Files
paperspec/spec/08-fidelity.md
T
alexmandClaude ea0a7efa78 feat: zipped packages, target set, checklists, audit fixes and a stress test
Features for the next six months (accessible HTML export, paperpack and
Lode on packages, the personal bundle):

- Zipped packages: the oracle writes and reads `.sct.zip` archives under
  the rules of spec/12 §12.2; validate.py and load.py accept them; a
  byte-identical zip of the tiny package joins the fixtures, and two
  broken archives (traversal, symbolic link) join fixtures/invalid.
- The target set of targeted F3 (spec/08 §8.10) is computed by the
  oracle and pinned by vectors/targets.jsonl under each digit rule; the
  rule now covers cells as well as sentences and the leaves of anchored
  containers, and defines a node's section.
- scienting-checklist/1 (provisional): a conformance check of an export,
  covered by a receipt as ["checklist", id, hash] (spec/08 §8.11), with a
  WCAG 2.1 AA fixture and its receipt.
- QA reports are signed like receipts (spec/08 §8.9).
- A `description` annotation body with role alt or long, so an
  accessible export can find a confirmed figure description; the layer
  fixture carries one, and the store's body-kind constraint allows it.
- scienting-spot-check/1, the seed payload, is registered with a schema.
- exercise, submission and grade are provisional until a course has
  used them.

Audit fixes, each with a vector or a broken package where one was
missing:

- Folder packages: symbolic links are refused on validate and load,
  manifest paths must not collide under case folding, listed files are
  matched by exact name, files are hashed in chunks under a size limit.
- Files are compared byte for byte, so CR LF, a byte order mark, NaN,
  a lone surrogate or invalid UTF-8 is refused instead of translated;
  integers beyond 2^53 are read as doubles, as canonical JSON writes them.
- Receipts: signature.key must equal by and the key must be valid when
  signed; a qa receipt's nodes must be the sample at the original's
  hashes, its scope must name qa.of, its errors must be sampled nodes;
  statuses are computed once, iteratively, so chains of spot-checks
  hundreds deep settle; Base64 and manifest shape errors become refusals
  instead of tracebacks; counts and the fidelity of roster-less packages
  are recomputed.
- The gate enforces rule 11 of spec/11 §11.4: changed nodes are
  validated against the node schema and the references between nodes
  are checked, so a set to a bad value, an insert with unknown
  properties or a delete of a referenced header cell is refused; a
  relabel to null drops the label; a reclassify drops foreign
  attributes; a split strips whitespace across runs; ids.jsonl order
  follows spec/04 §4.4 in every state; a move after itself is refused.
- Anchors: the relabel rule requires that nothing else changed below the
  node; an alias chain ending in a deletion no longer crashes resolve.
- Notation: a script after an empty group applies to the previous atom
  ({}^{14}C); letters are defined by Unicode block in spec/05 §5.4 and
  the oracle; \left) is complex.
- Page fingerprints delete whitespace before folding, so "10 -20" and
  "10-20" fold alike (spec/07 §7.1).
- Grades are summed with exact decimals; the reviewer error bound is
  computed with integer arithmetic.
- receipts and revocations are keyed by (doc_key, id) in SQL, since one
  revocation may be carried by several packages.

Checking:

- tools/stress.py: seeded randomized property tests over canonical JSON,
  normalization, notation, random correction walks, anchor resolution,
  Merkle roots, spot-check samples, error bounds, fingerprints, mutated
  archives and packages, rights and spot-check chains; CI runs it.
- check.py verifies every count and version the prose states against
  the repository, the zipped fixture and crafted archives, the checklist
  coverage, and nine more broken packages.
- CONTRIBUTING.md gains a release checklist.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
2026-10-14 20:07:22 -04:00

12 KiB
Raw Blame History

8. Fidelity, receipts and spot-checks

8.1 Levels of fidelity

Level Meaning
F0 converted One reading, not corroborated.
F1 resolved Every suspect in the node was answered against the page image and passed the gate.
F2 corroborated An independent witness agrees, by match_hash and complex keys, or the text is born digital, as a Typst or LaTeX source is.
F3 proofread A named person compared the node with the page image, and a valid receipt they signed lists the node at its current node_hash.

Levels apply to leaves. The pipeline records F0, F1 or F2 in a leaf's fidelity; F3 comes only from receipts. A correction that changes a leaf drops it back: the gate's crop evidence makes it F1 at most, until it is corroborated or proofread again.

8.2 Levels

A leaf's level is the highest of its recorded fidelity and the levels of every valid receipt (8.6) that lists it at its current node_hash. A receipt at a lower level never lowers a node; only a revocation, a failed spot-check or a correction does. A spot-check's qa: receipt (8.8) raises the nodes it sampled, except those its errors name. A document reports, in document.json, the lowest level among its leaves and the count at each level.

8.3 Receipts

A paper has thousands of nodes, and nobody signs thousands of times, so people sign receipts: one signature over a Merkle root of everything reviewed in a sitting.

Property Required Holds
format Yes scienting-receipt/1.
id Yes The record ID (4.6).
doc Yes The document key.
rev Yes The revision reviewed.
level Yes F1, F2 or F3.
scope Yes all, section:<id>, page:<n>, targeted:<name> (such as targeted:lode), or qa:<receipt id> for a spot-check.
nodes Yes [[id, node_hash], …], sorted by ID without repeats.
records No [[kind, id, hash], …] for other things signed in the sitting: annotation, digitization, correction, layer or checklist.
root Yes The Merkle root (8.5).
evidence Yes What the reviewer looked at: {kind, sha256, node?, page_fp?, box?, note?} with kind crop, witness or source-file, or {kind: "person", by}. A regenerated crop is different evidence.
by Yes The signer's roster key.
at Yes When it was signed.
qa For qa: scopes The spot-check (8.8).
export_of No The receipt this one re-issues under a keyed scheme.
keyed No The scheme its values are keyed under, when it leaves the team.
signature Yes {scheme, key, value} (8.4).

A receipt lists nodes, records, or both. A correction invalidates a receipt for the corrected node only, because the node's node_hash no longer matches; the rest still counts.

A receipt shared outside the team cannot carry plain node hashes, which would leak short texts, and its signature covers them. So an export re-issues it: a new receipt with export_of, values under the export's key, signed by the original reviewer or a team key with the export purpose.

8.4 Record hashes and signatures

A signed record's record hash is SHA-256 of the canonical JSON of the record without id and signature. Its id is the 13-character letter-first encoding of that hash. Its signing payload is the canonical JSON of {"format": <the record's format>, "hash": <the record hash>}.

Scheme value
ed25519 The RFC 8032 signature of the payload, in standard Base64 with padding.
ssh An SSH signature of the payload in the namespace scienting, armored.
gpg A detached, armored OpenPGP signature of the payload.
sigstore A Sigstore bundle for the payload, as JSON text.

signature.key is the signer's roster key and equals by. Self-hosted deployments sign with team keys or a private Sigstore deployment; the public Sigstore instance, whose log permanently records each signer's identity, is for public layers only. signatures.jsonl verifies one receipt and the RFC 8032 test vector.

8.5 Merkle roots

A receipt's items are ["node", id, node_hash] for each listed node, in order, followed by its records. The root follows RFC 6962: a leaf hash is SHA-256 of the byte 0x00 and the item's canonical JSON; an interior hash is SHA-256 of 0x01 and the two child hashes; a list of n > 1 items splits after the largest power of two less than n. RFC 6962 has since been replaced by RFC 9162, which defines the same tree. A root over one item is that item's leaf hash; a receipt with no items is invalid. The tree allows a proof that one node was covered without revealing the others; merkle.jsonl gives roots for one to seven items.

8.6 Receipt status

A receipt is:

  1. invalid when its root, ID or signature does not verify, or its signer's key was not valid when it was signed (registered_at ≤ at < revoked_at);
  2. otherwise revoked when a valid revocation names it, or names its signer with a period that contains its at;
  3. otherwise suspended when a valid spot-check of it found an error;
  4. otherwise valid.

Only valid receipts set levels. A suspended receipt stays suspended; its reviewer reviews the sitting again and signs a new receipt, which needs its own spot-check. A spot-check is itself a receipt, so its status may depend on its own spot-check; statuses are settled from the end of each chain, and the receipts of a chain that loops back on itself are invalid.

8.7 Revocations and the roster

A revocation, scienting-revocation/1, names one receipt, or a signer and a period {from, to}, with a reason, and is signed like a receipt. It is valid when its signature verifies and its signer is the revoked receipt's signer, the signer it names, or a key whose role in the roster is owner.

The roster, scienting-roster/1, lists the keys that may sign: each with an id, a person, a role, a scheme, its public key (Base64 for Ed25519), its purposes (sign, spot-check, export), registered_at and revoked_at. The role is the holder's role in the team that keeps the roster, owner, editor, reader or guest, as in a project (20). Because the roster carries it, a package that carries its roster carries everything needed to judge its revocations. A package MAY carry the roster its receipts need as roster.json, so its receipts verify without the store. The roster is meant to be the one a store verifies against; see the rationale.

8.8 Spot-checks

For each F3 receipt, a second reviewer checks a sample of its nodes and signs a qa: receipt. A seed the reviewer controls could be ground, by varying the receipt until the sample falls on careful work, so the seed is a signature nobody can steer:

  1. The seed is the checker's Ed25519 signature over the canonical JSON of {"format": "scienting-spot-check/1", "receipt": <the original's record hash>}, made with a key registered in the roster with the spot-check purpose before the original was signed. Ed25519 signatures are deterministic, so the checker cannot choose it, and the reviewer cannot predict it.
  2. The sample size is the smaller of n and the larger of minimum and ⌈rate × n⌉, for the original's n nodes; the defaults are a rate of 0.05 and a minimum of 10.
  3. Rank the original's node IDs by the hexadecimal SHA-256 of the seed's 64 bytes followed by the ID's UTF-8, then by ID, and take the first ones.

The qa object records of, key, seed (Base64), rate, minimum, the sample sorted by ID, and the errors found, each {node, kind?, note?}. The qa: receipt's nodes are the sampled nodes at the hashes the original lists, its scope is qa: and the original's ID, every error names a sampled node, and it is signed at or after the original. Anyone can recompute the sample; spotcheck.jsonl does. The payload's identifier, scienting-spot-check/1, is registered with the other formats so that no file ever carries it by mistake; it names a signing payload, never a file.

8.9 Reviewer error bounds

One sample proves little; what someone relying on the work needs is a bound on each reviewer's error rate across all their samples. With k errors in n sampled nodes, the bound is the one-sided Clopper–Pearson upper confidence bound at confidence c, by default 0.95: the rate p at which the probability of at most k errors in n is 1 − c, or 1 when k = n. It is reported to six decimal places, rounded up. With no errors the bound is about 3/n, so zero errors in 300 sampled nodes bounds a rate below 1%; bounds.jsonl gives cases.

A QA report, scienting-qa-report/1, states for each reviewer who signed a delivery the nodes sampled, the errors found, the confidence, the method clopper-pearson and the bound. It is the record a buyer keeps, so it is signed like a receipt (8.4): by names the roster key of the team that made the delivery, id derives from the record hash, and signature.key equals by.

8.10 Targeted F3

Most of a paper is never relied on, so a team proofreads the target set: the nodes its claims, data and quotations rest on.

  1. The anchored nodes: every live node that an extract, a dataset value, a confirmed claim-link, a draft quotation or a pinned citation anchors, each anchor resolved first (6.5). An anchored leaf is a target; an anchored container contributes its leaves.
  2. Every equation and equation-line.
  3. Under the project's digit rule (20): with all, every sentence and cell with a digit in a text or math run; with anchored-sections, only those in a section that holds an anchored node; with none, none. Digits are characters of General Category Nd. Digits in cite, xref and fnmark runs do not count, so citation marks, "Figure 3" and footnote marks pull nothing in. A node's section is its nearest ancestor-or-self of kind section, or the front or back matter when it has none.

The set is reported in document order, each node with the first reason that applies: anchored, equation or digits. Neighbors shown for context are not in the set, and nothing new comes to rest on a node below F3 without the node joining the set. targets.jsonl gives the set of fixtures/packages/hale2026harmonic under each digit rule.

8.11 Checklists

An export of a document, such as accessible HTML, can be checked against a standard, and that check is a record a buyer keeps. A checklist, scienting-checklist/1, holds the doc and rev checked; the standard, {name, level?, url?}, such as WCAG 2.1 at level AA; the export checked, {kind, sha256, size?}, so the record binds to one file; its items, each a criterion, its result (pass, fail or not-applicable), the method (manual, or automated with the tool and its version) and an optional note; and by and at. Its content hash is SHA-256 of the canonical JSON of the record without id, and its id is the record ID of that hash (4.6). A checklist carries no signature of its own: the person who checked signs a receipt that lists it as ["checklist", id, hash] (8.3), so one sitting covers a checklist together with the nodes read for it. The receipt covers the checklist only at that hash, so a checklist changed afterwards is uncovered. fixtures/records/checklist/ holds a checklist for an accessible HTML export of the example paper and the receipt that covers it. The format is provisional (0.4).