Papers
donto's own original research — novel concepts and results that come out of building a bitemporal, paraconsistent, evidence-anchored knowledge substrate. Not literature surveys (those are reports); these are the contributions only a contradiction-preserving substrate can make. Living drafts, written from the live system and revised as the work progresses. Each was drafted then adversarially fact-checked against the running donto-pg, with claims held to a four-tier candor scale (proven / measured-but-noisy / conjectured / speculative).
Last updated: 2026-06-13.
Program
The research agenda lays out ten original problems that become askable once a knowledge base holds contradiction instead of collapsing it. They share one engine — hold everything, align late, re-rank by reality — and the papers below are the worked-out contributions.
| paper | contributes |
|---|---|
| The donto research agenda | The full program: the ten problems, why donto is uniquely positioned for each, the experiment that would establish it, and how they compound into one engine |
Theory
| paper | contributes |
|---|---|
| The Bitemporal Sheaf | Cellular-sheaf cohomology over both valid-time and transaction-time — combining bitemporal databases and sheaf theory for the first time. H¹(valid,tx) distinguishes the world was contested from we were wrong then corrected; "re-rank by reality" becomes a spectral flow that drains transaction-H¹ while preserving genuine historical disagreement |
| Cycle contradictions | The conflicts pairwise donto_argument edges structurally cannot see: loops of individually-fine claims that can't be globally reconciled (the Kitty daughter-of vs spouse-of case). Sheaf holonomy detects + localizes them — with the d≥2 sign-carrying-maps correctness fact |
| Identity as a restriction map | Non-destructive entity resolution: never merge — identity is a learned restriction map + a glue/no-glue verdict, a reversible query-time decision. On contested data this dominates merge-based ER, which a wrong early merge ruins irreversibly |
| Standing as a dynamical system | Epistemic standing ⟨maturity, corroboration, contradiction-pressure, recency⟩ as a vector field steered by a feedback loop with reality (the STEER loop). Fixed points = settled facts, limit cycles = perennial disputes; standing-weighting may be the necessary condition that makes the bitemporal-sheaf convergence hold |
| Alignment debt | An information-economics of emit-free knowledge: when generation cost → 0, normalize-at-write becomes strictly worse than align-at-read. Defines alignment debt and the crossover Δ_norm > r·δ·φ; grounded in donto's 1.16M live predicates + the query-time fold engine (+8.7% recall, measured) |
Empirical & method
| paper | contributes |
|---|---|
| Answer-Shaping | The scalpel/hammer law from LoCoMo + LongMemEval on a weak reader: pre-computing the answer-relevant fact lifts focused-fact questions (preference 0.70→0.767) and degrades synthesis (0.70→0.63); aggregates-as-context = 86% of episodic accuracy at 3.8× fewer tokens |
| Decoupled anchoring | Separating what (extraction) from where (citation) into two always-on stages: run the cheapest extractor without losing faithfulness, and the citer becomes a hallucination filter. Anchor-precision over coverage — bogus spans driven to zero by construction (refuse, never fabricate) |
| Contested retrieval | Retrieval over a paraconsistent store that returns disagreement as a first-class result — "X (contested: 2 sources disagree, see Y)" — with a contestedness score from the sheaf H¹, instead of silently picking a winner |
| Generative-abundance metrics | Predicate proliferation as signal, not noise: a metric suite (proliferation, fold-rate, alignment-debt repayment) treating the freely-minted predicate space as the health signature of an emit-free KB; aggressive ingest-time typing destroys recoverable signal that query-time folding preserves |
| The verifiable epistemic trail | Agent memory as a bitemporal, evidence-anchored claim trail that is independently verifiable and replayable — reframing the agent-memory benchmark from accuracy to accountability ("what did the agent know, and when," as a query) |
See also: reports (literature grounding + benchmark studies) · benchmarks · open questions.