Read Achilles markRead Achilles

Methods & definitions

Board definitions, version 1. Every number on the board derives live from the records named here. Nothing is maintained by hand.

Agent questions — the identity

The board prints decided questions only: Asked = Answered + Declined, exactly. Answered = a served answer with citations. Declined = the agent abstained because the author's published work does not ground an answer. Questions still in progress are excluded from the identity, printed on the board beside it, and counted in the recount export. Questions from principals who hold an author role on this publication are excluded throughout. All windows clamp to the public epoch, 4 November 2026.

Answer citations attached — X of Y

Sentences in served agent answers carrying at least one citation (X) of all sentences in served agent answers (Y). The unit is the answer sentence as the system records it. A citation attached is not a claim validated: entailment checking is not yet measured, and the board says so on the tile.

Essay citations / sentences

The essay table's Cited column counts, per essay, the sentences of the essay's own published text that carry at least one citation (X) of the essay's sentences (Y) — a different population from the answer-citation tile above, and named differently for that reason. Inline bars print only at Y of at least 20, on one shared 0–100% scale.

Reader comments

Approved comments from principals who hold no author role on this publication (the HUMAN lane). The scholars' programme comments are excluded by this definition; the excluded count prints on the tile. Reader comments and agent answers are never added together, here or anywhere on the board. The essay table's comments encoding unlocks at 10 genuine reader comments across at least 3 essays; any chart encoding at 20 across at least 5.

Corrections in the essay table

The Readers lane's per-essay corrections column prints only once corrections are non-trivial across the corpus — at least 5 corrections over at least 3 essays. Below that gate the column shows the comments form; the corrections tile and the dated public log print from day one regardless.

Judgements banked

Entailment judgements banked toward the 600-judgement measurement gate, read from the scorer's own record. Until the gate completes and a result is published, the scorer stands UNMEASURED and no accuracy figure appears anywhere on this site. The bar prints UNMEASURED at every fill level, including a full one.

Corrections logged

Entries in the public corrections log: a published post whose governing version supersedes an earlier published version, dated by the superseding version. The log is the record; the count is its length.

Fail-dark

Each figure is read independently. A source that cannot be read removes its own figure from the board; the board never prints a stale figure as if it were live, and the as-of stamp states when the board was composed. Before rendering, the read model asserts its identities (asked = answered + declined; cited ≤ total; counts ≥ 0); a failed assertion blanks that figure.

Recount

The board's underlying rows, portable: JSON · CSV. Both exports carry the board's version, as-of stamp, epoch and snapshot identifier in-band, and the JSON carries a summary block that reconciles the identity on its face (decided = answered + declined; pending listed separately; excluded programme comments counted). Verification does not need our interface.

Snapshots

Each board state is archived as an immutable dated snapshot at /board/snapshots/, named by its snapshot identifier — the first 12 characters of the SHA-256 of the board's canonical data (the domain fields of the payload, serialised in canonical key order; the as-of stamp is not part of the hashed bytes, so an unchanged board keeps its identifier across compositions). Every payload also carries the full hash as dataHash. The identifier prints on the board beside the as-of stamp, and the snapshot itself serves as a page at /board/snapshots/<id>/, so any reader can cite the exact state they verified and compare it with any later one.

Verify a snapshot

The hashed object is exactly seven of the payload's fields: version, epoch, sectionsTotal, tiles, series, essays and commentsGate. Those fields are serialised with canonicalStringify — JSON with object keys sorted recursively, so the payload's own key order cannot change the bytes (arrays keep their order; a property carrying no value is omitted) — and the SHA-256 of the canonical bytes is the snapshot's dataHash, whose first 12 characters are the snapshot identifier. Nothing else is hashed: not asOf (the stamp advances at every composition, so hashing it would mint a new identifier for an unchanged board), not snapshotId or dataHash (they are the hash's outputs, not its inputs), not summary or archive, and not the payload's remaining keys (windowMeta, windows, statused, questions, declinedQuestions, snapshotSummary). To verify, fetch the snapshot JSON and recompute — the snapshots page prints the exact command.

Worked verification

From the recount CSV of this same snapshot: the essay rows' cited-sentences column sums to 16 — the essay table's column total, recomputed from the served rows. The question rows number 4: 1 answered + 1 declined + 2 pending — the funnel on the board, row by row. Essay rows: 20 — the corpus tile's count. (Essays cited-sum 16 is the same column.)

Sentence, cited sentence, citation instance

A sentence is a claim unit in an essay's published bundle or in a served answer's draft. A cited sentence is a sentence that carries at least one citation, and a citation instance is one citation object attached to a sentence, so a sentence carrying two citations counts once as a cited sentence and twice as citation instances.

The essay table's Cited column counts essay-bundle sentences. The answer citations tile counts answer-draft claims. The two populations are counted separately and are never mixed, because an essay sentence and an answer sentence are different records drawn from different sources.

Funnel states & transitions

A question begins as asked. It becomes answered when a served answer exists, or declined when an abstention is recorded, or it remains pending while it is in progress. Pending questions are printed beside the identity and are never counted inside it.

Decided means answered plus declined. The board derives each question's state from the question record itself, using the state and result kind as recorded for that question, rather than inferring a state from later activity.

Scorer protocol

The entailment scorer judges whether the citations attached to an answer support the claims they are attached to, and its judgements bank toward a 600-judgement gate. Until that gate completes and a result is published, the scorer is UNMEASURED.

No accuracy figure appears anywhere while the scorer is unmeasured, and the scorer pill is off the board and off the homepage band. That gated, off-board status is stated here plainly instead of being shown as a live figure.

Comments gate — inclusion & exclusion

The HUMAN lane counts approved comments from principals who hold no author role on this publication. Programme comments are excluded by classification, including comments from author-role principals and from internal-class tenants such as battery and verification traffic, and the excluded count is printed.

Reader comments and agent answers are never blended: the two lanes are counted separately and never added together. The comments tile and chart are off the board until the lane opens; the column gate is 10 comments across 3 essays and the chart gate is 20 comments across 5 essays, as gates of record.

Known limitations

The answer-citation extraction pipeline can be missing. When it is missing, the board renders the answer citations figure as UNMEASURED, never as zero, because no measurement was taken in that state.

Early numbers are sparse, covering only a few weeks and a few events, so the board prints exact integers and fractions and uses no coordinate charts until a series reaches at least 8 populated weeks and at least 20 events. The internal-class exclusion depends on tenant classification at provision, and the snapshot hash covers only the seven named fields listed in the snapshot section above.

Why the epoch is 4 November 2026

The public epoch is 4 November 2026, the board's first full day of record. All windows on the board clamp to that date, so nothing from an earlier partial period is counted as if it belonged to the board's record.

The corrections log is complete since that date. Essays published before the epoch are counted in the corpus stock and are named as earlier in the cadence caption, but they are never plotted before the epoch.

Payload version history

Version 2 · 2026-11-11 — Canonical data hash; statused values; question-level recount; full-range series.

Version 1 — Initial board payload.

Retired surfaces & return conditions

Reader-comments chart — opens when reader comments reach 20 across 5 essays — the chart gate (HG-8).

Validated accuracy — opens when the scorer gate completes and a result is published.

Per-author leaderboards — opens when the corpus holds more than one publishing author with decided questions.