE, IN TEN SECONDS
WHO THIS IS FOR
The tool does the broad, repetitive investigation and hands over a dossier. The archivist, not the model, decides whether the evidence is strong enough to act on.
The gate requires three independent domains, two distinct clue families, temporal and entity compatibility, no unresolved contradiction, cited decisive claims, and explicit human approval.
THE SCALE, FROM THE PEOPLE WHO COUNTED IT
The Library of Congress commissioned a census. Of roughly 11,000 American silent feature films, 14 percent, about 1,575 titles, survive in their original format. Another 26 percent of survivors were found abroad.
“The loss of American silent-era feature films constitutes an alarming and irreversible loss to our nation’s cultural record.”
Survival is not the whole problem. A reel that exists but cannot be named is not catalogued, not searchable and not prioritised for funding. The Library runs an annual workshop for exactly this, and it publishes its own hit rate.
| Year | Titles screened | Identified at the event |
|---|---|---|
| 2016 | 142 | 32 · 23% |
| 2017 | 180 | 52 · 29% |
| 2018 | 187 | 56 · 30% |
This is context, not a scoreboard. Their corpus is 187 films chosen because they are hard; ours is five with a sealed answer key. The two are not comparable, five cases support no rate at all, and this project makes no claim to outperform expert archivists. It is cited so a reader knows that the people who do this for a living identify well under half of what they screen, and that anyone promising near-certainty on this task is selling something. See the full register.
TRY IT NOW
No signup. The page mints itself a real API key the moment you press run, and calls the same public endpoints you would. There is no privileged internal path.
An investigation runs five agents and several live Parallel calls, so it takes 61 to 444 seconds, median 304 over 27 measured runs. Nothing is queued or cached, so every run is real.
HOW IT WORKS
Writes down the words and physical clues visible in the fragment.
Gemini multimodal emits a typed clue set; verbatim intertitles remain verbatim.
Searches old publications using the rarest exact words.
Parallel Search API receives two to five literal, diverse probes and returns long cited excerpts.
Checks alternate titles, then counts which archives actually list the film.
Parallel Task API performs multi-hop research against a JSON output schema; Parallel FindAll enumerates the named catalogues that hold it.
Tries to prove each candidate wrong.
Gemini directs fresh Parallel searches for date, entity and attribution contradictions.
Every page a claim quotes is re-opened and checked.
Parallel Extract fetches each cited page; a quotation that is not in the live document is refused by the gate.
Code checks every rule. An archivist must still approve.
A pure Python function exposes every boolean threshold as gate.passed and gate.failed.
THE STACK
Five Google Cloud services and all six of Parallel's product surfaces. Every one names the function that calls it, and reports whether it is reachable right now.
Agent Development Kit sequences the five roles and keeps each output separately inspectable. Gemini on Vertex AI reads the fragment and argues against itself. Controlled generation forces a typed claim structure, because prose cannot be gated. Cloud Run hosts it; Secret Manager holds the credentials.
Search hunts rare literal strings. Task researches holdings against a schema. FindAll names the catalogues that let us describe coverage honestly. Extract re-opens cited pages to confirm quotations. Task Group gives each candidate an independent attacker. Monitor keeps looking after we abstain.
WHAT IT WILL NOT DO
It never says “last surviving copy.” Searching every catalogue you can name tells you where a print is; it can never tell you no other print exists. The only permitted sentence is: “No additional holding was found across these named catalogues as of this search date”, followed by the names. Sentences that break that rule are caught in code and shown to you as violations rather than published as findings.
A fragment without legible, discriminating clues may remain unidentified, and the system will abstain. That is a result, not a failure, and you can leave a standing watch on it.
EVALUATION
Three arms on the same five Library of Congress fragments. A: Gemini alone, no web, no gate. B: those same answers put through the real gate. C: the full system with all six Parallel surfaces live. The baseline was measured first, before the credential existed, so its numbers could not be tuned after the fact.
| Measure | A · Gemini alone | B · + the gate | C · full system |
|---|---|---|---|
| Correct identities surfaced | 0 of 3 | 0 of 3 | 2 of 3 (this pass) |
| False-confident identifications | 0 | 0 | 0 |
| Identifications with no citable source | 7 of 7 | n/a | 0 |
| Answers unstable across repeats | 3 of 3 cases | n/a | 4 of 5 cases moved over 28 runs |
| Median latency | 15 s | 15 s | 383 s (this pass; 299 s over 18) |
Two results carry it. D02: the control returned three different Bray cartoons across three runs, every one at high confidence, none sourced, all wrong. The full system returned Dud Leaves Home (1919), the answer key exactly, on 9 decisive claims across 9 domains, citing the very Library of Congress record the benchmark was built from. D05 arrives mislabelled “Those who pay”; the control replaced it with two different wrong films, the full system surfaced Through the Breakers (1909), the documented correction, with a contradiction claim.
Arm C is a single pass. Six passes over the same five fragments, 28 runs in total, say something quite different, and we would rather you heard it from us.
| Across 28 runs | Count |
|---|---|
| Cases that gave the same verdict every time | 1 of 5 |
| Correct identities | 3 |
| False-confident identifications | 9 (16 on the strictest reading) |
| Candidates that were not a film at all | 7 (now 0, fixed) |
Runs that reached probable | 0 |
Two findings hurt. D02 returned four different leading candidates in four runs: the correct film once, the studio once, and two other films. And D04 returned the same wrong film, Un coin de Paris (1900), in three of four runs at five of seven thresholds: not noise, a reproducible misidentification. “Zero false-confident identifications” is a property of one pass, not of this system, and every surface that quoted it now says so.
We fixed both defects and measured again, and only one of them went away. Candidates that named no film went from 7 to 0. False-confident identifications did not improve. Constraining the candidate type turned disguised failures into visible wrong answers rather than removing them. The error rate was always this high; the non-films were hiding it. Both eras are published at /v1/eval/fix-comparison.
We could have shipped the good pass and said nothing. Publishing this is the only position consistent with everything else on this page. And one thing did hold across all 19: no run ever reached probable, because that threshold needs a human the API cannot supply. Every wrong answer above arrived as candidates with its failing thresholds attached. The defensible claim is not that this system is not wrong. It is that it does not assert what it cannot support, and shows its working.
D04 abstained where the answer key says Buying a cow. The research worked: 13 domains, “LOUVRE” and “TABAC” correctly transcribed. Then ten citations were rejected: eight because the compiler paraphrased instead of quoting Parallel verbatim, two because Extract could not find the quote on the live page. No decisive claim survived, so it abstained. The strictness that stops this system asserting an identity on its own is the same strictness that cost this one. We are not loosening it; a system that accepts paraphrased citations has given up the only thing it offers.
Opened exactly once, after the implementation is tagged eval-freeze-*, with the misses published beside the successes. Hashes are published now so the set cannot be quietly changed. Download the manifest.
PRIOR ART
Our documented search found screenshot identifiers, archive cataloguing products, and crowdsourced identification projects, but did not find a production system combining multimodal clue extraction, live open-web historical investigation, adversarial verification, global holdings research, and claim-level provenance. This is a documented search result, not a claim that private systems cannot exist.
HONEST RISKS
The evaluation fragments have known answers; real orphan material may not. Search coverage is uneven outside digitised English-language sources. Exact phrases can be reused. Human approval is mandatory because the gate measures evidence completeness, not historical truth.