sed-i/mania-pattern-annotations

osu!mania pattern annotations

Snapshot v4 uses publication schema beatmap-lens-annotations version 4.

This snapshot contains 600 human judgments, 16402 agent judgments, and 2969 source identities (annotation and required calibration sources).

Export implementation: GitHub commit d6b4210550ba.

v4: 2,860 complete sections

This release retains all 2,860 requested complete sections using explicit metadata-only source access. Each source retains its original SHA-256, beatmap ID, title, artist, mapper (creator), difficulty and retrieval URL. The original beatmap page is https://osu.ppy.sh/beatmaps/{beatmap_id}. Current public downloads may differ from the annotated version; retrieval availability is independent of label admission. Always match the original SHA-256 before using chart bytes.

Source	Complete sections
Published v3, combining human and machine labels	736
Other eligible local sections outside v3	59
October 6 campaign	1,023
October 7 campaign	1,042
Total	2,860

A complete section has current, independently supported machine judgments and/or effective human judgments covering all five Foundation dimensions at the same source SHA-256, start, end and playback rate. Human cells take precedence, including masked uncertainty. Agreeing human records retain their identities and must be deduplicated when weighting training examples. Machine judgments retain separate opt-in method tables and are not human gold.

The snapshot also retains eligible individual tag judgments in incomplete sections. Missing dimensions are unreviewed, never inferred negatives. The October campaigns delivered 1,175 and 1,225 sections respectively; their 152 and 183 incompletely supported sections are not counted as complete. The extra 23 and 42 fully supported sections beyond the campaigns' first-1,000 exports are included.

October labelers requested gpt-6-astra with medium reasoning; different fresh Astra auditors requested high reasoning. They ran in 25-section batches with normal service tier using codex-cli 0.160.0. The frozen skill bundle SHA-256 is bf56212bc04c99b4bcc5736917285680f0fba756864bd8222c8a22ca3a186299; source commit 22e5c84f5cacb8493bdab5f1d0fdc09c5373dc60 pins the roles and eight harness tools. The 94 and 98 final worker records have matching shared input and skill hashes. Historical methods retain their own recorded versions and public artifact pins. Same-model independent audit can retain correlated errors; this release makes no held-out benchmark or population accuracy claim.

No source bytes, audio, model weights, raw traces, or private source paths are included. Source identity remains the original SHA-256; changed downloads must never be substituted for annotated versions. All previous published tags remain unchanged. Thanks to LuckyCosine7042 for moral support.

Source retrieval observations

Checks completed 2026-10-09T04:12:26.578868+00:00. Of 2,969 retained source identities, 2,845 had at least one public retrieval URL returning the exact original SHA-256; 124 did not among the checked official and mirror URLs. These are dated observations, not permanent availability guarantees. The table below identifies the unavailable original versions; their metadata and labels remain included.

Original beatmap page	Original source SHA-256
3784899	001c80d21ff2270b447ad0a8ed65dcba174454e7a97dd9235daa6a7fb3efc5ea
4131923	0117abaf472e50b0636b82adddee1a82400bb1fcd0ee7beacadb9dbe4d897530
787944	02a133d033de4e3ff12d2344244df7783312dddd90dac3927a85038b2899cdb0
2726536	0433373d40d3408ea741ad6ca039bb92a27b2f5f7ff4dd850a6a9fceab7e3d1d
3575842	06238799b5f0cc5eb19da5aff4d149e4bb9f6fdb79c7b1927aaf98b93ad4a81c
1562395	081be2d9648b3f86fd574642360dd77853b719e6687d584edc09168195361ba9
3566273	0c0b6696f1623db93eb93a18c7b501723e55e86711d37a2e5033b5e27362d7d3
4652408	0d36f6ae7995987c1dc1280883e4f3ef476ffd9435d609691698b98cc32ebd3e
3392371	0d9732ac5003fcc4766ae2f4a1c7c205060b7257de0010bcdc25ec50d51b499d
2325441	0d9d985e0f38f026083a16bbf401541f1b5cce16b818d3270699c89f465f08de
3456997	0f0aad5abcdd2a0ac98adbd5b14304935dcec4f764bb9370d3854eaa286fa165
4340994	0f1f60add4cdf43d5a8832cc0c3ee32a282c994caa08493a41922bf49d8e272e
4344809	10ac675c939b7ae360a4062451528979f144cf7f802f4786800ef42f7acb84b6
2025496	14a91ec4cd3333a363506dce61432279b42920e34e5f14c34e94837b085f05c7
3253451	16594b5718e28c8ef27ed5c54b953ebbb6bcf2af135301bd73ef577c1ffa7270
3784896	18230f811484a9d9ee177c37916e7adbe3adcfc3d2d80e439bedc21f17e2d932
293911	1fdac8c2fd8f344b7b91e973246fd47888aecee696868bfc14b0f0f9996c32a1
4953374	205d4739a154f8a562e273ee1d95188184f86023487c8b1de88595554f17b29b
3782005	219dbb277ba33ad9a5c80b4942c4b427c4bd1a7627c50b9894388c3e79e1478d
4638677	21ae806b843a151e7c16ebf40c01c5a4f6a533cdb6016be36a3c9f3c66f21a17
1632981	2298b1b33155985956c9c071ba6b10764040dbb35de8499c2228a10a03876ab7
3468088	267ea5d814e8d3a1a051d7cc2825aa01648cbd167dec48d7ad36aa3de1ded087
4324192	2769511d64e8430a339173167b0010b756f3f33444049cb9348695da4700a9dd
2314533	2b3f752109a4d3ecfd988db3dcf643a074874f4bb93ee9322df1aac9ec684661
4027045	2cc6744d72ab618ab1dd391d44b7fc815c3e76d43a8d798e6c9a33b386dd71e7
1025070	2ccc8455b1e67ca99e5baa95f811e570837137b30c45998459f4d59b2d31e5e0
3606971	2d36b95c6fe056e863190f9c60a0cda398d4165724ed20994839cbb323421ce8
2595777	322bf0c0724908a992ca0c724435032e0e00a45dba538d13e0dfe81342e9368e
3973282	33b889f015bfbbe06475c6cc32b508a5532c5f3063380b4f819d5fdfe22de6b4
1086328	365c8629a8804935f4e55b85f71835351aeeab439d054215b71fb3992dc5f833
2603455	372a0b1a393da7399489bb4e8e619acd53dad0e2529438818713597fba4cbf7b
2060036	373ab9a72e742e85f7eb6cfbca3cc37c74090e306740136c809f73f46031a6bb
3673087	3bf83a2b1b1a7eaab5a827e99c79ee423ba4df013b334a6b6220cc5d87d03b8f
2948731	3d6a1022e842571fdefd081357945300f6998a09e1059c2e3dd9975978b00c1d
1966772	4b548c27ba78a6ee5a49584fe00a2ad5010799d004edb73104a16c37754fe1df
3413575	4de059498707570124d0ae65fda49c3f25ae58a50b2853926089f450adfeeb98
3583899	4e4cd85f672b5c11e8a974ac7288d8715c021a4faea17008165241253a83d05b
1862336	50c5df5c255446aa7e478eea830e84a3d8479435fe92fcdfc49623ae9a6d985c
1535762	510bd80681a8f6ecc8b2653342a02ff47b4740642b497a8c5ba10a81b9d81c66
2909745	516ed28478d943513ff4e3f08116fea2e3b42d8a2c86c531f9f3d0c6eb04f5f5
3556052	5559b7b6bf9630bab83c966f4ae4b4ed15a0275c90e32addd77981ea0b8eb8b8
4793996	592e5a4d6494c2ff367b40d8b458189603892cb30b44b3756d5fdb94c0d45d40
4210902	59be0abb9a7d574932b64e7b9c69de6a77abfe7bfbbf9fa72b21d3945c4e90da
4575874	5b34df05a1c64eb4cf058110b21a6f4c6204e9dea7a8390ec58625d6a4bee56f
3987715	5e1e58cec28a4287124ddd23c81abb28579bf0ac83117bc002173a58c03a4983
1379737	6805de2890d0c9fe4cfa59e2e2c3f983b1e98f33c3e411cf46e85c0a1ea225e4
3253520	6c0a571e04e24aa9450463ff2ea97ad4d01ba0eb0a0db4e2cf8c0253441724aa
4849054	6c9bd30c19b3ec89de63b48b6a323b7f8a77b454335d964564db36a9eab0c8da
2749117	6d795acc5bc28fc81eadd9b35aae681f8107320c0f8cc3e1b5930f5eaf351596
3635424	6f5dceb4896b293ac52bfefbe7844b009356410cec88c852228c291a5546cefe
3534660	6fa5ba39bf5ff1dd802c45bbce36d332ac758bb9134a8a2685077efae5af6ffd
3440041	710d317f286d89bd5677eedc63647ca2e35fe5fc31fe1d50724e13374ca0a4b6
3525702	723002d93eb1d22fb6052dbb791fa8ed4502339dee9a52b7c4c075c15adc1ef6
3457180	768f8562c8d115c4c8a94c35bdd595df36591bf31bf28fb92d33b562572ed688
3663061	7710cbb6a505b640ee58ae08b3dbdb22598b1270e582dd8b3e1ca8ce10cc95af
2103854	78591b986f4f6193e8a012fbc2767260b1773f7021b61ffc7f3a08b40c8e057a
4757028	78bef19ff8855e42198a5d17ad740bd3c20d080e0bb7a5a0c249c692c0afd8df
2912769	7bf200162801822a76944b0cbe098a3bdde5fc525b7399c61c5b98f29d7d1fce
3306565	7e0b5e9b00a447477caa17219044716de04918d32815721d7a8f1dffb2a82fbc
3861828	7e47bcd4f040be5aea153e356fa355ea2533553bb7d7be1507a04999db4ad941
3475520	7ec5862c52ef435c6ef58525e073d4c049047bffc5fce27d5ddd7895e6e512c9
3679235	804d618d05fa6fb1fd529379affa1cdec8f8ba39f979a32ca1534d5f9515173e
3198320	825b136425aec0ead29a3fba1157d2d77e9c58b15549ee0692fdc860f936b78f
2940967	830c898b2109062d575a6574a7e8c311a1c0ac4426fa38f18c84879fae508af8
4246481	8a19dbe80e5c161d21bc13135701ba6fbeeeccf42c2873de75509cf713ae344f
656901	8be8b535e9103938adf63f3351ef425459e39b05691a73cea91d82c9eac81fad
3313202	8d63c8f1b97260523a488398a48c1c0c11f9ed64396ef03e5e908513f3845702
3674684	8df1846ff297e424c44f0c16a41965e32d2324b28d7c5aa032d42fcf981664cb
4072432	90831730c33a148e9e64e4ebc8ce7e54566ef5ae5e04d2546466c0324cc9906a
3260220	917825677f34f0a6c435cd8bb38ea231ada216454f4f280f70d51e7c722e40cb
4793985	9188ec613f0501ea8e6db7cb363c8e3cd7862cec8572f14fd300b3a88bbf9bc0
2877369	98ad7e0cc64131f1bda8738263c55a8e21a06160f5219d543fc2aa33037b186f
3441203	9a87966e5c6dc17b325d1638c60ae85c414aaef523ef6e3c8b75d23d2c71e7e6
2468487	9a94afbd13854c303311484467d1414e055be7f5133e132dd929fb652f88bdb6
2705089	9c832975f3f27b5906f1b35bb6b549b1655a48d9dc741376ab98be58bf71974f
1367425	9cb921a5abd3bf7030e7f89b4768a4035385b837d3856ece7afd15568f3eb441
2727798	a11269b098ccd65e3b54435cab48b829631394f0e3742912538152247f2639a6
3678863	a8fe667b2c7a81d9d50c1de5888ec555a580bdb3935f9f8f6b13b9a58b3db63c
3771148	ab34a680bf33289af9add327046814c54a255ed0f9fdae7665587fc42b51112f
3965313	b44d5a4ee56d0ec6b0ae4d27985d33653b010b552227b56c36ffd4c4de1ab4ae
4316771	b4be24b80e86eeb5297257a3242ab62b0f7d9ab83d685adfcbc575fb4abd6dc1
3569190	b5460b52ed52d3c3c7cf3ed531f854715a20051075a191b794aaf99a94424a01
2817250	b6da5c329b6dc2f14894dfccdda316f9dc97479c19cc251f7b9a0b155ab66e22
1104278	be04abb40105ba2f13e53384e9a4594d79d220c153cb41ba3821b4b6910bd0ff
3467739	be530f3bbe67f87c60efca4302b5c87232d36b8e2da8dfcc0765ac89de685cba
1126788	c03d50fdeb14edec6530f38b98b42ea7f6f278a9b27d681a07d238bce68f7eae
4142170	c0f214da6a542d4ee16a6093a527a65699e3ec4c1a4baff10f5d05fedaf25e72
3200765	c0f62e3b8f20c8ace89d25b33b3692295fc62810dd5ca5ee0dff1bedcce70a93
3703413	c2d801efc9ad3de411e50e2b5edf27f2523e69c555f1e36619533d8ea4f0eb23
2056576	c303d6b8f262ecd39cd18c0c2dca40a42fdae1b524126f41af39ce8a6f670e77
3457474	c459454bc7426c816db91fd649548ac204fc294c3fcdc626249d50122edd88c2
4276089	c6e49c0cce47c1b86513306cff629e753cc4ef7ded2b456f6333a3cf26fd715f
3142765	c79909da9ee90be414baa2387e973a60fa0d07aee447da3a48df4ece067e6a74
4283840	c90c921cbac165c0842880a0ce2195bec5c7cc2f118b8360537545503f0d5c8e
3490578	c9670abc7dd654808808cddadab2da57d5de3e74f0bf6f2e34f25c34eb19dec7
3897575	cbca50a223ccc772965ddd23c70232461f8bcc29b49f18605a04c98b5be648f5
2468707	ce2ee5d5fdb2fe9c3619ff446c292c2f2f37c12adeab50ce3c48de7c809d2010
5370523	ceb036a373ccae4ccac55f942aeb16c4b64399950c806add8a86181212ae781d
2182529	cf44532f3cb1e3f257fead5b88b53b89799efd9741edcd02dde642a2ca2d88ed
3497908	d098ca97fbabaa1c88e6fa1d29c6343b791a30e9c32c04a5fab6446ffc5d3a47
3621175	d13f18bec794955550c1d2b0ce1420259bc5e665938890e6cc21ee532c300d11
2006488	d4a1ca9c904d96862d00dd27f5f26a45744a9160815c867ca2348802bd84bb46
4440091	d83a9dfbbc8876967074a2cacdda877d2499e1d450bd3c4228a1b42f13526e63
2525713	d951d7a29d15714fe008df58c14b071ec1ec061cc4cbee7ce1fb748d5d1ce448
3441142	d9e42114cf3868121c7df0c503e97518f7c39ff69be417bb2db0a350fdf47a82
3872699	e01a435b202aaca223aa67001b11a69b8598de2cc055f79fb7aba9013bebb93c
3185607	e2ac3b977c43e8cdfde66380f58301375868c99be0baa82fee0975ecd8dde2ce
3793375	e40d14df7c11564095ef2a1e2ed61f1ae847d4bbd1cd8431585febb6bdef5e3a
3415961	e90203943cc9632fdf3b636bb1cd439f91cb9f8a07a9c9a2ca97f3563e47e4d0
2555476	e93c3cd9c98d3ab647da49ed7da15fa496b67996d0e1ea8ababcd9c276c592b7
3708561	eb37c714ad42e0e7d00b440361088008c1e0e6d9711e1f3000ca59d18cda7acd
3709974	eb84a0323d01916d8653dadc2e75d296c150c44fdb96831f4768c2e2cd2a49f6
3519359	ec07907256fdbcaf71a69134323a46d7dc73055051d6c074d5408a3fe4670584
5285158	eed10384f893e6efefe58bb86c281e6f170a75eae332976e8fa7c773baa949b3
1287062	f337e1ca9926f3ca618b7ef254c511d2f00cb911bc286ce23fe9a380d42e63c7
3617507	f44450ec122f849de2d9c7f98613a8c0c9f1c863df7f8b51804e6b5679f2e795
3281799	f6ce62b725e06f2c367b57a675eea535d1e7c0d9c8b2e20b2fe639668ec76c57
1938851	f7612c708cb56d133557f08b47e84f23a56cdd54bb1f321a26f044f4b19f898d
2281313	f8d80fbff200b8c1d5fd711854814879023c4a52e74b1f95a8c36a8870034b11
1108288	f9893bb1156313d1efe7cd5ce3500085569fc799d465c2c23cc825d575781b64
3988022	f9fd4083240406a3d12ff39d4b852ba398f99a5ddbda10b86b84080e20da13f6
4202630	fbd9db4a4c598df464a99e045e3890d8e45e081a4387bde246975278008f2e6d
1170730	fcfc0a990443731bc8ec755277e68c82c67e8f017add2a2807f62f558a86667b
3781967	fea747d1a67014b9f59bdf091a3d3f6eb8065f11ead500852724f83ca04825ed

Source identity and availability

This release uses policy.source_access: metadata-only. It publishes the original source SHA-256, beatmap ID, title, artist, mapper (creator), difficulty, and retrieval URL even when the original chart version is no longer publicly retrievable. The original beatmap page is https://osu.ppy.sh/beatmaps/{beatmap_id}. Publication does not require those URLs to serve matching bytes; an existing or working URL is not a verified copy of the annotated version. The collector still validates the original local bytes and source-backed judgments. Exporter, Foundation, and method artifact checks remain required.

Before applying a label, source time range, or note reference to a downloaded chart, verify its SHA-256 against source_sha256. A failed download or hash mismatch means that chart version cannot be used with these labels. Never substitute the current chart version. Retrieval observations in the release notes describe only the stated check period, not permanent availability.

Normative Foundation definitions:

• f-15fa68913bdb2bf3

The default human configuration contains effective human judgments. Agent configurations are explicit opt-ins and must not be silently combined with human gold. The selected machine tables exclude exact cells with effective human judgments; ancestry remains available through provenance.

This is a positive-first, partially exhaustive annotation resource. A row is one tag assessment over one exact source interval, in original source milliseconds, with half-open [start_ms, end_ms) boundaries. Tags are independent and multiple tags may be prominent. presence is present, absent, unresolved, or unreviewed. Only present has salience (supporting or prominent); other rows have null salience. Missing rows are unreviewed, never negatives. Rejected proposals do not manufacture absence labels.

from datasets import load_dataset
human = load_dataset("sed-i/mania-pattern-annotations", 'human',
                     revision='<full-HF-commit>', split='full')
supervised = human.filter(lambda row: row['presence'] in ('present', 'absent'))

The split is named full; this release makes no held-out benchmark or measured accuracy claim. agent-reviewed describes independent audit, not human authority or calibrated confidence.

sources contains metadata, exact source hashes, and retrieval references, not beatmap bytes, notes, audio, or images. For source_ref.kind == 'hf', retrieve path from the declared corpus repository at the full commit; record_key, when present, identifies the record in that file. For content-addressed, retrieve uri. For osu, retrieve the official https://osu.ppy.sh/osu/{beatmap_id} URI. For url, retrieve the declared HTTPS uri, such as a public mirror holding the original revision. The official locator is mutable, as are general URL locators, and neither promises archival availability; the original source_sha256 remains the identity. Verify the retrieved original .osu bytes against that hash before using time ranges or source-line note references. A mismatch must fail instead of substituting a newer chart. Corpus access and format are owned by its publisher.

The manifest references immutable public Foundation and method artifacts on GitHub. Frozen Foundation hashes and public artifact hashes are deliberately distinct when calibration bytes have been replaced by references. Definitions, skill text, raw agent packets, and workspace journals are not embedded here. details retains typed context, evidence, boundary uncertainty, transition, exemplar role, section ID, and original human rationale. Public annotator identities are omitted from annotation tables.

Judgment authority, source/Foundation compatibility, and auxiliary evidence freshness are separate. Included rows have compatible source/Foundation binding. auxiliary_evidence_status is snapshot-relative: current, changed, untracked, or not-applicable. Changed or untracked machine ancestry does not revoke a human judgment. Human confirmation does not certify an ancestor's rationale. Historic snapshots remain fixed as workspace judgments evolve.

Machine release policy: auxiliary evidence admitted = ["current"]; partial method provenance admitted = false. Selected agent rows require independent supporting audit and an effective agent-reviewed or accepted state. See manifest.json for selected methods, omissions, lineage, file hashes, and public evaluation references, when supplied.

Checks of schema, identity, and provenance do not measure labeling accuracy. No numeric quality certification is implied.

playback_rate is the judgment's execution rate: 0.5, 0.75, 1, 1.25, or 1.5. All stored ranges and note references remain in original source milliseconds. Performance durations and intervals divide by this rate; BPM multiplies by it. The same source, scope, and tag at different rates are distinct judgments, not contradictory labels. Historical schema-v1 rows implicitly use 1x. Split related rates of the same source together during evaluation.

The exported annotations and accompanying dataset documentation use the MIT license in LICENSE. Externally referenced beatmaps remain subject to their own terms; this snapshot does not distribute or relicense their contents.

Machine admission is restricted to 3430 exact handoff IDs listed in policy.agent_handoff_ids. Other runs of the same method are excluded as handoff-not-selected. This batch restriction does not filter human records or their ancestry.

For training, identify an exact judgment cell by (source_sha256, start_ms, end_ms, tag_id, playback_rate). The human table preserves distinct effective observations, including agreeing duplicates at the same cell. Deduplicate agreeing human assessments by this key before weighting examples; retain their record IDs as provenance. Keep unresolved and unreviewed cells masked from supervision and do not replace human uncertainty with a machine label. Conflicting supervised human assessments at the same cell block publication.

policy.human_precedence is true: otherwise-eligible machine rows are excluded whenever any effective human projection row has the same exact cell, including unresolved or unreviewed human assessments and rows from different handoffs or claims. This check uses the full effective human projection, even when a human row is omitted by another release gate. Different source hashes, interval boundaries, tags, or playback rates remain distinct. exclusions.agents.human-precedence counts the omitted machine rows. The canonical human table and separate opt-in method tables remain separate; no merged gold table is published.

cell_id groups the exact (source_sha256, start_ms, end_ms, tag_id, playback_rate) independently of record, authority, or method. All original record_id values remain. Select authority and method layers first, then give an agreeing cell one training weight; do not count duplicate records or overlapping machine methods as independent examples. The exporter supplies no universal weight. Resolve contradictory selected machine assessments explicitly. Related source versions, overlapping sections, rates, and exemplar dependencies also need split controls beyond exact-cell grouping.

observation_sha256 binds a human row to its canonical workflow observation when available; it is null when unrecorded and for machine rows. The row binds its current assessment and notes together. details.proposal_changes lists actual differences from the original proposal named by handoff_id/claim_id; witness and context comparisons use exact note-tuple sets. An empty list means unchanged, not independently checked. Null means no comparison is supplied. For example, assessment without witnesses or rationale records a revised label retaining the ancestor selection and explanation; it does not validate that explanation for the new label.

details.human_revision separately binds the immediately preceding human observation ID/hash and lists changed_fields for the latest revision, including direct-human revisions. Both comparisons cover tag, assessment, scope, rate, review context, witnesses, context notes, rationale, section ID, boundary uncertainty, transition, and exemplar role. A label change followed by a notes-only revision retains the cumulative assessment change against the proposal while the latest human delta lists only witnesses. Metadata-only revisions have empty changed_fields and new observation identities; null means no predecessor comparison is supplied. Only effective human observations are exported, with predecessor identities and hashes rather than private journal bodies.

details.evidence_review records selection origin, source pointers, and draft operations. The nullable legacy fields selection_reviewed/rationale_reviewed preserve declarations only when already recorded; current saves do not require or manufacture them. Null means unrecorded, never an inferred human selection. Operations distinguish automatic full-scope filling, explicit arrangement selection, manual edits, and range filtering. Their list describes the saved draft, not a complete note-by-note action log. Agents select notes while making their judgment; no post-hoc explanation or omitted-note reason is required. A recorded review declaration does not certify minimal or sufficient evidence or establish prediction accuracy.

In the section annotation delivery pipeline, context_note_refs is automatically the supplied review-context notes minus witnesses. It is surrounding material, not a selected negative set. Unselected notes have no implied semantic-negative label or per-note omission reason. Whole-scope witnesses, partial chords, disconnected witnesses, and entering long notes may all be valid. Subset size or coverage alone does not establish evidence quality or salience. Recover the complete source section and review context, including unselected intervening rows and long notes crossing its boundary, for evidence research.

auxiliary_evidence_status concerns freshness of referenced human exemplars, not whether this row's explanation still supports its current label. Inherited or empty explanations do not invalidate final human labels or create an explanation-review obligation. Independent machine audit is not human note-level gold and does not certify minimal or sufficient evidence. Keep evidence rationale, human decision rationale, and audit explanations out of salience-prediction inputs because they can disclose the target. Preserve manifest human_evidence_refs, including references with no published record_id, when constructing evaluation splits; exact source and observation hashes remain dependency identities.

human_confidence preserves the human's explicit high or low confidence in the saved assessment. Null means unspecified historical confidence and is never inferred as high; machine rows always have null confidence. Confidence is part of the canonical observation hash. A confidence-only revision creates a new observation identity and records confidence in details.human_revision.changed_fields; it does not change details.proposal_changes or the claim's assessment. Human confidence is separate from auxiliary evidence freshness and evidence-review metadata.