ClaimSet 48 fields + 45 fields

research/ClaimSet

ids list of object

This record's own identities: the outside addresses that say which real thing this record is (arXiv id, DOI, ORCID, Apple episode, homepage). Each address belongs to exactly one record in the whole store; a second record holding it means the two are one thing and must be merged. Order is meaningful: the first provider identity of a Media record is its preferred recording.

no description declared

required

url string matching /^https?:\/\/.+$/

Canonical URL whose shape must match the selected source. HTTPS is required except for a homepage or canonical-url whose description records the failed HTTPS attempt with HTTPS-FAILED: followed by the observed failure; HTTP homepages render as plain links, never fetched or embedded.

path string matching /^\/.+$/

Absolute path into a $-prefixed skill source's owned store; omitted when the location derives from another identity on this record.

metadata object OR object OR object OR object OR object OR object OR object OR object OR object OR object OR object OR object OR object

An artifact pointer's provider comes from its source's registry entry unless metadata.provider names a saved-copy owner: site-warc on canonical-url or doi, or query-lesswrong on a canonical-url or related-url forum post. New saved-copy pointers state their owner explicitly; historical canonical-url captures may omit it.

form 1 object OR object

Provider-backed arXiv absence or artifact result.

required

form 1 object

Provider-backed absence of an arXiv preprint for this Paper.

required

unavailable exactly "preprint"

What the provider could not supply and the only discriminator of this metadata arm, such as pdf; a provider-specific marker key in its place is banned.

required

reason one of: arxiv-search-miss

The provider-typed cause of that absence drawn from this subject's closed vocabulary, such as closed; free text or a cause invented at the call site is banned.

required

recheck string matching /^(snapshot-v\d+|final)$/

Where this absence sits in the recheck cycle: the arXiv search snapshot that last looked, such as snapshot-v298, or final once the paper is over 18 months old and is never re-checked; any other value is banned.

required

form 2 object

Pointer to a query-arxiv-owned PDF artifact, validated when it was saved.

required

artifactPath non-empty string

Absolute path into that provider's own artifact store, such as <provider-store>/papers/<id>.pdf; a relative path or a record-local copy is banned.

required

form 2 object OR object

Provider-backed DOI acquisition result.

required

form 1 object

Provider-backed absence of usable PDF bytes for a DOI Paper, whether no source was recorded, the selected host refused them, or every open-access source failed and the mirror fallback refused (mirror-refused).

required

unavailable exactly "pdf"

What the provider could not supply and the only discriminator of this metadata arm, such as pdf; a provider-specific marker key in its place is banned.

required

reason one of: closed | no-direct-pdf | doi-not-found | pdf-host-http-403 | mirror-refused

The provider-typed cause of that absence drawn from this subject's closed vocabulary, such as closed; free text or a cause invented at the call site is banned.

required

form 2 object

Pointer to the validated full text proving a DOI Paper: a query-unpaywall PDF, or a site-warc capture of the publisher's own page when provider is site-warc.

required

artifactPath non-empty string

Absolute path into that provider's own artifact store, such as <provider-store>/papers/<id>.pdf; a relative path or a record-local copy is banned.

required

provider exactly "site-warc"

Set only when the publisher serves no PDF anywhere (Distill is the case): the pointer then names a site-warc capture of the publisher's own article page instead of a query-unpaywall PDF, accepted only after every PDF route failed. Absent means the ordinary query-unpaywall PDF proof.

form 3 object

Pointer to a query-openreview-owned PDF artifact, validated when it was saved.

required

artifactPath non-empty string

Absolute path into that provider's own artifact store, such as <provider-store>/papers/<id>.pdf; a relative path or a record-local copy is banned.

required

form 4 object

Provider-backed absence needed to keep an affiliation-less Person visibly blocked.

required

unavailable exactly "institution"

What the provider could not supply and the only discriminator of this metadata arm, such as pdf; a provider-specific marker key in its place is banned.

required

reason one of: openalex-institution-absent

The provider-typed cause of that absence drawn from this subject's closed vocabulary, such as closed; free text or a cause invented at the call site is banned.

required

form 5 object

OpenAlex-reported citation count for this paper.

required

citedByCount integer >= 0

OpenAlex citation count observed when this surface was last read; counts differ by index, so it is never copied onto another source's row.

required

form 6 object

Semantic Scholar-reported citation count for this paper.

required

citedByCount integer >= 0

Semantic Scholar citation count observed when this surface was last read; counts differ by index, so it is never copied onto another source's row.

required

form 7 object

no description declared

required

finished exactly true

Apple Podcasts has counted at least one full play of this episode, read from the user's library at the row's checkedOn. Written only as true: a play count above zero never goes back down, so absent means not finished as far as anyone has looked. This flag records completion, not its date; live play counts and playheads stay in Apple's library.

lastPlayedAt string matching /^\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}(?:\.\d+)?Z$/

Apple Podcasts' last-played time for this episode, as a UTC ISO-8601 instant from the user's library. Written by the listening pass alongside finished when Apple supplies it, including for an episode already marked finished. A missing provider date leaves the stored value unchanged. This is the last play, not the time the episode was first completed; the Media page shows it as Last played.

form 8 object

no description declared

required

artifactPath non-empty string

Absolute path into that provider's own artifact store, such as <provider-store>/papers/<id>.pdf; a relative path or a record-local copy is banned.

required

provider one of: site-warc | query-lesswrong

Owner of the saved copy. query-lesswrong is permitted only for a LessWrong, Alignment Forum or EA Forum post and holds the full post followed by every API comment. site-warc, or an omitted provider on a historical row, names the retained page capture. A new save never overwrites historical evidence.

form 9 object

no description declared

required

artifactPath non-empty string

Absolute path into that provider's own artifact store, such as <provider-store>/papers/<id>.pdf; a relative path or a record-local copy is banned.

required

provider exactly "query-lesswrong"

The saved-copy owner for this LessWrong, Alignment Forum or EA Forum post. The pointer includes the complete post and comment thread and never changes the Paper's canonical full text.

required

form 10 object

no description declared

required

artifactPath non-empty string

Absolute path into that provider's own artifact store, such as <provider-store>/papers/<id>.pdf; a relative path or a record-local copy is banned.

required

provider exactly "site-warc"

no description declared

required

form 11 object

conf-talks session identity for one Media document.

required

sessionId non-empty string

Stable conf-talks session identity defined by the provider skill.

required

checkedOn string matching /^\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}(?:\.\d+)?Z$/

When this surface was last read AND adjudicated as this record's own, as a UTC ISO-8601 instant; paired with a leading VERIFIED: in the description and never stamped by an auto-attach - except on a citation-count row (openalex-work, semanticscholar-paper), where it is when the provider reported the count and carries no VERIFIED: stamp, and on an apple-podcasts-episode row carrying finished: true, where it is when the user's Apple library was read and likewise carries no VERIFIED: stamp.

description non-empty string

What the surface supplied or a caveat specific to it.

state one of: stub | hydrated | enhanced | highlighted

Derived state; each level includes every requirement of the levels below it: stub for a mechanical import, hydrated when its Paper carries an arXiv or DOI identity, enhanced when that Paper has validated full text, and highlighted when enhanced holds and the user highlighted this ClaimSet.

required

claims list of object

Stable append-only claims shared by every reproduction of this paper.

required

id non-empty string

Stable record-local claim id used by reproduction verdict rows.

required

statement non-empty string

Precisely located paper claim shared by every reproduction.

required

Checks

researchLinksRule - every link matches its source-specific URL and metadata contract
_nonPersonHomepageLinksRule
programProfilePersonOnlyRule
claimSetClaimIdsAreUnique - record-local claim ids remain unique

Defects

Checked 200 of 6,768 records (the per-store cap is 200), so this is a sample, not a whole-store verdict.

No defects in the 200 records checked.

Where it lives

Bound by	research-records
Declared in	`~/mnt/mdr/skills/research-records/assets/schemas/claim-set.ts`
Binding	`claimSetDataStore`
Directory	`~/data/autoresearch-reproduction/claim-sets`
Files	`*.yaml`
Referenced as	`claim-set`

Specimen record 15 lines

~/data/autoresearch-reproduction/claim-sets/abas-2026-tur-dpo.yaml

1slug: abas-2026-tur-dpo

2state: hydrated

3claims:

4 - id: c1

5 statement: TUR-DPO improves GSM8K exact-match accuracy from 58.7% (DPO) to 62.8%, and MATH-mini exact-match from 33.4% to 36.0% (Table 1).

6 - id: c2

7 statement: On GSM8K reasoning graphs, TUR-DPO reduces average cycle count from 10.3 to 7.6 per graph and dangling nodes from 16.9% to 12.2%, while increasing minimal path coverage from 61.5% to 69.3% (Table 3).

8 - id: c3

9 statement: TUR-DPO reduces Expected Calibration Error from 0.101 (DPO) to 0.087, with the largest reduction (2.8 points) in the highest-confidence bin (0.9-1.0) (Table 9-10).

10 - id: c4

11 statement: Ablations show removing the uncertainty weighting term drops GSM8K win-rate from 62.8% to 60.3%, and removing the topology reward drops it to 59.6%, versus a graph-size-only baseline of 57.7% (Table 21).

12 - id: c5

13 statement: TUR-DPO matches or exceeds PPO-RLHF on reasoning tasks while requiring only ~15% more compute than vanilla DPO (42 vs. 48 GPU-hours on 8xA100s to reach a 65% win-rate target) (Section on Computational Overhead, Table 11).

14 - id: c6

15 statement: TUR-DPO is evaluated across mathematical reasoning (GSM8K, MATH), compositional reasoning (BBH), factual QA, summarization (TLDR faithfulness), and dialogue (HH), with human evaluation showing inter-rater agreement rising from kappa=0.66 (DPO) to 0.71 (TUR-DPO) (Table 7).