A research report on agentic polyperspectival decomposition, paraconsistent holding, and the architecture of machine serendipity — grounded in a 10-area literature study and four adversarial critiques.
Strip the founder's idea down to its load-bearing claim and it is not about volume at all. It is not "a million facts about a thing." The premise is depth of decomposition through many lenses, in service of relationship discovery:
Use AI agents to break any entity, text, or event down through the full spectrum of human analytical lenses — philosophical, linguistic, temporal, causal, mereological (part/whole), teleological (purpose), ethical, aesthetic, economic, ecological, semiotic, phenomenological, social, material — each pushed toward the utmost of human understanding. The payoff is not the facts inside any one lens. It is the relationships between entities that emerge at the intersection of lenses — connections no human ever drew, because no human holds all the lenses at once over all the entities.
Two things make this precise enough to engineer. First, the unit of value is a relationship, not a fact — specifically a relationship that becomes visible only when entity A's decomposition under lens X is laid alongside entity B's decomposition under lens Y. Second, the generative mechanism is the cross-product: many entities × many deep lenses, mined at the seams. The fact-extraction is just scaffolding; the gold is in the joins.
I will call this system the lens engine, and the broader research program polyperspectival decomposition. Both names matter. "Lens" keeps faith with the founder's own metaphor and with a 90-year tradition (Ranganathan's facets, Pustejovsky's qualia, Nietzsche's "affects") of treating an analytical viewpoint as an aperture that reveals one structure and suppresses others. "Polyperspectival" insists on the plurality and on the fact that the perspectives are incommensurable by design — that is where the value lives. A third name worth holding in reserve is bisociative substrate, because, as we will see, Koestler already named the core move in 1964 and donto's paraconsistent store is the first place that move can be persisted rather than performed once and discarded.
The honest reframing the founder has already half-made is the correct one: this is a serendipity / latent-structure / discovery engine, and its success metric is novel-and-verified-and-valuable relationships per unit of curation cost, never fact count. Everything below is in service of that single metric — and of being brutally clear about where it has been pursued before, where it breaks, and what is genuinely unclaimed.
The founder's instinct that "no one has thought to do this" is, at the level of the idea, false — and that is the most useful gift this report can offer, because it means the hard-won lessons of seven distinct research traditions are available for free. The components are old. Let me map the lineage honestly, tradition by tradition, naming what each already achieved.
The single most direct ancestor is Don R. Swanson's Literature-Based Discovery (LBD). His 1986 concept of undiscovered public knowledge is almost verbatim the founder's thesis: knowledge logically implied by the union of two existing literatures, unknown only because no one read both. Swanson formalized it as the ABC model — if literature reports A→B (Raynaud's disease ↔︎ elevated blood viscosity) and a separate, non-co-citing literature reports B→C (fish oil ↓ blood viscosity), then a testable A→C link (fish oil treats Raynaud's) is latent in the disjoint corpora. He published the fish-oil/Raynaud's hypothesis in 1986; a 1989 trial confirmed it. His 1988 "Migraine and Magnesium: Eleven Neglected Connections" found convergent multi-path support — eleven independent indirect links — for magnesium-deficiency→migraine, later clinically supported.
These are existence proofs that the method finds real, non-obvious, testable discoveries from text alone. The field then split discovery into the two modes that map exactly onto the founder's use-cases: open discovery (start from A, fan out to find unexpected C's — "serendipity mode") and closed discovery (given a suspected A and C, find the B-paths that explain it — "verification mode"). Swanson & Smalheiser's Arrowsmith, Hristovski's BITOLA, Petric's RaJoLink, and Cambridge's LION LBD (2019, ~27M PubMed abstracts) built ranking machinery — Jaccard, normalized PMI, symmetric conditional probability, chi-squared, log-likelihood. SemMedDB / Semantic MEDLINE (Kilicoglu & Rindflesch, NLM) replaced bag-of-words with ~130M typed predications (Drug-X TREATS Disease-Y), a direct precursor to donto's quad structure (NLM deprecated it 31 Dec 2024).
What LBD learned, and donto must inherit: ranking is the entire game. LION honestly reports that true targets often sit at rank 56–299 (closed) or rank 15–120,000 (open) — the signal is real but buried. The field's only non-gameable evaluation is time-slicing (train on pre-year-Y literature, test whether top-ranked links were actually published after Y). And the field is in a reproducibility crisis: the 2025 "Make LBD Great Again Through Reproducible Pipelines" movement (SKiM and others) exists precisely because results swing wildly with corpus, preprocessing, and split choices. Crucially, a 2023 Bioinformatics critique (Sebastian/Moreau) showed time-slicing's gold standard is "dominated by meaningless co-occurrences (Ebolavirus + Professional Burnout)" — the true-discovery fraction is "unknown and likely low." After four decades, the validated trophy case is still essentially Swanson's two hand-curated findings, found by a careful human reasoner, not by combinatorial enumeration. That is a warning written in the founder's own handwriting.
The founder's "connections no one thought of, because no one holds all the lenses at once" is, almost word for word, Arthur Koestler's bisociation (The Act of Creation, 1964): every creative act — the comic Haha, the scientific Aha, the artistic Ah — is perceiving one situation in two self-consistent but habitually incompatible matrices (M1, M2) at once. Ordinary thought is association (within one plane); creativity is bisociation (the collision of two). "The value is the relation that springs from intersecting two frames" is the founder's thesis, stated sixty years ago.
And it was operationalized. The EU FP7 BISON project (2008–2012; Berthold, Lavrač, Mladenić, Borgelt; Springer LNCS 7250, 2012) built BisoNets (heterogeneous evidence graphs) and three formal bisociation types — bridging concepts, bridging by structural similarity (analogy), bridging by graphs (connecting subgraphs) — and tried to make "bisociativeness" a rankable score. Its running tool, CrossBee (Juršič, Cestnik, Urbančič, Lavrač, ICCC 2012, crossbee.ijs.si), takes two document sets and ranks candidate bridging b-terms by an ensemble of heuristics, then lets a human verify. That is donto's intended workflow — over-generate cross-context links, curate the rare valuable ones — already built, and already stuck on the same wall: too many candidates, the expert is the bottleneck.
Running alongside is Fauconnier & Turner's Conceptual Blending (The Way We Think, 2002): two input mental spaces project selectively into a blend via a generic space, and the blend develops emergent structure present in neither input. This is the precise theoretical name for "a relationship that emerges only at the intersection of lenses." It too was made algorithmic — Goguen's category-theoretic Unified Concept Theory, Pereira's Divago, and the EU COINVENT project (Schorlemmer, Kutz, Confalonieri, Pease) modeling blends as amalgams / colimits. Their hard-won lesson: the bottleneck is selection/optimality among a combinatorial explosion of possible blends, not generation. And Margaret Boden's taxonomy (The Creative Mind, 1990) gives the founder a roadmap: cross-lens links are mostly combinational creativity; pushing a lens to its limit is exploratory; a lens that rewrites another's assumptions is transformational.
Comparing two entities under a lens is, formally, an analogy. Dedre Gentner's Structure-Mapping Theory (1983) says an analogy maps a system of relations from base to target, governed by the systematicity principle: prefer deep, interconnected higher-order (causal/mathematical) structure over surface attributes. The Structure-Mapping Engine (Falkenhainer, Forbus & Gentner, 1989) returns not just correspondences and a score but candidate inferences — novel claims projected from base onto target. That candidate-inference step is the literal generator of "a relationship no one thought to draw."
Douglas Hofstadter's FARG tradition (Fluid Concepts and Creative Analogies, 1995; Surfaces and Essences, 2013) makes the radical claim that all thinking is analogy-making, and that representation is constructed under pressure, not given — Copycat's Slipnet/Workspace/codelets/temperature is the cognitive analogue of donto's identity-as-hypothesis. Holyoak & Thagard's multiconstraint theory (ACME/ARCS) adds that analogy is governed by soft, purpose-weighted constraints, not strict isomorphism.
This tradition scaled. Hope, Chan, Kittur & Shahaf's "Accelerating Innovation Through Analogy Mining" (KDD 2017, Best Paper) learned purpose and mechanism embeddings so analogies could be mined from messy patent corpora; SOLVENT (2018) extended it to scientific papers (background/purpose/mechanism/findings). This is the closest existing relative of donto's vision at the document level — but it uses one coarse facet schema (purpose × mechanism), not a full spectrum, and holds no contradictions. The 2023–2026 LLM wave reopened it: Webb, Holyoak & Lu (Nature Human Behaviour 2023) found "emergent analogical reasoning" in GPT-4; Lewis & Mitchell (2024) showed it collapses on counterfactual (permuted-alphabet) variants where humans stay robust — evidence that LLM analogies are often training-data echoes, not structural mapping. YARN (2026) splits the difference with the template donto should steal: LLM-abstraction-then-structural-alignment beats both pure LLM (0.46 vs 0.41) and pure SME (which scores below random, 0.17) on far analogies.
S. R. Ranganathan's Colon Classification (1933) and its PMEST facets — Personality, Matter, Energy, Space, Time — were the first analytico-synthetic scheme: decompose a subject into orthogonal facets, then synthesize compound subjects no one enumerated in advance. The deep claim — a small set of orthogonal facets composes to an unbounded space — is the founder's combinatorial generativity, stated in 1933. PMEST is the historical ancestor of modern faceted search (Marti Hearst's Flamenco).
The emergence engine is Rudolf Wille's Formal Concept Analysis (1981; Ganter & Wille 1999). From a binary object × attribute table, a Galois connection yields formal concepts (maximal extent/intent pairs) ordered into a concept lattice — emergent concepts the analyst never named — plus a canonical basis of attribute implications (A→B). This is literally a machine that surfaces latent concepts and rules from a matrix. Its multi-relational extension, Relational Concept Analysis (Rouane-Hacène, Huchard, Napoli, Valtchev, 2013), iterates over several object sorts linked by relations — structurally what cross-entity, cross-lens discovery aims at. Bo Xiong's 2026 Lattice Representation Hypothesis even argues LLM embeddings already encode FCA-style concept lattices (71–83% F1 recovering concept-attribute relations). The brutal caveat: classical FCA needs exact binary incidence, is noise-sensitive, and is worst-case exponential.
The deepest formal warrant that "decompose every entity through fixed lenses" is principled rather than arbitrary comes from semantic-decomposition primitives. Wierzbicka's Natural Semantic Metalanguage reduces all meaning to ~65 universal primes — so two concepts from different cultures become comparable at the prime level (the discipline of a shared decompositional alphabet, without which cross-entity discovery degrades to string matching). Pustejovsky's Generative Lexicon (1991/1995) is the most lens-like precedent: its qualia structure assigns every noun four Aristotelian modes of explanation — formal (kind), constitutive (parts — mereology), telic (purpose — teleology), agentive (origin — causation). A book is formal=physical-object, constitutive=pages/text, telic=read(x), agentive=write(x). And its dot-object (book = PHYSICAL•INFORMATION — one entity legitimately under two co-present types) is a near-exact precedent for donto's "identity is a hypothesis." Schank's Conceptual Dependency, Jackendoff's Conceptual Semantics, Dowty's proto-roles, and the modern, graded Universal Decompositional Semantics (White et al., 2016–2020 — many scalar, confidence-weighted, orthogonal property layers on one graph) all teach the same throughline: once two entities are decomposed into a shared primitive vocabulary, latent cross-entity relations (shared telic purpose, matching proto-role profiles) become computable rather than guessed. The standing objection — Fodor & Lepore's atomism ("decomposition smuggles in unverifiable structure") — is the one the founder must answer: does a machine-proposed decomposition carry falsifiable content, or just rename the problem?
This is the tradition the founder most needs and least expects, because it answers the question which generated connections are valuable — and it contradicts a naive reading of the vision. Granovetter's "Strength of Weak Ties" (1973): novel information flows across bridges, not within dense clusters. Burt's structural holes (1992; "Structural Holes and Good Ideas," AJS 2004, the Raytheon study): people whose networks span holes have a measurable vision advantage; "the creative spark on which serendipity depends is to see bridges where others see holes." Uzzi, Mukherjee, Stringer & Jones, "Atypical Combinations and Scientific Impact" (Science 2013, 17.9M papers) is the keystone scoring result: the highest-impact papers are not the most novel — they sit in the high-conventionality + high-tail-novelty quadrant. A bedrock of convention with one sharp atypical intrusion is 2× more likely to be a hit. Pure novelty underperforms.
Foster, Rzhetsky & Evans (ASR 2015) showed scientists rationally avoid risky bridges (the impact premium doesn't offset the chance of being ignored) — the market gap a machine can exploit, bearing risk humans won't. Wang, Veugelers & Stephan (2017) showed the most novel papers are systematically undervalued short-term and cited mostly in "foreign" fields — the strongest external validation of "no single evaluator holds all the lenses." And Chen, Ding & Evans ("New Directions Emerge from Disconnection and Discord," 2021) found new directions emerge where clusters are both disconnected and in discord — direct empirical warrant that a contradiction frontier is a map of where novelty pays off, not noise to be resolved. Sourati & Evans (Nature Human Behaviour 2023) built "human-aware" random walks that avoid the crowd to surface "alien" hypotheses — improving prediction of real discoveries by up to 400%.
Finally, the systems that already do most of the founder's pitch. SciAgents (Ghafarollahi & Buehler, MIT; Advanced Materials 2025) samples random (not shortest) paths through a 33K-node ontological KG and runs an Ontologist → Scientist → Critic team to surface "hidden interdisciplinary relationships previously considered unrelated" — the single most architecturally aligned prior system. Google DeepMind's AI Co-Scientist (Nature 2026) runs Generation / Reflection / Ranking / Evolution / Proximity (clusters specifically to prevent collapse into one line of thought) / Meta-review agents with an Elo tournament of simulated debates, and produced wet-lab-validated results: an AML / liver-fibrosis drug-repurposing candidate (~91% of a scarring response blocked at Stanford), and an independently re-derived antimicrobial-resistance mechanism that had taken an Imperial lab years. Robin (FutureHouse) ran end-to-end to a validated dry-AMD candidate (ripasudil) in ~2.5 months. mat2vec (Tshitoyan et al., Nature 2019) proved latent future relationships are already encoded in a corpus — word2vec over 3.3M abstracts recommended thermoelectric materials years before discovery, with no symbolic lenses at all.
Here is the lineage in one table:
| Tradition | Key figures / systems | What it already achieved | What it did not do |
|---|---|---|---|
| Literature-Based Discovery | Swanson 1986; Arrowsmith; LION; SKiM | Cross-literature A→C discovery, clinically validated; open/closed modes; time-sliced eval | Many lenses; persist rejected candidates; identity-as-hypothesis |
| Bisociation / Blending | Koestler 1964; BISON/CrossBee; Fauconnier-Turner; Boden; COINVENT | Named the core move; built rankable cross-domain bridging; emergent-structure theory | Scale; full lens spectrum; paraconsistent hold |
| Analogy / structure-mapping | Gentner SME; Hofstadter; Hope/Shahaf/SOLVENT; YARN | Candidate-inference generation; facet-embedding mining at scale; LLM-abstract-then-align | >1 facet schema; contradiction-holding; evidence anchoring |
| Faceted classification + FCA | Ranganathan 1933; Wille 1981; RCA 2013 | Orthogonal-facet generativity; automatic emergent concepts + implications from a matrix | Noise tolerance; web scale; agentic fill |
| Semantic decomposition | Wierzbicka NSM; Pustejovsky GL/qualia; UDS | Per-entity lens decomposition (4 qualia); graded multi-property layers; dot-objects | Full spectrum beyond linguistics; paraconsistency |
| Science of value | Granovetter; Burt; Uzzi; Foster/Evans; Chen/Ding/Evans | Computable value signals: brokerage, z-score atypicality, discord-as-generative | A substrate to apply them to machine-proposed edges |
| Agentic discovery | SciAgents; AI Co-Scientist; Robin; mat2vec | Generate-debate-rank-evolve with wet-lab-validated novel discoveries | Persist the firehose; hold contradictions; formal certification |
The gift is plain: the discovery idea, the bisociation move, the facet idea, the value math, and even the agentic generate-rank-evolve loop are all prior art. "No one has thought of this" is false at the level of every component. Which is exactly why the genuine novelty must be located precisely.
The four adversarial critiques converge — all at confidence ~0.6–0.72, all verdict partially-holds — on the same answer. The novelty is not any single ingredient and not the discovery itself. It is a specific, defensible combination:
Agentic, many-deep-lens generation × paraconsistent, evidence-anchored, bitemporal HOLDING of the entire speculative (and mutually-contradictory) relationship frontier × a formal/empirical VERIFICATION ladder that promotes the rare valuable survivors — operated at substrate scale.
Let me be exact about each leg, and honest about what the critiques refuted.
The defensible core: generate-HOLD-verify, where every prior
system can only generate-and-discard. This is the strongest,
most consistently affirmed point across all four critiques. SciAgents
generates and discards into prose. LBD/analogy engines rank candidates
and throw away the rest. AI Co-Scientist keeps its tournament population
only in memory for one run. Cyc had the contradiction-tolerant
store (microtheories) but no agents and famously collapsed under manual
curation. donto is the first place where the 99.9% of
machine-proposed relationships that everyone else discards can
persist — as hypothesis_only edges, with typed
argument edges
(supports/rebuts/undercuts),
evidence-anchored to source bytes, on a bitemporal store, re-rankable
forever as the corpus and lenses evolve. The science-of-value tradition
supplies the killer argument for this: Chen/Ding/Evans showed
discord is generative — so a consistency-enforcing
store would destroy the very signal that predicts new
directions. donto's contradiction frontier is not a tolerated
defect; it is a priority map. This is genuine white space, and
it is the moat.
The qualified leg: paraconsistency + identity-as-hypothesis
is the right home, but today the moat is mostly schema, not
populated reality. Critique #3 is the sharpest and most useful.
It verified the substrate is real and at scale (39,496,776 current
statements, genuinely bitemporal), and that the design beats a "vector
DB + reranker" in principle for the stated job — an ANN index
has no native typed edge, no first-class "this claim attacks that
claim," no provenance-on-the-assertion, no way to assert two
contradictory things as co-equal legal state. The lineage is legitimate
and should be cited, not claimed: paraconsistent belief
revision / LFI; bipolar/weighted argumentation (Dung; Toulmin/Pollock
supports-rebuts-undercuts); nanopublications (assertion
+ provenance + publication-info, with a 2024+ line using provenance to
detect contradictory claims); RDF-star / named graphs; Cyc
microtheories; and most precisely Standpoint Logic,
purpose-built so "multiple viewpoints [can] be integrated into the same
ontology, even when certain viewpoints may hold contradicting beliefs."
But the critique grep-counted the live store and found
zero
supports/rebuts/undercuts/hypothesis/contradiction/sameAs
predicates among 39.5M statements; the store is ~96% genealogy facts;
the user's own memory confirms "only 3 of ~80 Caroline-line kinship
triples have evidence_links, all in test contexts." The
differentiator is, today, an affordance, not an exercised
reality. Worse: the critique read
extraction.py and found that the six apertures run
independently and are deduped — there is no cross-lens
intersection / relationship-discovery step in the code at all.
The discovery engine the vision presupposes does not yet exist; the
substrate is being justified by a generator that hasn't been written.
This is the most important correction in the entire report.
What the critiques refuted or seriously qualified — integrate these or the vision dies:
Novel ≠ valuable (the ideation-execution gap). Si & Hashimoto (arXiv 2506.20803): LLM ideas rated more novel than experts (5.78 vs 4.91), but after 43 experts spent 100+ hours each executing them, LLM scores collapsed (overall −1.976, novelty −1.049) while human ideas barely moved. The apparent novelty advantage was an evaluation artifact. Volume of "connections no one drew" is the expected signature of spurious links, not of discovery (Calude & Longo on big-data spurious correlation). The verifier, not the generator, is where all value and all unsolved difficulty live.
The bitter lesson (Sutton). A hand-authored taxonomy of "philosophical / teleological / semiotic…" lenses is exactly the engineered human structure that scaled end-to-end learning has repeatedly out-classed. mat2vec found latent thermoelectrics with plain gradient-boosting on embeddings — no lenses, no symbolic store. A frontier LLM prompted "find a non-obvious connection between X and Y across economic and ecological framings" may extract the same links more cheaply. If a vanilla frontier model matches the lens engine, the lenses are dead weight. This is the crux baseline the pilot must beat.
The precision/noise deluge is made worse by many lenses, not better. Multiplying lenses multiplies the candidate cross-product; multiple-comparisons / false-discovery-rate math (Benjamini-Hochberg) guarantees the expected count of spurious-but-surprising "significant" hits approaches certainty. The novelty/precision tradeoff is empirically adverse (serendipity recommenders: novelty and accuracy directly negatively correlated). LLM self-validation does not certify truth ("even unanimous agreement among SOTA LLM critics does not guarantee scientific accuracy"). The Lean-4 overlay certifies shape/consistency, not empirical truth — it cannot supply the missing external signal.
So the honest verdict, integrated: the discovery idea is old and the generation half is increasingly commoditized; the defensible white space is narrow but real — an agentic, machine-generated, paraconsistent claim store with first-class evidence-anchoring and a formal verification overlay, operated at scale on a domain whose evidentiary structure demands contradiction-preservation. The genealogy / native-title domain (descendant-restricted records, irreducibly conflicting witnesses, legal stakes where every source-attestation must be preserved paraconsistently, and where the user's own standing rule is "no authority is ground truth") is the strongest possible first home, because there contradiction-holding is not a clever architecture choice — it is the actual requirement.
Here is a concrete design that takes every correction above seriously.
The semantic-decomposition tradition (Pustejovsky's qualia, NSM's primes, UDS's scalar layers) and faceted classification (PMEST) teach two non-negotiable design contracts: lenses should be as orthogonal as possible (value lives in synthesis across facets, not within one), and each lens should emit structured, typed, graded, falsifiable descriptors, not prose. Organize the spectrum into five families, anchored where possible to an existing formal lens so each is more than a prompt:
| Family | Lenses | Anchored to | Emits (typed descriptor) |
|---|---|---|---|
| Constitutive (what it is, is made of, is for, came from) | formal/kind, mereological (parts), telic (purpose), agentive (origin), material | Pustejovsky qualia; BFO continuant/occurrent | qualia 4-tuple; part-of lattice; function frame |
| Dynamical (how it changes, causes, persists) | temporal, causal, processual, modal/counterfactual | UDS event aspect (telicity/dynamicity); Jackendoff GO/CAUSE; bitemporal | event graph; causal DAG fragment; valid-time interval |
| Semiotic-linguistic (how it means, signs, frames) | linguistic, semiotic, rhetorical/pragmatic, narrative | FrameNet frames + roles; AMR; NSM primes | filled frame; predicate-argument graph; sign relation |
| Normative-evaluative (worth, ought, beauty) | ethical, aesthetic, teleological-value, legal | argumentation (Toulmin/Pollock); standpoint logic | value-claim + warrant; defeasible norm edge |
| Situated (its place in larger systems) | social, economic, ecological, political, phenomenological | structural holes / network position; D&S descriptions | role-in-network; flow/exchange edge; experiential frame |
Two design rules from the critiques. First — dynamic, fine-grained lenses beat a fixed generic list. Solo-Performance-Prompting (Wang et al., NAACL 2024) found dynamically-identified personas crush fixed ones (Codenames 79% vs 38%); a treaty rewards "legal/temporal/diplomatic-game-theory," a poem rewards "prosodic/semiotic/phenomenological." Let an agent select the lenses an entity actually rewards, drawn from this organized palette, rather than always running all of them. Second — subjective lenses (aesthetic, ethical, phenomenological) have no convergent ground truth; emit them as distributions of defensible readings tagged perspectival, never as facts. Per the bias-variance-diversity theory (Wood, Mu, Brown, JMLR 2023): diversity helps only among individually competent members, so gate each lens-output by a competence check and budget frontier-model capability — cognitive synergy emerges only at GPT-4-level (SPP).
Each selected lens runs as a specialized agent (CAMEL/AutoGen-style infrastructure) producing scalar, confidence-weighted, orthogonal descriptors on the entity's node — the UDS pattern, which is exactly what a paraconsistent, hypothesis-weighted substrate wants. To defeat representational collapse ("Representational Collapse in Multi-Agent LLM Committees": same model under N persona prompts gives cosine 0.888, effective rank 2.17/3 — illusory diversity), structurally vary models, temperature, and tools per lens, and measure semantic diversity (effective rank / pairwise distance), discounting near-duplicates. Heed YARN: use the LLM to abstract, then a structural step to validate — never trust the raw LLM mapping (Lewis-Mitchell counterfactual collapse). Heed Fodor: every descriptor must be evidence-anchored to a source byte so the question "real content or relabeling?" has an answer.
This is the step that does not yet exist in the code and must be built. It is not per-lens fact extraction then dedup. It is: given the descriptor-decorated nodes, generate edges between nodes that become visible only at lens seams. Concrete generators, each with prior-art backing:
hypothesis_only edge — SME's candidate-inference mechanism,
the literal "relationship no one drew."Never score a relationship with one number (Kotkov's decomposition). Carry four separate axes, computed cheaply before any expensive agent or Lean touch:
supports/rebuts argument-edge disagreement —
Chen/Ding/Evans).Critically: plausibility is a gate, never the ranking (LLMs over-produce plausible-invalid hypotheses). And multiple-comparisons control is not optional — make the false-discovery-rate budget an explicit, tunable parameter; an engine proposing millions of links manufactures apophenia by construction. Recycle rejected links as negatives (the HITL-KG pattern).
A funnel, narrowing, with cost rising at each rung — so cheap filters must do the heavy lifting:
supports/rebuts/undercuts edges
accumulate; bipolar argumentation frameworks adjudicate which claim wins
under a given lens at query time (identity-as-hypothesis).Paraconsistency lets the engine hold the firehose without collapsing — converting "decide value at generation time" (everyone else's binding constraint) into "accumulate speculative links and let evidence arrive asynchronously." Identity-as-hypothesis lets the same substrate discover relationships under different identity resolutions — a genuinely unexplored degree of freedom, and exactly where many false LBD links come from (spurious co-occurrence of ambiguously-grounded entities). Bitemporality lets a speculative edge be held as legal state now and retro-confirmed or rebutted by later-ingested evidence without rewriting history — BioDisco's temporal evaluation, native. And the contradiction frontier is the Chen/Ding/Evans priority map. No vector DB + reranker can do this without being rebuilt into a graph/claim store.
Combinatorial explosion. N entities × L lenses × pairwise relations is astronomically large, and almost all candidates are junk. Generation is trivially cheap (Weitzman/Arthur/Kauffman: the supply of combinations is effectively unbounded). The entire moat is the triage filter. SciAgents and Co-Scientist only tame the single-lens version (bounded random sampling, Elo pruning); many-lens intersection needs a fundamentally better candidate-generation theory, or it drowns before any triage.
Novel ≠ true ≠ valuable. The deepest correction. The ideation-execution gap (Si-Hashimoto) and TruthHypo's explicit novelty↔︎validity tradeoff (more creative ⇒ more hallucinated) show the most surprising link is disproportionately likely to be wrong. "No one thought to draw this" is the expected signature of a spurious link. Pareidolia by construction.
Evaluation has no ground truth. By definition, "a connection nobody made" has no label set. Time-slicing only validates connections that eventually got published (and its gold standard is mostly meaningless co-occurrence — Sebastian/Moreau). Value is subjective, experienced, and delayed — Wang/Veugelers/Stephan show the highest-value bridges look worthless short-term, so any greedy score discards them. There is no formal definition of "a discovery," no shared benchmark.
The bitter lesson. The single most dangerous critique. If a frontier LLM, prompted directly, surfaces the same valuable cross-domain links without the lens taxonomy and without the symbolic store, the elaborate pipeline is refuted. mat2vec is the existence proof that embeddings already capture latent structure cheaply. The lenses must demonstrably add discoveries the single-pass baseline misses (ablation: drop lenses, value drops) — or they are scaffolding a model already internalizes.
The profound-sounding-but-vacuous failure mode. LLMs fluently generate confident, beautiful, false cross-domain connections. Without per-lens competence gating, mandatory evidence-anchoring, FDR control, and an external verifier, the engine is a serendipity-shaped hallucination generator, and the paraconsistent store becomes "a landfill of unfalsifiable speculation."
Cyc is the cautionary ghost. A hand-built, microtheory-based, contradiction-tolerant, context-relative ontology that consumed 2000+ PhD-years and collapsed under curation load. The lens engine's bet is that agents — not knowledge engineers — fill the lenses cheaply, and that verification is largely automated. If either fails, donto becomes Cyc with extra steps.
The thread tying all six: generation is solved; discrimination is not. The founder's vision succeeds or fails entirely on the verifier.
A phased plan that builds the missing discovery step first, beats the bitter-lesson baseline before scaling, and proves itself with retrospective time-slicing.
Phase 0 — Make the moat real (close the schema-vs-reality
gap). Before any new generation, populate and exercise
the differentiating machinery on a bounded corpus: write
supports/rebuts/undercuts
argument edges, hypothesis_only flags, and
evidence_links for a few thousand relationship triples so
they are load-bearing, query-optimized state, not test-context
curiosities. Critique #3's grep returning zero must become a non-trivial
fraction. This is the precondition for everything.
Phase 1 — Build the cross-lens intersection step.
Extend the 6 apertures from independent per-entity extraction to a
genuine relationship-hypothesis generator (§4.3): start
with shared-qualia bridging + SME candidate-inference + bounded
random-path sampling over one bounded corpus. Emit edges as
hypothesis_only, evidence-anchored. This is the engine the
vision is actually about and which the code does not yet contain.
Phase 2 — The scorer + FDR gate. Implement the four-axis score (§4.4) as cheap pre-filters: novelty (substrate lookup), unexpectedness (single-lens primitive-prediction subtraction — the ablation), Uzzi z-score + Burt brokerage + discord over the quad graph, Bayesian surprise. Make the false-discovery budget an explicit knob. Recycle rejections as negatives.
Phase 3 — The verification ladder. Wire survivors into argumentation adjudication → Lean-4 shape certification (FCA implications are Lean-shaped) → human/empirical review of the thin top slice.
The falsifiable pilot. Pick a domain with cheap ground truth — the strongest is the genealogy / native-title corpus (bridging documents are checkable; contradiction-preservation is a hard requirement; "no authority is ground truth" is already the user's posture). Then run retrospective time-slicing, the field's only non-gameable metric:
hypothesis_only relationship edges.Report precision/value (edges surviving independent verification), not rated novelty. Pre-register the FDR budget and splits (the LBD reproducibility lesson). If the lens engine beats (a)–(c) on verified-valuable edges at a controlled false-discovery rate, the white space is real and executed. If it ties (a), the lenses are dead weight and the honest pivot is to the substrate-and-curation product — narrower, but still genuine white space.
Suppose Phase 0–3 succeed and the pilot beats all three baselines. What then?
The near horizon is modest and concrete: a standing discovery engine for one domain — a genealogy substrate that, as each new record arrives, re-ranks its entire dormant frontier of speculative kinship and identity hypotheses, surfacing the bridging documents a human researcher would have found only by holding every source at once. Because contradictions are held, not collapsed, it does the one thing the field's own ethics demand: it preserves every source-attestation paraconsistently and lets the lens (genealogical vs legal vs DNA) decide which reading wins at query time. That alone is a real product in a domain where the user already lives.
The middle horizon generalizes the operating model. Every prior
discovery system is one-shot: generate, rank, discard. donto's
bet turns one-shot generation into a compounding asset.
Each new lens, each new entity, each new piece of evidence
re-illuminates the entire held frontier — the adjacent-possible
explosion (Kauffman's TAP equation) working for you rather than
against you, because the speculation persists. A relationship proposed
in 2026 and held as hypothesis_only can be confirmed by
evidence ingested in 2031, retro-dated correctly by bitemporality,
certified by a Lean proof whose shape was unprovable until a 2029 lens
was added. This is the first architecture where machine
serendipity accumulates.
The far horizon — earned, not promised — is the one the founder is reaching for: a discovery engine for latent structure across all human knowledge. The science-of-value tradition proved (Wang/Veugelers/Stephan) there is a measurable surplus of value in cross-lens bridges that human, discipline-bounded evaluation leaves on the table — precisely because no human holds all the lenses. Foster/Evans proved scientists rationally avoid the risky bridges. A machine can bear the risk humans won't, hold the firehose humans can't, and re-curate forever where humans curate once. If — if — the verifier is good enough to keep the false-discovery rate bounded, the engine becomes a standing map of where the next interdisciplinary directions are most likely to pay off: a contradiction frontier that is, literally, Burt's structural holes drawn across all of recorded thought.
But earn the vision by saying the quiet part: the bottleneck is not imagination, and it never was. It is the verifier. The founder's deepest, most counterintuitive asset is not the many lenses — those are increasingly cheap, possibly free in a frontier model. It is the paraconsistent substrate that lets verification be lazy, asynchronous, and compounding instead of eager and one-shot. That is the genuinely new thing. Build the verifier, prove it with time-slicing, and the lenses will have somewhere worthy to point.
Grouped, annotated, with URLs. The starred (★) works are the ones to read first.
The direct ancestor — Literature-Based Discovery
The cognitive mechanism — bisociation, blending, creativity
Analogy — structure-mapping and its modern fusion
Faceted classification and Formal Concept Analysis — the emergence engine
Semantic decomposition — the lens-per-entity precedent
The science of value — which connections are worth it
Modern agentic discovery — the shipped state of the art
KG link prediction — the cheap first-pass generator
Multi-perspective reasoning — and its failure modes
Serendipity and surprise — how to score the rare gold
The contradiction-holding substrate — cite, don't reinvent
The essential warnings — read these with the vision, not after
Final word, in the founder's interest and not in flattery: the lens idea is old, the bisociation move is old, the discovery dream is forty years old, and the agentic generate-rank-evolve loop has already produced wet-lab-validated discoveries this year. None of that is the moat — and believing it is the moat is the fastest way to build Cyc again. The genuinely unclaimed ground is narrow, real, and yours for the taking: an agentic, machine-generated, paraconsistent, evidence-anchored claim store that holds the entire speculative frontier without collapsing and re-curates it forever — first proven on a domain (genealogy / native-title) whose evidentiary structure actually demands contradiction-preservation. Build the verifier, build the missing cross-lens intersection step, beat the bitter-lesson baseline with retrospective time-slicing, and you will have made the first place where machine serendipity compounds. That is worth doing.