Does donto Work? — 105 Queries Against an Abundance-Extracted Knowledge Graph

An empirical stress-test of the donto substrate after abundance extraction: 21 analytical lenses, 105 queries, with full results and an honest good/bad analysis. 2026-06-03.

What this is. We extracted ~12,500 evidence-anchored RDF-style facts from 7 frontier-conflict event documents (Univ. of Newcastle Colonial Frontier Massacres dataset) into donto — a paraconsistent, evidence-first knowledge graph — using abundance extraction (emit free/untyped, ~1,000–2,300 facts and ~1,300 predicates per event, defer typing/alignment to query time). Then we asked: does it actually work, and is the abundance useful or just noise? We ran 105 distinct queries across 21 lenses (a 22-agent workflow), verified the strongest claims by hand, and a final judge synthesized the verdict. Every query, its result, and the analysis are below. Companion reports: the extraction-system report and the abundance vision, same /research/ dir.

Method

• Corpus: ~12,500 live statements, 6,111 distinct predicates, 725 entity-subjects across 10 contexts (7 distinct events + 3 re-extractions of event 10690 / 11176). One donto_statement table (~39M rows total); queries scoped by indexed context equality, live = upper(tx_time) IS NULL.
• Queries: 105 queries across 21 lenses, each lens an independent agent running ~5 queries and capturing the actual rows.
• Headline pass rate: 103/105 queries returned real, on-point rows (98%); 56 flagged genuinely surprising.

Verdict — does donto work?

YES — donto works as a queryable substrate, and I verified the strongest claims directly rather than trusting the lens self-reports. Every scoped query returned real, on-point rows in ~1s over a 39M-row table; the context-equality scoping discipline is the thing that makes it usable. The corpus is genuinely a knowledge graph, not a text blob: 12,589 live statements / 6,111 distinct predicates / 725 subjects across 10 contexts, with real entity IRIs, real cross-event recurrence, and real typed edges. The headline differentiator — paraconsistent holding of contradictory claims — is unambiguously real and was the single best thing I confirmed: for one event subject (ex:attack-wilmot-lagoon-near-mt-larcomb-1855) the graph simultaneously holds killedCountPerMurrayLetter=11, killCountPerCapricornian5Sept1925="over 100", countLowerBound=100, AND countDiscrepancyWith="Murray reports 11 killed and 3 wounded; Boles Jnr reports over 100 killed", plus epistemic hedges (killCountUnclearAssignment, killingsOnlyInRetrospectiveAccount, murrayLetterOnlyDescribesRecoveryNotKillings). A vector store would have collapsed these to one chunk; a normal KG would have invalidated-on-conflict and kept one winner. donto keeps all of it as legal state and lets you query the disagreement itself. The big honest caveat: it is a paraconsistent EVIDENCE STORE with WEAK entity resolution, not a clean linked-data graph. (Evidence ANCHORING, by contrast, is sound — see the correction note below: every claim is span-wired to a retrievable source snippet.) It is excellent for forensic/humanities synthesis by a human or an LLM; it is NOT yet ready for fully automated reasoning without a query-time normalization/dedup layer.

Correction (2026-06-03, post-publication). The original adversarial pass reported its single worst finding as "evidence links are wired at the weakest tier — all point at target_run_id, with target_span_id=0 and target_document_id=0, so the source snippet is not dereferenceable." That finding was FALSE — a measurement artifact. target_span_id, target_run_id, and target_document_id are UUID columns; the original query compared a UUID column against the integer literal 0 (and read a NULL UUID as "0"), which inverted the reading. Re-checked directly against live donto-pg: across the corpus's evidence links (15,078 live, all link_type='extracted_from'), every link has target_span_id SET, ZERO point at target_run_id, and ZERO at target_document_id. Every span carries real surface_text plus byte offsets and resolves span → revision → document. So evidence anchoring is SOUND — 100% span-wired and dereferenceable (claim → span → revision → document), exactly meeting donto's fact → evidence_link → span → revision → blob promise. The corrected sections below replace the original ones; verification SQL is archived at donto-align/evidence-link-correction/verify-evidence-link-tier.sql.

Is abundance useful, or noise?

USEFUL, with a real and quantified cost. The abundance is not noise — I sampled the long tail and it holds: of 6,111 predicates, ~81.7% are singletons, but they are well-formed domain predicates (no stop-words, no debug/temp junk, no parse garbage), each carrying a genuine event-specific fact. The payoff is that abundance lets you ask questions a normal KG/vector store cannot: not just "how many died?" but "how did the reported death count drift across 58 years of sources?" The deep extraction (ctx:test/ingest-verify/10690) literally computed deathCountTrendOverTime="increasing (2 in 1855 to 12 in 1913)", earliestSourcesMoreConservative=true, interSourceDisagreementOnDeathCount="severe (1-12)", deathCountRange="1-12" — these are DERIVED meta-facts about the historiography, the kind of thing you would normally pay a historian to notice. That is abundance earning its keep. BUT the cost is concrete and I measured it: (1) predicate fragmentation — same concept under multiple spellings (rdfType=899 / rdf:type=95 / "rdfType"-with-space; rdfsLabel=829 / rdfs:label=67), so a consumer MUST normalize at query time or miss data; (2) positional-suffix pseudo-arrays — the LLM emits layer1..layer6, level1..level6, link1..link6, nameVariation1..4, associatedEvent/1/2/3 instead of one multi-valued predicate; (3) IN-BAND CROSS-SOURCE ATTRIBUTION — which of several sources made a given claim is baked into predicate names (killCountPerCapricornian5Sept1925) and literals ("11 men killed per John Murray letter") rather than a reified claim→source edge, so a cross-source toll join cartesian-products (every toll × every source) instead of joining cleanly. (This is NOT an evidence-anchoring problem — every claim IS span-wired to a retrievable source snippet; see the correction note above. It is specifically about per-source attribution.) Net: abundance is a feature for a smart consumer that defers typing/alignment to query time (exactly donto's stated thesis), and a liability for a naive one.

What's GOOD — where donto shines

• Contradiction preservation — the core thesis is real and load-bearing here. Multiple incompatible casualty counts, dates (Christmas Day 1855 vs after 27 Dec 1855), locations (4 miles vs 15 miles), and event terminologies (massacre/attack/outrage/slaughter, one per source) coexist as live facts with explicit countDiscrepancyWith/dateDiscrepancyWith/eventTerminologyBySource edges. This is the one thing a vector store and a winner-takes-all KG structurally cannot do.
• Derived historiographic meta-facts — deathCountTrendOverTime, sourceDepth ('quinary (based on Leith Hay's...')', sourceTemporalDistance ('uncertain (58 years later, third person recollection)'), reliabilityRating ('not particularly reliable'). The extractor reasons ABOUT the evidence, not just from it.
• Causal/reprisal chains as graph edges — reprisalFor, provokedReprisal, triggeredBy, subsequentReprisal, plus killedToTriggerDeathRatio='at least 100 killed in reprisal for 5 deaths' and full 6-step causationChainDetailed prose. You can traverse trigger→reprisal→secondary-reprisal across events.
• Cross-event entity recurrence — ex:nmp appears across 5 contexts (127 statements), Frederic Charles Urquhart and Burke Pastoral District across 4, building usable prosopographies and command hierarchies (Captain O'Connell → Lt Murray → Sgt Boles → Constables) from independently-extracted documents.
• Decolonial framing as data — euphemismDecoded ('paid the penalty → killed', 'dispersed → killing'), perpetratorsFramedAs='the avengers' vs victimsFramedAs='the black murderers', colonialJustification, framingBias. Period bias is captured as structured fact rather than editorialized away — directly usable for critical historiography.
• Honest uncertainty — possibleEvent=true, dateDisputed=true, allegedlyUnprovoked, certaintyMarker='confirmed by depositions'/'supposed'. The graph marks unknowns visibly instead of fabricating confidence.

Best findings (verified)

• VERIFIED: deathCountTrendOverTime = 'increasing (2 in 1855 to 12 in 1913)' on ex:attack-rannes-1855 (ctx:test/ingest-verify/10690), alongside earliestDeathCount=2, latestDeathCount=12, deathCountRange='1-12', earliestSourcesMoreConservative=true, interSourceDisagreementOnDeathCount='severe (1-12)' — the extractor performed longitudinal source analysis and emitted derived meta-facts about how the historical record itself drifted. This is the strongest argument that abundance > a flat KG.
• VERIFIED paraconsistency on one subject: killedCountPerMurrayLetter=11 AND killCountPerCapricornian5Sept1925='over 100' AND countLowerBound=100 AND countDiscrepancyWith='Murray reports 11 killed and 3 wounded; Boles Jnr reports over 100 killed' — all live, all queryable, none collapsed. plus self-aware hedges like murrayLetterOnlyDescribesRecoveryNotKillings=true and killingsOnlyInRetrospectiveAccount=true.
• VERIFIED dense identity hypotheses capturing real OCR/colonial-renaming ambiguity: maconachie/maconochie/maccronerky → john-mcconachie; kliediewarry/kalidawarry/kalidawarry-waterhole → kaliduwarry; james-bowles-nmp/mr-bol/white-sergeant-nmp → james-boles-nmp. These are exactly the messy variants a frontier-records historian fights, surfaced as queryable likelySameAs/possiblySameAs edges (with confidence grading: possiblySameAs for the uncertain ones).
• VERIFIED structured evidence links EXIST, are dense, AND are anchored at the STRONGEST tier — 15,078 live evidence links across the corpus (all link_type='extracted_from'), and every one points at a target_span_id, with ZERO at target_run_id and ZERO at target_document_id. Each span carries real surface_text plus start/end byte offsets and resolves cleanly span → revision → document, so a claim dereferences to its exact source snippet via a join — donto's fact → evidence_link → span → revision → blob promise is met at the span tier, not the run tier. (The original pass reported the opposite — span=0, doc=0, points at run — which was a UUID-column-vs-integer-0 measurement artifact, now corrected.)
• MEASURED the abundance cost: predicate fragmentation is real (rdfType=899 vs rdf:type=95 vs a space-prefixed rdfType; rdfsLabel=829 vs rdfs:label=67) and the LLM fakes arrays with positional suffixes (layer1..6, level1..6, link1..6, nameVariation1..4). A cross-source toll join cartesian-products because tolls and sources both hang off the event subject with no claim→source reification.

What's BAD — where it falls short

• Provenance attribution (which SOURCE made a claim) is in-band, not reified. NOTE: this is a narrower point than the original report's #1 "bad" finding, which was withdrawn — evidence ANCHORING is fine (every claim is span-wired to a retrievable snippet via fact→evidence_link→span→revision, see the correction note above). What is NOT structured is cross-source attribution: a claim resolves to a source document, but WHICH of several sources asserted a given fact is encoded in predicate names (killCountPerCapricornian5Sept1925) and literal text ('11 men killed per John Murray letter'). That is human/LLM-readable but not a clean per-source edge, so a "who said which toll?" query can't join on it cleanly (see next bullet).
• No structured claim→source reification, so multi-source facts don't join cleanly. A natural query ('which source said which toll?') cartesian-products because both the toll and the source hang off the event subject as independent predicates. You must parse predicate names/literals to recover who-said-what. (This is the genuine provenance gap; it is about source reification, NOT about whether evidence is dereferenceable — it is.)
• Predicate fragmentation forces query-time normalization. Same concept under rdfType / rdf:type / ' rdfType', rdfsLabel / rdfs:label, describedAs / describedAsHaving / etc. Any query that doesn't regex-normalize silently undercounts.
• Positional-suffix pseudo-arrays (layer1..6, level1..6, link1..6, nameVariation1..4, associatedEvent/1/2/3) — the LLM uses numbered predicates as a fake array, which is both fragmentation and a sign of unconstrained emission. These should be one multi-valued predicate.
• Entity resolution is hypotheses, not resolved. likelySameAs edges are dense and bidirectional but require union-find consolidation at query time; the same pair can be linked by both aliasOf and likelySameAs to different targets. Until resolved, ex:combo-james vs ex:combo-jimmy vs ex:colin-killed remain distinct nodes.
• Subjects are NOT stable across extraction runs — the same Rannes attack is ex:attack-nmp-rannes-1855-09-23 in one context and ex:attack-rannes-1855 / ex:ranges-station in others; locations vary (Rannes vs Ranges). Cross-extraction-consistency must be done with loose regex + manual inspection.
• 64.5% of objects are literals, not IRIs (verified ~35.5% IRI), so graph TRAVERSAL is shallow — much of the value is text packed into object_lit that you read, not edges you walk. The in-degree distribution is also dominated by the boolean literal 'true' (2620), which drowns real entity hubs unless filtered.

Killer examples (concrete query → result)

• PARACONSISTENCY (verified, ctx:.../46860): query subject=ex:attack-wilmot-lagoon-near-mt-larcomb-1855 → returns killedCountPerMurrayLetter='11', killCountPerCapricornian5Sept1925='over 100', countLowerBound='100', countDiscrepancyWith='Murray reports 11 killed and 3 wounded; Boles Jnr reports over 100 killed', killCountUnclearAssignment='not clear whether 11 killed were all associated with the second reprisal or with both'. Both numbers held live; the conflict is itself a first-class fact.
• DERIVED HISTORIOGRAPHY (verified, ctx:test/ingest-verify/10690): subject=ex:attack-rannes-1855 → deathCountTrendOverTime='increasing (2 in 1855 to 12 in 1913)', earliestDeathCount='2', latestDeathCount='12', deathCountRange='1-12', earliestSourcesMoreConservative='true', interSourceDisagreementOnDeathCount='severe (1-12)', earliestToLatestSourceSpan='1855 to 1975'. No vector store produces this; it is a computed claim about source drift.
• IDENTITY UNDER OCR/RENAMING (verified): predicate likelySameAs → maconachie/maconochie/maccronerky→john-mcconachie, kliediewarry/kalidawarry/kalidawarry-waterhole→kaliduwarry, james-bowles-nmp/mr-bol/white-sergeant-nmp→james-boles-nmp. Real spelling-variant resolution surfaced as graded hypotheses (likelySameAs vs possiblySameAs).
• TERMINOLOGY-BY-SOURCE (lens-reported, consistent with verified discrepancy machinery): ex:attack-rannes-1855 eventTerminologyBySource → 'massacre (Morning Bulletin 1912)', 'attack (dataset)', 'outrage (Empire Oct 15)', 'slaughter (De Satge 1901)'. Same event, four framings, each attributed — exactly the meta-data a historian needs.
• THE JOIN THAT EXPOSES THE LIMIT (verified): joining toll predicates to source predicates on ctx:.../46860 returns a cartesian product — 'over 100' paired with capricornian-1925-09-05-p60 AND james-boles-jnr AND john-murray-nmp AND capricornian-newspaper indiscriminately, because there is no claim→source reification. The who-said-what is real but only recoverable by reading the predicate name/literal, not by a clean SQL join.
• EVIDENCE-LINK REALITY CHECK (verified, corrected): joining donto_evidence_link → donto_span → donto_document_revision → donto_document over the corpus → 15,078 live rows, ALL link_type='extracted_from', ALL targeting target_span_id (ZERO at target_run_id/target_document_id), every span carrying real surface_text+offsets that resolve to a source document (e.g. offset 1958–2024 → 'Rockhampton Bulletin 10 July 1873, p3; Telegraph, 16 July 1873, p3'). Evidence wiring is near-total in coverage AND anchored to a retrievable source span — donto is genuinely 'evidence-anchored', not merely 'evidence-aware'. (The original reading of 'span=0, doc=0, points at run' was a UUID-vs-integer-0 artifact and is withdrawn.)

Per-lens scorecard

Appendix — all 105 queries & answers

Per lens: each query's intent, the actual sample rows donto returned, and whether it worked / surprised.

contradictions

VERDICT: donto SUCCESSFULLY handles the contradictions lens. It genuinely holds both sides of conflicts: / / STRENGTHS: / 1. Paraconsistent by design: The graph explicitly stores contradictory claims as separate facts. Query 1 returned 20 contradiction predicates (dateDisputed, dateDiscrepancyWith, countDiscrepancyWith, conflictsWith, threeConflictingDates). These are queryable, precise, and a…

cross-event-entities

EXCELLENT SIGNAL. The cross-event-entities lens reveals that donto's graph is HIGHLY QUERYABLE and reveals genuine frontier-war structure: (1) Key institutional entities like ex:nmp (Native Mounted Police) recur across 5 different event contexts with 127 total statements—providing rich, usable profiles of colonial agents. (2) Named individuals like Frederic Charles Urquhart appear in 2+ distinct e…

euphemism-decode

The euphemism-decode lens WORKS EXCELLENTLY. Donto captures the scholarly core: period euphemisms (dispersal, lesson, punish, paid the penalty) paired with decoded meanings (killing, mass killing, reprisal killings). The graph structure includes direct mappings, victim/perpetrator framing reversals, source-anchored quotations, explicit bias annotations, and causal chains linking euphemisms to casu…

motive-tactics-resistance

STRONG PERFORMANCE. The donto graph successfully extracted and linked: / / 1. Motive inference: The model identified distinct motive types (reprisal, punishment, resistance to encroachment, NMP troop retaliation) for different actors—Aboriginal attackers, settler reprisal parties, and NMP detachments. The "motivationTargetingNmpOnly" statement shows nuanced analysis of intra-colonial tension (…

provenance-epistemics

DONTO HANDLES PROVENANCE-EPISTEMICS VERY WELL. The graph is NOT noise. Findings: (1) Rich attestation network: 73 attestedBy statements show witness/source linking; 37 corroboratedBy statements explicitly track cross-validation. (2) Source taxonomy working: Predicates like sourceDepth, sourceReliability, sourceTemporalDistance are populated with semantically meaningful values (primary/seco…

casualties-quantities

donto handles the casualties-quantities lens well. The graph successfully extracts and stores casualty counts (numberOfPeopleKilled, reportedDeathTollPerSource) with explicit qualifiers (approximate, lowerBound, disputed), tracks source multiplicity (Murray vs Boles Jnr vs Capricornian with differing figures), and resolves body treatment in resolvable chains (buriedBy→agent→further metadata). Coun…

nmp-structure

donto successfully reconstructs NMP COMMAND HIERARCHY and CAMP TOPOLOGY from frontier-conflict events. Queries 1, 2, 4, 5 returned clean, on-point structure: Captain O'Connell → Lt. John Murray → Sgt. James Boles → Constables; 4 named camps (Gladstone, Rockhampton, Burke River/Boulia, Cloncurry) with stationed officers and geographic containment. Query 3 (trooper recruitment) FAILED: regNo (regist…

kinship-social

donto's kinship-social lens is HIGHLY QUERYABLE and GENUINELY USEFUL. The graph successfully captures: (1) genealogical parent-child links with correct directional predicates; (2) rich marriage records including dates, locations, and relationship variants; (3) occupational hierarchies with rank and status distinctions; (4) employment chains and land ownership networks showing economic power struct…

spatial-jurisdiction

donto successfully reconstructs the spatial-colonial hierarchy (site → pastoral run → pastoral district → colony) with clean, queryable containment predicates. Station ownership chains are solid. Native title boundary links exist and are valuable for matching colonial events to modern land claims. Aboriginal country assignments (onCountryOf, countryIncludes) are present but signal is weak due to a…

temporal-sequence

SOLID GRAPH, genuinely useful for temporal analysis. The donto extraction captured: (1) Precise dates at multiple granularities (ISO dates, years, day-month phrases). (2) Robust event sequencing (precededBy/followedBy) across multi-event conflict chains. (3) Rich reprisal networks (reprisalFor, causedBy, triggeredBy, subsequentReprisal) that link trigger events to responses with full temporal cont…

causation-second-order

DONTO HANDLES THIS LENS EXCEPTIONALLY WELL. The causation-second-order graph is QUERYABLE and USEFUL. / / STRENGTHS: / 1. EXPLICIT CAUSAL PREDICATES: The extractor invented rich, domain-specific predicates (causedBy, triggeredBy, inResponseTo, provokedReprisal, organisedAfterKillings) that directly encode causal intent. Not generic RDF—tailored to frontier violence narratives. / / 2. MULTI-HOP NAR…

novel-predicates

Donto's novel-predicate long tail is GENUINELY USEFUL and semantically coherent. Of 6,111 unique predicates extracted across 12,589 statements, 4,995 (81.7%) appear exactly once—a extreme abundance phenomenon. However, these are NOT noise: they are carefully formed, domain-specific predicates (aboriginalLaborExploitation, aboriginalPartyTrackedBySheepTrail, accusedOfFencingOutAboriginalPeople) tha…

sources-as-entities

Donto handles the sources-as-entities lens VERY WELL. The graph successfully captures: (1) source entities as first-class RDF subjects with metadata (publication dates, names, reliability ratings, source types); (2) verbatim quotes attributed to sources via quotedAsSaying; (3) source-to-event mappings (reports, reportsKillCount, reportsSurvivorCount); (4) attestation links (attestedBy); (5) explic…

bias-framing-language

STRONG PASS. Donto successfully captures period-source bias and framing as structured facts, not as editorial judgment. The graph models: / (1) Asymmetric characterization—Aboriginal subjects as "murderers," "savage," "wild"; colonials as "avengers," "victims," "performing duty" / (2) Inflammatory vocabulary—"outrage," "depredations," "barbaric"—reliably extracted and attributed to events / (3…

identity-hypotheses

VERDICT: EXCELLENT QUERYABLE GRAPH. The identity-hypotheses lens is genuine, rich, and immediately actionable. / / STRENGTHS: / - ~20 core identity predicates (sameAs, likelySameAs, aliasOf, alsoKnownAs, nameVariant, spellingVariant, nameVariantOf, possiblySameAs, hasSpellingVariant, nameSpellingVariant, spellingVariantOf, etc.) are consistently present and semantically coherent / - Identi…

analytics-topology

Donto IS queryable and captures real semantic structure, but with significant limitations. STRENGTHS: The graph correctly identifies focal entities (massacre events, attack IRIs, witnesses), respects domain predicates (discrepancy tracking, quoted testimony, name variants), and shows measurable variance across extraction contexts (ingest-verify is 2.5x denser than agnostic in statements/subject). …

single-event-deep

donto handles single-event-deep VERY WELL. Event 10690 is fully reconstructable across 5 critical axes: geographic/identity anchors, timeline and casualties, people (NMP killed by name and regiment), source provenance (primary letters 1855, 7 newspapers spanning 58 years, secondary sources to 1975, archive reels), and explicit contradictions (6 sources disagree on death count; Christmas Day date e…

cross-extraction-consistency

Donto returns REAL data on core facts (event date, victim count, victim names agree across runs), but the graph is severely fragmented by predicate variance and entity deduplication failure. Same facts are encoded with ~10+ different predicate names (killedIn vs killedInEvent vs hasKilledCount vs numberOfPeopleKilled vs killedPerEmpireOct15). Same victims get different IRIs (ex:combo-james vs ex:c…

weapons-methods

The weapons-methods lens is highly functional and genuinely useful for historical violence research. Donto successfully extracted weapon types (Snider rifles, spears, poison), violence methods with source attribution, specific injury descriptions (named individuals with wound locations), and kill counts per event. Weapon entities are modeled as RDF objects with operational properties (stacked nigh…

surprising-emergent

VERDICT: donto is GENUINELY USEFUL for this corpus. The surprise-emergent lens reveals a working, queryable knowledge graph. Strengths: (1) witness reliability chains are captured as semantic relations, not just name-date pairs; (2) perpetrator-victim-location triangulation works well—I can trace who killed whom on whose country; (3) temporal disputes are captured as first-class facts (dat…

adversarial-junk

donto does NOT produce significant junk. The extraction is FAITHFUL rather than hallucinating false confidence. Booleans are intentional semantic markers (possibleEvent, dateDisputed). Duplicates (3–5 counts) are expected cross-referencing of shared entities. Predicates are well-formed, descriptive, domain-appropriate camelCase (no corruption, no debug/temp prefixes). The most "noisy" finding—ille…

Conclusion & implications

donto works as a queryable, paraconsistent, evidence-first substrate, and abundance extraction earns its keep — it answers questions (held contradictions, cross-source death-toll drift, decoded euphemisms, reprisal causation, prosopography) that a vector store or winner-takes-all KG structurally cannot. The honest caveats are concrete and actionable, not fatal:

1. Evidence anchoring is SOUND — every claim dereferences to its source snippet (CORRECTED). The original report's #1 "bad" finding — "evidence links are wired at the weakest tier (point at the run; span=0, doc=0; not dereferenceable)" — was FALSE, a UUID-column-compared-to-integer-0 measurement artifact. Re-verified on live donto-pg: of the 15,078 live evidence links, every one points at a target_span_id (zero at target_run_id/target_document_id); every span carries real surface_text + byte offsets and resolves span → revision → document. So claim → exact source snippet is already a join, not a string-parse — donto's fact → evidence_link → span → revision → blob model is met at the span tier. The remaining gap is narrower and is item #3: cross-source attribution (which of several sources made a claim) is in-band, not a reified claim → source edge.
2. Predicate fragmentation is real — same concept under multiple spellings (rdfType=899 vs rdf:type=95) and positional pseudo-arrays (layer1..6, nameVariation1..4). A consumer must normalize at query time (exactly donto's defer-to-query-time thesis) — but a query-time alignment/dedup layer is now a hard requirement, not optional.
3. Cross-source attribution should be reified — evidence anchoring is already sound (item #1), but which source made a claim is baked into predicate names (killCountPerCapricornian5Sept1925) rather than a reified claim → source edge, so cross-source joins ("who reported which toll?") cartesian-product. A claim → source edge would fix it. This is now the highest-value provenance fix (it replaces the original report's withdrawn "weakest-tier" claim as the real provenance gap).
4. Abundance is a feature for a smart (LLM/human) consumer and a liability for a naive one — which is precisely donto's stated design point. The 98% query pass-rate and the derived historiographic meta-facts (deathCountTrendOverTime = 'increasing (2 in 1855 to 12 in 1913)') are the proof it's worth it.

Generated 2026-06-03 from a 22-agent, 105-query workflow over the live donto substrate.