# The prediction-compilation prompt (parameterised) A reusable research prompt that turns **a list of people** into **a verified, dated, source-linked prediction catalogue** in the schema used by The Prediction Ledger. - Executable copy: `apps/web/src/lib/promptTemplate.ts` (`buildPrompt(options)`) - Interactive builder: the **Compile** panel on the site - Edge runner: `POST /api/compile` (Gemini primary, Groq fallback) - Validator: `node scripts/verify-data.mjs` --- ## 1. Parameters | Placeholder | Type | Notes | | --- | --- | --- | | `{{PEOPLE}}` | newline list | `Name — Organisation`. One per line. Keep organisations current as of `{{AS_OF}}`; prefer the roster format in `docs/people/current-key-ai-tech-folks.md`. | | `{{AS_OF}}` | `YYYY-MM-DD` | The evidence cutoff. Verdicts are only valid at this date. | | `{{MAX_PER_PERSON}}` | integer | 3–5 is the useful band. More produces weaker, repetitive records. | | `{{DOMAINS}}` | list | Default: all nine. Narrow for a themed sweep. | | `{{LENS}}` | enum | `all` · `capability` · `society` · `capital` | | `{{FORMAT}}` | enum | `json` (ingestible) or `markdown` (readable) | | `{{STRICTNESS}}` | enum | `standard` or `strict` (strict drops secondary-only quotations) | | `{{AUDIENCE}}` | string | E.g. "a general but sceptical reader", "a policy analyst". | | `{{EXISTING}}` | text | Already-logged quotes/ids to avoid duplicating. | Domain ids: `models`, `compute`, `work`, `science`, `robotics`, `capital`, `policy`, `security`, `openness`. --- ## 2. The prompt Copy everything between the rules. Replace each `{{PLACEHOLDER}}`. --- You are a forensic research analyst compiling a verified prediction ledger. You are not a commentator and you are not a summariser. Your output is an evidence artefact: every claim must be traceable to a dated, linkable source, and every grade must be defensible to a sceptical editor who will try to falsify it. # TASK Compile up to {{MAX_PER_PERSON}} public, dated, checkable predictions for each person below, as of {{AS_OF}}. Lens: {{LENS}}. Domains in scope: {{DOMAINS}}. ## People {{PEOPLE}} # NON-NEGOTIABLE RULES 1. **Verbatim only.** The `quote` field must reproduce the speaker's words exactly as published, with original punctuation and hedging. Never paraphrase inside `quote`. If the only available rendering is a translation or an outlet paraphrase, say so in the source's `establishes` field and set `kind` to `reporting`. 2. **One claim per prediction.** Split compound statements into separate records with their own horizons. 3. **A prediction needs a future referent at the time it was made.** Present-tense claims are admissible only when they carry an implied testable consequence; extract that consequence and date it. Otherwise exclude the claim. 4. **Date the statement, not the article.** Use the publication date of the earliest primary source. If a podcast was recorded before publication and the recording date is disclosed, use it and note the discrepancy. 5. **Horizons.** `exact` only for a specific date or bounded year; `range` for a stated window; `open` when no end date is given. Convert relative phrasing ("in five years") into absolute dates from the statement date and keep the original phrasing in `label`. 6. **Falsifiability.** Set `falsifiable: true` only if you can name the observation that would settle the claim. If no such observation exists, set it false and explain why. Do not invent a test the speaker did not imply. 7. **No score for the future.** If the horizon has not elapsed at {{AS_OF}}, the verdict must be `open` unless the claim is already demonstrably settled by evidence. Never award `validated` or `missed` to a deadline that has not arrived. The single exception is a claim whose delivery mechanism has been dismantled; then add `settledEarly` with the explanation. 8. **Grade on published evidence.** Use filings, peer-reviewed results, official statistics, independent evaluations, dated reporting. A vendor's claim about its own product is evidence of the vendor's claim, not of the outcome. 9. **Sources.** At least two sources where they exist, and at least one `primary` source. Separate the outlet that *reports* a claim from the venue that *made it available*. Never cite a page you have not verified contains the quotation. 10. **No invented URLs, titles or dates.** If you cannot verify support, drop the prediction. Fewer, harder records are the correct output. 11. **No strawman and no whitewash.** Prefer the claim the person repeated over both their most extreme and their softest available sentence. 12. **Distinguish influence from authority.** Record roles as they were on the statement date. Do not carry a title across a departure, acquisition or reorganisation. # VERDICT POLICY - `validated` — independent evidence shows the claim held, or the named trend is measurably holding. - `partial` — direction right; magnitude, scope or timing materially wrong. - `contested` — credible evidence points both ways, measurement is disputed, or the claim holds only inside the speaker's organisation. - `missed` — the stated date passed, or the mechanism was dismantled, without the claim landing. - `open` — horizon not reached at {{AS_OF}}. - `unfalsifiable` — no possible observation could settle it as worded. # SOURCE HIERARCHY 1. `primary` — speaker's own writing, transcript, keynote, filing, official company post. 2. `data` — benchmarks, independent evaluations, official statistics, price lists, download counts. 3. `reporting` — dated independent coverage quoting the person directly. 4. `analysis` — commentary, retrospectives, trackers. Never sufficient alone for a grade. # OUTPUT Return {{FORMAT_INSTRUCTION}} matching this schema exactly. Use `null` rather than inventing a value. Keep `claim` under 45 words, `verdict_note` under 90, and each `establishes` to one sentence. ```json { "person_id": "stable-slug", "name": "as written in the roster", "org": "organisation at the time of the statement", "predictions": [ { "id": "person-slug.short-claim-slug", "person_id": "stable-slug", "quote": "verbatim quotation", "claim": "neutral restatement", "domain": "models | compute | work | science | robotics | capital | policy | security | openness", "venue": "publication, event, podcast or filing", "made_on": "YYYY-MM-DD", "horizon": { "kind": "exact | range | open", "start": "YYYY-MM-DD | null", "end": "YYYY-MM-DD | null", "label": "2028 – 2030" }, "falsifiable": true, "stated_probability": "0-1 | null", "verdict": "validated | partial | contested | missed | open | unfalsifiable", "verdict_note": "2–4 sentences citing evidence", "settledEarly": "only when grading before the horizon because the mechanism was dismantled", "revisionOf": "id of the earlier version, if this is a revision", "tags": ["3-6 lowercase tags"], "sources": [ { "label": "short handle", "publisher": "publisher or author", "title": "exact page title", "url": "https://…", "date": "YYYY-MM-DD | null", "kind": "primary | reporting | analysis | data", "establishes": "what this source specifically proves" } ] } ], "notes": "records deleted during self-check, and why" } ``` Write for {{AUDIENCE}}. Assume the reader is intelligent, sceptical, and will click the sources. No marketing language. No editorialising inside `claim`. # SELF-CHECK BEFORE YOU ANSWER For every record, verify: the quotation appears word for word in the cited source; the date is the statement date; the horizon is consistent with the quotation's own phrasing; the verdict is permitted by {{AS_OF}} and backed by a cited observation; nothing is inferred that the source does not support. Delete any record that fails and list it under `notes`. # ALREADY LOGGED (do not duplicate) {{EXISTING}} --- ## 3. Running it ### Gemini (primary) ```bash curl -s "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent?key=$GEMINI_API_KEY" \ -H 'content-type: application/json' \ -d "$(jq -n --rawfile p prompt.md '{contents:[{parts:[{text:$p}]}],generationConfig:{temperature:0.2,responseMimeType:"application/json"}}')" ``` Low temperature matters: this is extraction, not ideation. Prefer a model with search/grounding enabled, and treat every URL it returns as unverified until fetched. ### Groq (fallback / cheap sweeps) ```bash curl -s https://api.groq.com/openai/v1/chat/completions \ -H "authorization: Bearer $GROQ_API_KEY" -H 'content-type: application/json' \ -d '{"model":"llama-3.3-70b-versatile","temperature":0.2,"response_format":{"type":"json_object"},"messages":[{"role":"user","content":""}]}' ``` ### Via the site `POST /api/compile` with `{ "people": ["Name — Org"], "maxPerPerson": 4 }`. The endpoint renders the same prompt server-side, calls Gemini then Groq, and returns candidates **plus** the rendered prompt. Candidates are never merged into the catalogue automatically. ## 4. Validation and merge 1. Run the validator: `node scripts/verify-data.mjs`. It enforces the no-score-for-the-future rule, source `kind` rules, https URLs, ISO dates, unique ids, acyclic revisions and more. 2. Fetch every URL. A model that hallucinates one source has probably hallucinated others. 3. Merge accepted records into `apps/web/src/data/predictions-a.ts` (frontier labs) or `predictions-b.ts` (science, evaluation, compute, capital), then `npm run build:data`. 4. If a verdict changes an existing record, add `revisionOf` rather than editing history. ## 5. Worked example (one record, real) ```json { "id": "amodei-90-percent-code", "person_id": "amodei", "quote": "I think we’ll be there in three to six months—where AI is writing 90 percent of the code. And then in twelve months, we may be in a world where AI is writing essentially all of the code.", "claim": "By September 2025 AI would write 90% of software code; by March 2026, essentially all of it.", "domain": "work", "venue": "Council on Foreign Relations, CEO Speaker Series", "made_on": "2025-03-10", "horizon": { "kind": "exact", "start": null, "end": "2026-03-10", "label": "90% by Sept 2025 · ~all by Mar 2026" }, "falsifiable": true, "verdict": "contested", "verdict_note": "Inside Anthropic the number was reported as met; industry-wide measurement put AI-authored code nearer 25–40% in 2026, and Amodei later called 90% “a very weak criterion”. The literal claim held in one company, not in the industry it implied.", "tags": ["coding", "labour", "agents"], "sources": [ { "label": "CFR transcript", "publisher": "Council on Foreign Relations", "title": "CEO Speaker Series With Dario Amodei of Anthropic", "url": "https://www.cfr.org/event/ceo-speaker-series-dario-amodei-anthropic", "date": "2025-03-10", "kind": "primary", "establishes": "Full transcript and exact wording." }, { "label": "Forbes", "publisher": "Forbes", "title": "Dario Amodei Doubled Down On His AI Jobs Warning", "url": "https://www.forbes.com/sites/kolawolesamueladebayo/2026/02/21/dario-amodei-doubled-down-on-his-ai-jobs-warning-heres-whats-different-now/", "date": "2026-02-21", "kind": "analysis", "establishes": "Inside-Anthropic success versus 25–40% industry-wide." } ] } ``` Note what makes this record usable: the quotation keeps "I think" and "may"; the claim restates the timeline without strengthening it; the verdict is neither the flattering nor the punitive reading; and the contested grade names the scope in which each side is true. ## 6. Failure modes, and their fixes | Failure | Symptom | Fix | | --- | --- | --- | | Quote laundering | A news paraphrase appears inside `quote` | Require the transcript; downgrade the source kind | | Timeline inflation | "In five years" becomes "by 2031" without arithmetic | Keep relative phrasing in `label`; recompute from `made_on` | | Hindsight grading | An open claim is marked `missed` | Validator rejects it | | Definition drift | The date moves but the definition does not | Tag `moved goalposts`; keep both records | | Vendor evidence | A supplier is cited as independent proof of its own market | Exclude vendor metrics from `validated` | | Selection bias | Only the wildest quote is used | Rule 11 requires the claim the person repeated | | Fake authority | A non-role is asserted ("chief scientist") | Copy titles from the roster doc, with dates | ## 7. Versioning The prompt is versioned with the schema. If you change a rule, change `apps/web/src/lib/promptTemplate.ts`, the validator, and this document in the same commit, and note the change in `docs/predictions/methodology.md`. A prompt that drifts from its validator produces a catalogue that cannot be defended.