donto — a contradiction-preserving substrate for knowledge in the age of generative abundance

2026-06-17. A full LaTeX treatise (6 pages) accounting for a two-day autonomous research run: what donto is, what was actually built and measured, and whether it can be the substrate for the future of AI and knowledge. Honest throughout — it corrects its own headline where the measurement disagreed with the story.

📄 Read the full PDF treatise →

This is the consolidated, citable version of the work detailed in the companion web report, Claim + evidence-span recall. The PDF covers:

• What donto is — a bitemporal, paraconsistent, evidence-first claim substrate built for the era where generating typed knowledge is free and the hard problem is holding an unbounded, contradictory firehose without collapsing it.
• The instrument — why benchmarks (LoCoMo, LongMemEval) are used as steering signal, not a leaderboard; the reader-ladder finding (retrieval is complete, the reader is the cap); and the definitive full 1,539-question LoCoMo run (53.3%, reader-bound).
• What was built — the full LongMemEval claim layer (~115K evidence-anchored claims); the claim + evidence-span recall unit, designed, validated, and deployed into the live /recall API; the extraction-infrastructure fixes; and the operational saves (a 19-hour runaway query killed, disk 99%→39%).
• The correction — a cross-reader check that refuted the tempting headline: claims+spans beat the dump only with a weak reader; with a strong reader on a dump that fits, the dump wins. The robust wins are token-efficiency (3.5–4.8×) and claims+spans > bare claims. The accuracy win requires the regime where the context cannot fit — an honest tie in the budget-constrained proxy, with the decisive _m test still open.
• Can donto be the future of AI and knowledge? — separated honestly into what is warranted by what was built (a non-collapsing, evidence-first, token-efficient memory substrate, deployed) and the larger bet (an epistemic operating system for contested reality), with a frank account of the distance between the two.

Every number in the document is measured, not asserted.