How this was made, and what to distrust.

Large language models will happily write a thousand pages about the Peloponnesian War. Whether any of it is true — whether the model was guessing, paraphrasing, or fabricating — is usually invisible to the reader. The output looks the same either way.

This project inverts that. Every year between 3200 BCE and 2025 CE was researched through a single, locked research prompt that forces the model to declare its own confidence, cite its own sources, and record the evidence that would disconfirm its claims. The output is a structured JSON record. Not prose. A ledger.

The frontend you are currently reading exists to make that ledger legible. Every certainty tier is rendered visibly. Every disconfirming note is kept adjacent to the claim it qualifies. Gaps in the record are drawn as gaps — not smoothed over.

Each year JSON is validated against a strict schema called ICCRA. If a year fails validation, it is re-researched; nothing is hand-patched in the corpus files.

Categories are restricted to twelve; certainties to five. Any drift — a compound like political | military— is normalised by a migration script, never by the frontend pretending it didn’t happen.

The production run uses Claude Sonnet 4.6 against the Anthropic API directly. Two alternate models were trialled and archived:

Both archives live in outputs/haiku_experiment/ and outputs/gemini_experiment/ — kept, not deleted, so the model-choice justification has receipts.

Layer 1 is the 5,226-year year-level corpus. Unit: one year. Question: what happened in year N, and how confident are we. Shipped 2026-04-13.

Layer 2 is a scholarly-evidence deep-dive per era, produced through the Scite MCP against peer-reviewed literature with retraction enforcement. Unit: one era (10–400 years). Question: what does the peer-reviewed record say about the core claims of this period. Currently covers eras 01–20 of a planned 50, with zero hallucinated citations across 10 agent runs.

Read the Scite MCP whitepaper →

• Fill empty years with plausible-sounding prose. If a year has no documented events, the page shows the era context and says so.
• Smooth legendary or traditional material into confirmed material. Those tiers are labelled in the margin, not rendered as body claims.
• Hide disconfirming evidence behind a click. It renders adjacent to the claim it qualifies, with an oxblood rule.
• Guess at enum values the model produced outside of the schema. Unknowns render as unknowns, not as a default.

You will see a small strip at the bottom of every page. It states how many events are on the page, how many sources back them, how many are contested, and when the underlying research was produced. It is the transparency argument made ambient.