genes.apexpots.com / research source: donto-claim-substrate-appendix-2026-06-02.md

Claim-Substrate Report — Research Appendix (2026-06-02)

Claim-Substrate Report — Research Appendix (raw findings)

Companion to the iteration-3 report. Structured output of the 5-area study (2026-06-02).

resume-job-skills-intelligence

The skills-intelligence market runs on a handful of large public skill taxonomies that all model skills as graphs of related nodes, exactly the structure a claim substrate would emit. Sizes: ESCO (EU) ~14,575 skills/competences linked to 3,039 occupations across 28 languages; Lightcast Open Skills ~32,000-34,000 skills in 31 categories, refreshed every 2 weeks from 1B+ job postings; LinkedIn Skills Graph ~39,000 skills, 374,000 aliases, 200,000+ explicit skill-to-skill links ("knowledge lineages"), spanning 875M people and 59M companies; O*NET (US gov, free) 1,016 occupations with ~35 skills + 277 descriptors per the Content Model; SkillsFuture Singapore ~11,000 skill competencies across 34 frameworks/38 sectors. Crucially these are skill-to-skill relationship graphs (prerequisite, alias, adjacency), not flat lists, and the newest research (Skill-LLM, SkiLLMo, "LLMs as zero-shot ESCO matchers", ESCOX) uses LLMs to EXTRACT and NORMALIZE skill claims from free text against ESCO with ~79-91% precision. That extraction step IS donto's typed-claim ingestion in another vocabulary.

The commercial layer is large and consolidating around "talent intelligence" / knowledge graphs. Eightfold AI: ~$96.6M ARR 2024 (up from $58.8M 2023), ~$2.1B valuation, $410M raised, pitches an AI "Talent Intelligence Platform" with a deep-learning talent graph inferring latent/adjacent skills. Beamery: ~$112.8M ARR 2024, hit unicorn $1B in Dec 2022 on a "Talent Graph", but laid off 12% (2023) then ~25% (2024) chasing profitability. SeekOut: ~$25.2M ARR 2024, was $1.2B (2022) now ~$435M, sources from 700M+ profiles, cut 30% staff in 2024. Draup: $20M Series A, talent+sales intelligence on 80,000+ sources for workforce/reskilling planning. hiring.cafe: scrappy aggregator scraping real employer ATS feeds (Greenhouse/Lever/Workday) with GPT-based per-job summaries and natural-language search, free to seekers. The pattern: the unicorns are CASH-CONSTRAINED and their graphs are proprietary/opaque deep-learning embeddings; explainability and evidence-grounding are exactly where they are weak. Underlying market: talent-acquisition software ~$22-26B in 2025, AI-recruiting segment ~$3.2B growing ~12% CAGR; skills-based hiring claimed by 85% of employers in 2025 (up from 57% in 2022), though only a tiny fraction of hires are actually affected (the rhetoric/reality gap is itself a wedge).

JSON Resume (the founder's asset): an MIT-licensed open standard (schema v1.0.0, the resume-schema repo ~2.4k stars, jsonresume.org monorepo) with a clean typed structure already (basics, work, education, skills{name,level,keywords}, projects, awards, languages, etc.) plus a draft job-schema.json. Adoption is "10,000+ developers / tens of thousands of users", 400+ themes; the registry renders any GitHub Gist named resume.json at registry.jsonresume.org/ and already has nascent AI "suggestions". So the founder controls BOTH a typed-resume corpus AND a draft typed-job schema and the registry distribution channel, the two sides of the matching graph in a format that is already half-decomposed into claims.

Academic person-job fit has moved from CNN/RNN co-attention models (PJFCANN) to graph neural nets (Fine-Grained Semantics-Enhanced GNN 2025, graph adaptive fusion 2025) and to LLM re-ranking. The state-of-the-art honest admission is the donto thesis verbatim: ConFit v2/v3 (ACL 2025/2026) note embedding rankers "lack controllability and explainability as the ranking process happens entirely in latent embedding space" and bolt on LLM re-ranking to recover interpretability; bias is real (one benchmark 72% male). That is the precise gap a claim substrate fills: evidence-anchored, contradiction-aware, re-rankable matching with a human-readable WHY.

Players / datasets / taxonomies / standards:

Volume verdict: High-volume typed extraction is unambiguously HELPFUL on this domain, more clearly than almost any other, but only when split across the two layers. WHERE IT HELPS (extraction/decomposition): a single resume bullet ('Led migration of monolith to microservices on AWS for a 5M-user fintech') legitimately contains dozens of typed claims, explicit (AWS, microservices, leadership) AND implicit/inferred (distributed systems, observability, PCI/regulatory exposure, team-scale, domain=fintech, seniority signal). The founder is right: that latent typed structure is vast, and extracting it densely is pure recall gain, the raw fuel. Denser per-document claim graphs are PRECONDITIONS for discovery: you cannot find a hidden-candidate-to-job (A-C) link without the intermediate skill/role/trajectory (B) nodes existing (Swanson ABC). Skill-adjacency and career-path graphs only become rich at volume, the network effect is real, and the open taxonomies (ESCO/Lightcast/LinkedIn) prove the relationship structure exists and is large (200K+ LinkedIn skill links). So volume at the typed-extraction layer is the asset. WHERE IT BECOMES NOISE (relationship-hypothesis layer): the cross product resumes x jobs x skill-combinations is combinatorially huge, and naively materializing every candidate match edge produces overwhelming false-discovery (most skill overlaps are spurious, most implicit-skill inferences are weak, most 'matches' are degree-of-overlap noise). Here volume MUST flow through a verifier/ranker that scores by evidentiary support, novelty, contradiction-value and verification cost, re-ranks on new evidence, and gates low-evidence match hypotheses as hypothesis_only. The reconciliation: VOLUME FEEDS THE VERIFIER. More extracted claims => denser graph => more candidate relationships => but the verifier decides which candidate edges become surfaced matches. Volume and the verifier are complementary, not opposed; the failure mode is skipping the verifier and treating every extracted relationship as a finding. A useful concrete guardrail: every emitted match/implicit-skill claim must be individually evidence-anchored and falsifiable, claims with no backing bullet are hypothesis_only and never shown as confident matches.

Design implications:

Honest caveats:

linguistic-decomposition-dimensionality

The founder's intuition is literally true and measurable. Modern NLP defines a stack of ~12 distinct, gold-standard, FALSIFIABLE annotation layers over the same text, each with public datasets and inter-annotator agreement. They are not redundant — each emits a different TYPE of claim. Layer by layer, every token and every predicate gets decomposed many times over. Universal Dependencies annotates EVERY token with 6 fields: form, lemma, 1 of 17 universal POS tags, a bundle drawn from ~30 morphological feature types (200+ possible values: Case has 40+, Tense, Mood, Aspect, Voice, Number, Person, Gender, Definite, etc.), a HEAD pointer, and 1 of 37 dependency relations. That alone is ~5-6 typed claims per token before you touch meaning. Then Abstract Meaning Representation (AMR 3.0, ~59K gold sentences, LDC2020T02) re-encodes the same sentence as a rooted directed graph of concepts+relations (roughly one node per content word plus :ARG/:mod/:time edges). Then PropBank (every WSJ verb, sense-numbered e.g. leave.01 + numbered args ARG0-ARG5) and FrameNet (1,000+ frames, 13,000+ lexical units, named roles like Buyer/Seller) label predicate-argument structure / semantic roles. Then word-sense disambiguation tags content words to WordNet (content words average ~3-8 senses; SemCor is the gold corpus). Then coreference + named-entity + entity-linking (OntoNotes 5.0 layers all of these over 400K words across 3 languages). Then Universal Decompositional Semantics (UDS1.0, decomp.io) replaces categorical roles with 16-18 GRADED real-valued proto-role properties per predicate-argument edge (instigation, volition, awareness, sentience, change-of-state, existed-before/during/after...) PLUS predicate-level factuality (did it happen?) and genericity (kind/hypothetical/dynamic) — i.e. each single edge becomes ~20 scalar claims. Then temporal: TimeML/TimeBank events + TIMEX3 + TLINKs using Allen interval relations; TimeBank-Dense shows the combinatorial blow-up directly — 36 documents, 1.6K events yield 5.7K temporal links (~10x denser than sparse TimeBank, because you label ALL event/time pairs). Then discourse: PDTB-3 (53,631 annotated discourse-relation tokens, 3-level sense hierarchy) and RST/eRST (rhetorical trees). Then multiword expressions (PARSEME 1.3: 455K sentences, 26 languages, 62K+ verbal MWEs). Plus sentiment/stance and presupposition/implicature layers.

So a realistic count for ONE rich 100-word paragraph: ~100 tokens x ~5 UD claims = ~500 syntactic/morphological claims; +~80-120 AMR nodes/edges; +~10-20 predicates x (sense + ~3 args + ~20 UDS graded properties each) = several hundred semantic claims; +coref chains, +entity links, +5-15 temporal events/links, +3-8 discourse relations, +MWEs, +sense tags. That is conservatively 1,000-2,000+ typed, falsifiable annotations per paragraph WITHOUT any free-form prose. At corpus scale this is not hand-waving: SemMedDB is the existence proof — SemRep extracted 96.3 MILLION typed subject-predicate-object predications from 29.1M PubMed abstracts, and that database is the literal substrate of automated biomedical hypothesis generation. So "millions of typed linguistic facts from a corpus" is not aspirational; it is below the demonstrated ceiling.

Crucially for donto's thesis, these typed claims DO enable cross-text relationship discovery, and the mechanism is exactly Swanson's ABC / literature-based discovery: shared intermediate typed facts are the bridge. Swanson's Raynaud-fish-oil discovery worked because the bridging B-concept (blood viscosity) appeared as a typed claim in BOTH literatures even though only 4 of 489 articles co-mentioned A and C — the discovery LIVED in the intermediate typed facts, not the prose. Shared FrameNet frames, shared WordNet senses, shared entity links, and shared AMR concept nodes are precisely the join keys that let two texts that never cite each other be connected. This is the honest reconciliation of the volume debate: dense typed extraction is the FUEL (high recall, the raw B-facts must exist or no bridge is findable), and the danger is purely at the relationship-HYPOTHESIS layer where all-pairs combinatorics (TimeBank-Dense's 1.6K events -> 5.7K links, and far worse cross-document) demand a verifier/ranker. Volume feeds the verifier; they are not opposed.

Players / datasets / taxonomies / standards:

Volume verdict: HIGH-VOLUME TYPED EXTRACTION HELPS, decisively, at the decomposition/recall layer and as the supply of intermediate B-facts for discovery. The founder is correct and the numbers back him: ~12 distinct gold annotation layers, ~5-6 UD claims per token, ~20 UDS scalar claims per pred-arg edge, conservatively 1,000-2,000+ typed falsifiable annotations per rich paragraph, and SemMedDB's 96.3M predications prove corpus-scale typed extraction is already real and already drives discovery. Denser typed graphs strictly enable MORE discovery, because Swanson-style bridges only exist if the intermediate typed facts were extracted in the first place (4-of-489: the discovery was invisible in prose, visible only in the shared typed B-concept). So volume = recall = fuel: extract aggressively, multi-layer, with provenance. It becomes NOISE at exactly one place: the relationship-HYPOTHESIS layer, where you combine claims into NEW candidate edges. There the math is combinatorial (TimeBank-Dense: 1.6K events -> 5.7K links in 36 docs; cross-document entity pairs are quadratic), so undisciplined generation produces a false-discovery flood. The fix is not less extraction — it is a verifier/ranker between the dense typed-claim layer and the asserted-relationship layer (rank by novelty x plausibility x evidentiary-support x contradiction-value x verification-cost; re-rank on new evidence). Concretely: typed EXTRACTION should be maximal-recall and volume-loving; typed RELATIONSHIP-GENERATION should be precision-gated. They are complementary, not in tension. The one thing that IS pure noise regardless of volume is FREE-FORM PROSE output from a lens — it has no join key, no falsifiability, no provenance type, so it can neither feed the ranker nor bridge two texts. Every lens must emit rows in a declared typed-claim family or it is decorative.

Design implications:

Honest caveats:

volume-vs-precision-reconciliation

The literature cleanly splits volume into two layers, and the founder is right on the first while the prior report was right on the second. LAYER 1 — typed extraction / graph density (volume HELPS): Knowledge-graph completion and link prediction degrade sharply as graphs get sparser. Common-neighbor and triplet-closure methods "falter in sparse graphs since they rely on closed triplets," and commonsense-KB completion models "implicitly assume densely connected graphs, with performance degrading quickly as graph density is reduced" (Malaviya et al., arXiv:1910.02915). Real KGs are brutally sparse (commonsense and biochemical graphs avg degree ~2); denser benchmarks (FB15k-237 vs WN18RR) yield materially better link prediction. The mechanism is exactly Swanson's ABC: an A→C discovery is impossible unless the intermediate B-facts physically exist in the graph. If extraction is shallow, the bridge term is simply absent and the true latent link is unrecoverable — a recall ceiling no downstream ranker can lift. So the founder's "vast latent typed structure of one text" intuition is the correct objective at this layer: more typed claims = denser graph = more reachable true links = higher recall. This is the raw fuel, and starving it is the one mistake you cannot fix later. LAYER 2 — unverified relationship hypotheses (volume HURTS without a gate): The Calude-Longo theorem ("The Deluge of Spurious Correlations in Big Data," Foundations of Science 2017) proves via Ramsey / Van der Waerden theory that sufficiently large databases MUST contain arbitrary regularities purely as a function of size, independent of the data's nature — "most correlations are spurious," findable even in randomly generated data. Candidate relationships between N entities scale ~O(N²) (or worse for typed/path hypotheses), so the number of false candidates grows combinatorially while true links grow ~linearly: the signal-to-noise of unverified candidates collapses as you scale. This is the classic multiple-comparisons regime where even Benjamini-Hochberg FDR control (E[V/R]) is needed and Bonferroni FWER is "too conservative to be useful" (BH 1995). The applied evidence agrees: OpenIE's documented coverage-vs-utility tension ("cover more information... at the cost of utility and compactness," noisier triples at high recall), and the fact that automatically constructed KGs "inevitably bring in plenty of noise" requiring trustiness/error-detection layers (Entropy 21(11):1083). RECONCILIATION: volume and precision are NOT opposed — they live at different layers. Volume at extraction feeds the verifier; the verifier (scoring + evidence + contradiction + human-in-loop) is what makes volume at the hypothesis layer safe instead of a deluge. donto's architecture is literally the missing gate: every claim is evidence-anchored (so a hypothesis carries its provenance / "why"), contradictions are first-class (a generated link that rebuts existing evidence is a feature, not corruption), and the ranking step (novelty/plausibility/support/contradiction-value/verification-cost) IS the FDR controller. The boundary is therefore concrete: extract maximally (high recall, accept noisy typed claims, they're individually source-checkable), but NEVER promote a generated relationship to "believed" state without passing the scorer + evidence threshold; hold the rest as hypothesis_only.

Players / datasets / taxonomies / standards:

Volume verdict: HELPS at the typed-extraction/decomposition layer; HURTS at the unverified-relationship layer. Concretely: (1) HELPS — denser extraction raises recall of true latent links because discovery is bridge-dependent (Swanson ABC: no B-fact, no A-C link) and completion models provably degrade on sparse graphs (avg degree ~2 is where they fail). Extract everything, including implicit/inferred typed claims; each is individually source-checkable, so extraction noise is recoverable. This is the founder's correct intuition and donto is currently STARVED here (4.7% evidence coverage). (2) HURTS — candidate relationships scale ~O(N^2)/combinatorially while true links grow ~linearly, and Calude-Longo proves large data MUST contain spurious regularities by size alone, so the believed-relationship set degrades toward noise unless gated. THE BOUNDARY is a state transition, not a volume cap: a typed claim entering the graph from extraction = allowed at any volume (it's evidence-anchored and stays a claim). A generated relationship being PROMOTED from hypothesis_only to believed = must pass the scorer (novelty/plausibility/support/contradiction-value/verification-cost) under an explicit FDR budget. Volume feeds the verifier; the verifier is what makes volume safe. They are sequential, not opposed.

Design implications:

Honest caveats:

claim-hypothesis-substrate-prior-art

The "claim/hypothesis/evidence substrate" is well-trodden territory — donto must cite it, not claim it. The shipped reference point is Wikidata: 120M+ items, ~1.65B statements, with a real claim-status model — every statement carries qualifiers (time/role/context), a list of references (each a "snak"/source), and a RANK of preferred/normal/deprecated. Crucially, deprecated statements are KEPT, not deleted ("known to be erroneous but still listed... in order to prevent them being constantly added and removed") — i.e. Wikidata already ships a contradiction-tolerant, source-attributed, three-valued claim store at billion-scale. ~73% of statements have provenance. Nanopublications are the academic gold standard for the per-claim envelope: three named RDF graphs — assertion + provenance (where it came from) + publication-info (who/when minted it) — with immutable Trusty-URI hashing; >10M published, decentralized registry (Nanopub Registry / Nanodash / Knowledge Pixels). The 2025 "knowledge-provenance" extension adds a 4th graph specifically to track an assertion aggregated from a body of supporting AND conflicting evidence with truth values — i.e. nanopubs are actively bolting on what donto has natively. Bucur et al. (2021) "super-pattern" formalizes scientific claims in logic to auto-DETECT contradictory claims (evaluated on the Cooperation Databank). So the claim-level provenance + contradiction-detection problem is solved-ish; what nobody ships is contradiction as permanent first-class legal state.

On typed-claim extraction (the VOLUME layer): SemMedDB/SemRep is the scale proof — 130M+ subject-predicate-object "semantic predications" from 37M+ PubMed abstracts, concepts normalized to UMLS. It is the canonical "decompose a corpus into millions of typed claims" system — and it was DEPRECATED Dec 31 2024 (final v43), leaving a vacuum donto's extraction layer directly addresses. SciClaim (EMNLP 2021) is the fine-grained schema: typed graph annotations (causal/comparative/predictive/statistical/proportional associations + qualifying attributes), 12,738 labels, >2x the label density of prior datasets — concrete proof that one text yields dense typed structure (validates the founder's "vast latent structure" intuition). ClaimsKG (28,383 fact-checked claims → 6.6M triples) shows claims-with-truth-ratings as a queryable KG. FEVER (185,445 claims labeled SUPPORTED/REFUTED/NOTENOUGHINFO with evidence-sentence sets) is the canonical evidence-anchored verification dataset — directly analogous to scored resume↔︎job matching ("this skill-claim meets this requirement").

On argumentation + re-ranking (the HYPOTHESIS-LIFECYCLE layer): AIF (Argument Interchange Format) is the standard ontology for typed argument edges — it already distinguishes REBUT (attacks a conclusion) vs UNDERCUT (attacks the inference itself), exactly donto's supports/rebuts/undercuts; AIFdb holds 14,000+ argument maps / 160,000+ claims. The re-ranking story has a flagship 2025-26 exemplar: Google DeepMind's AI Co-Scientist (Nature, May 2026) runs a literal "tournament of ideas" — Generation → Reflection (peer-review for correctness+novelty) → Ranking (pairwise Elo debates) → Evolution (refine top hypotheses, re-enter tournament) → Proximity (dedup). This is precisely "generate hypotheses, rank by novelty/plausibility, re-rank as new evidence arrives" — but it runs in-memory per-session over literature, NOT over a persistent paraconsistent substrate. Swanson's ABC literature-based discovery (fish-oil↔︎Raynaud's) is the foundational A-B-C relationship-generation pattern, and the key reconciliation point: ABC discovery REQUIRES the intermediate B-facts to already exist in the graph — denser typed extraction (volume) is the fuel that makes A→C discovery possible. Inconsistency-tolerant KG reasoning is a 2025 survey area (paraconsistent logics, QC-negation/QCDL avoid the principle of explosion) — academically mature, but as research, not shipped infrastructure.

Players / datasets / taxonomies / standards:

Volume verdict: Volume HELPS unambiguously at the typed-extraction/decomposition layer and the prior art proves it: SemMedDB shipped 130M predications from 37M abstracts and SciClaim showed one text yields >2x prior typed-label density — the founder is correct that the latent typed structure of even one text is vast, and denser graphs are a precondition, not a luxury. Swanson's ABC is the clinching argument: you cannot discover A→C unless the intermediate B-facts physically exist in the graph, so under-extraction directly caps discovery. For jsonresume→jobs, volume is genuine network effect — more resumes+jobs → richer skill-adjacency and career-path priors → better hidden-candidate discovery. So at the recall/fuel layer, more typed claims = strictly more discoverable relationships. Volume becomes NOISE at the relationship-HYPOTHESIS layer, where combinatorics explode (N typed claims → up to O(N^2) candidate relationships) and false-discovery-rate dominates — this is exactly why every serious system in this space pairs generation with a VERIFIER/RANKER: the AI Co-Scientist's Elo tournament + reflection reviewer, FEVER's evidence-sentence requirement, Bucur's logical super-pattern for contradiction, time-aware evidence ranking for fact-checking. The reconciliation the report must state plainly: volume FEEDS the verifier, it does not compete with it. Donto's job is to make the extraction layer as high-volume as possible (the fuel) AND to put a ranker (novelty/plausibility/evidentiary-support/contradiction-value/verification-cost) and ideally a Lean-certified typed gate between candidate generation and accepted relationships. A lens or extraction pass that raises recall of typed B-facts is always valuable; a step that emits unranked, unverified A→C relationships at volume is the decorative/dangerous case the prior art uniformly guards against.

Design implications:

Honest caveats:

cross-domain-application-survey

There is a real, large, mostly-unsolved market for a substrate that holds source-grounded typed claims, keeps contradictory claims as legal state, treats identity as query-time hypothesis, and generates+ranks relationship hypotheses. The clearest existing proof is biomedical literature-based discovery (LBD): Swanson's ABC model is exactly "generate A-C relationship hypotheses from a dense graph of A-B and B-C claims," and the canonical fuel is SemMedDB — ~98M semantic predications (subject-predicate-object) auto-extracted by SemRep from ~29M PubMed abstracts. SemMedDB demonstrates donto's volume thesis (you need the intermediate B-facts to exist before any A-C hypothesis can be found) AND its core gap (SemMedDB stores predications but has weak first-class machinery for contradiction, provenance-as-byte, and identity-as-hypothesis; conflicting predications are noise, not preserved legal state). Every domain surveyed splits the same way: volume is unambiguously good at the typed-extraction/recall layer (more claims = denser graph = more candidate relationships) and dangerous at the hypothesis layer (combinatorial blowup of candidate A-C links demands a ranker/verifier or the false-discovery rate buries the signal).

The strongest "this genuinely needs all four properties" domains are: (1) LBD/drug-repurposing (SemMedDB, DRKG, Hetionet) — contradictions between studies are scientifically real, identity-as-hypothesis matters because gene/drug synonymy is messy, evidence-anchoring to the PMID/sentence is the whole game; (6) financial-crime/AML beneficial-ownership — OpenSanctions (2M+ entities, 337 sources), OCCRP Aleph + FollowTheMoney schema, Open Ownership; entity resolution IS identity-as-hypothesis (is this "John Smith" the sanctioned one?), false positives are the #1 industry pain, every match must be explainable to a regulator, and ownership changes over time (bitemporal); (4) OSINT/investigative journalism — leaked datasets contradict each other and official records, confidence tiers (confirmed/probable/possible) are already standard practice, "why do we believe X" must survive legal/editorial review; (2) legal case-law — citation graphs (Caselaw Access Project, 6.7M cases) with TYPED edges (cites/distinguishes/overrules/follows) are literally a paraconsistent argument graph where an overruled precedent stays in the record as superseded-not-deleted (bitemporal valid-time), and "what was good law on date T" is the killer bitemporal query.

The flagship jsonresume→jobs case is the cleanest non-genealogy proof because matching is literally scored relationship-discovery over typed claims: resumes and jobs both decompose into skill/role/seniority/domain/education/trajectory claims (ESCO ~3,000 occupations + ~13,900 skills; O*NET ~900 occupations as the public typed vocabularies); the match is an evidence-anchored edge ("you meet this requirement BECAUSE this skill-claim came from this resume bullet"); it is contradiction-aware (resume claims senior, tenure dates say junior; two roles overlap in time); it is bitemporal (skills decay, the market re-prices them, so old match-hypotheses must RE-RANK when a new job posting or a new resume arrives); identity-as-hypothesis handles duplicate/variant profiles and inferred-vs-stated skills; and volume genuinely produces network effects (more resumes+jobs → richer skill-adjacency and career-path graphs → hidden-candidate and "adjacent role you didn't know to apply for" discovery, which is exactly Swanson ABC applied to careers). It is monetizable today and proves donto is domain-neutral.

The honest caveats: existing players already do a lot. SemMedDB, Hetionet/DRKG, OpenSanctions/Aleph, Mem0 (agent memory with temporal validity windows), ESCO/CareerBERT — none of them combine all four properties, but each does its slice well, so donto's wedge is the COMBINATION (paraconsistent + evidence-byte + identity-as-hypothesis + bitemporal + re-ranking hypothesis lifecycle), not any single property. The places where contradiction-preservation is decorative rather than load-bearing are domains where one authoritative source dominates and disagreement is rare/uninteresting (much of routine supply-chain provenance, single-EHR clinical data without cross-source conflict). Those make weaker example projects.

Players / datasets / taxonomies / standards:

Volume verdict: Volume HELPS, decisively, at the typed-extraction/decomposition layer in every domain, and the founder is right that one text's latent typed structure is vast: SemMedDB (~98M predications from ~29M abstracts) is the existence proof that you cannot do Swanson-ABC discovery until the intermediate B-claims have been extracted at scale — denser graphs strictly enable more discoverable A-C relationships. Same logic in jobs (more skill/role claims per resume → richer skill-adjacency and career-path graphs → hidden-candidate and adjacent-role discovery = network effects), patents (more SAO function-claims → more cross-domain analogies), and AML (more ownership/payment claims → more reachable shell chains). Volume becomes NOISE at the relationship-HYPOTHESIS layer, where it is combinatorial: N typed claims yield ~O(N^2) candidate A-C edges, and without a ranker/verifier the false-discovery rate buries the few real findings — this is precisely why LBD work added contextual/semantic constraints beyond raw co-occurrence, why AML lives or dies on false-positive suppression (OpenSanctions' explainable logic-v2 matcher), and why scientific claim-verification (SciFact) exists as a separate stage. Reconciliation: volume FEEDS the verifier, it does not oppose it. Decompose maximally (high recall, cheap), then rank/verify ruthlessly (high precision, evidence-anchored). The danger isn't extracting too many claims; it's PROMOTING unranked claim-combinations to asserted relationships. donto's bitemporal re-ranking turns volume into a strength even at the hypothesis layer, because more evidence arriving over time monotonically improves the ranking of a fixed hypothesis set rather than multiplying it.

Design implications:

Honest caveats: