Research Dossier

This dossier is the synthesized output of a multi-source, fact-checked research pass on dent8’s design space (agent-memory integrity, belief revision, formal verification, provenance/tamper-evidence, and academic positioning). It is the entry point; the depth lives in the topical docs it links. Facts here were produced by parallel web-search agents, then adversarially fact-checked (16 load-bearing facts: 14 supported, 1 refuted, 1 qualified). Corrections from that pass are folded into the topical docs and noted below.

Executive summary

1. dent8 is principled, not novel-per-primitive. Every individual mechanism — bitemporal validity, supersession-not-deletion, contradiction edges, provenance, deterministic replay, hash-chained logs — is prior art (Zep/Graphiti, PROV, SQL:2011, event sourcing, Certificate Transparency). These foundations predate and are independent of Zep; the overlap with Zep is convergent evolution on shared sources, not derivation. The defensible contribution is the combination as substrate plus typed authority-weighted supersession. → related-work.md

2. Belief revision is the right formal identity, and it’s a clean fit. dent8 is a belief base (Hansson), not an AGM belief set; it should deliberately violate Recovery; its contested state is genuine paraconsistency (inconsistency ≠ triviality); authority-vs-confidence is entrenchment vs probability; edges are TMS justifications. This is the “fresh and inspiring” lens and dent8’s strongest defensible framing. → belief-revision.md

3. A layered formal-verification stack is achievable today. proptest/bolero → Kani (bounded) → Stateright (concurrency) → optional Creusot/Verus (the fold). Frame honestly as “property-tested + bounded-model-checked + concurrency-model- checked,” never blanket “formally verified.” → formal-verification.md

4. The honest gap is narrowing. Authority arbitration is now implemented in the core fold (the headline differentiator vs Graphiti’s recency-only resolution), with an exhaustive non-resurrection proof. The freshness evaluator and serde + canonicalization + hash chain are also built and tested in the library. The write-path firewall (enforced at EventStore::append), the DB-verified Postgres adapter, the CLI/MCP runtime, and the evidence-edge retraction taint (ADR 0010, dent8 derive/verify + the poisoned_source_retraction eval) are built — the differentiator is now a running, demonstrated system (see STATUS.md). The analytical/export lane is also built (dent8 export → Parquet for DuckDB), and the file-based scenario-family golden corpus is seeded (evals/); what remains on the eval side is cargo-fuzz. → roadmap.md, threat-model.md

5. Publish in two stages. A model + belief-revision-semantics workshop paper is factable from the core semantics and runnable slice; the systems/security paper should wait for stronger operational evidence: fuzzing/model checking, operated witness deployment, and clearer production identity operations. → paper/outline.md

6. Defensible novelty is compositional, and an adversarial pass killed every single-primitive fact. The surviving directions route through authority arbitration — implemented. All three top directions now have a built+tested v0: rank 1 (verified non-resurrection, exhaustive + Kani harness), rank 2 (policy-counterfactual replay: EpistemicPolicy + replay_*_with_policy + diff_states), and rank 3 (earned entrenchment: authority-weighted corroboration + unearned_supersessions audit + the challenge-survival half — ADR 0015: rejected challenges recorded as fact.challenge_rejected feeding FactState.survived_challenges, on by default, plus the opt-in earned-supersession gate DENT8_ENTRENCHMENT_GATE). Rank 3 also now feeds survived challenges into arbitration (ADR 0017: the gate weighs earned entrenchment = corroboration + survived challenges). → novelty.md

What’s useful for the project (concrete adoptions)

Adoption	Source/standard	Where
Belief-base + Recovery-non-postulate framing	Hansson; AGM	ADR 0005
Paraconsistent contradiction tolerance (“contested”)	LFI / paraconsistency	belief-revision.md
Authority-as-entrenchment arbitration	AGM entrenchment	roadmap.md / ADR 0007
LFI hard-alarm tier on canonical facts	Logics of Formal Inconsistency	belief-revision.md §Adopt-3
Sorted-key serde_json canonical form (not JCS)	project invariant; RFC 8785 kept as contrast	ADR 0004
RFC 6962 domain-separated leaf/node hashing	Certificate Transparency	storage.md
W3C PROV-DM export mapping	W3C Recommendation	related-work.md
Layered verification (proptest/Kani/Stateright)	AWS systems-correctness	ADR 0006
MINJA/PoisonedRAG poisoning fixtures	attack literature	threat-model.md

Fact-check ledger (corrections applied)

• Refuted: “Prusti (uniquely) verifies unsafe code.” → Deductive verifiers (Creusot/Verus/Prusti/Aeneas) target safe Rust with unbounded functional correctness at high spec cost; the unsafe-coverage framing was wrong. Fixed in formal-verification.md.
• Qualified: RFC 8785 is an Informational RFC (Independent Submission), not Standards Track; property names sort by UTF-16 code units. dent8 deliberately ships a sorted-key serde_json form instead, and documents that it is not JCS in storage.md / ADR 0004.
• Critic corrections folded in: confidence float hazard was overstated (Confidence is u16; only FactValue::Json was at risk, now canonicalized via the CanonicalJson newtype); the temporal-validity matrix cell downgraded ✓→◐ (at the time: no valid_to, freshness-only; the freshness evaluator, and in v0.2.0 valid_to intervals plus --as-of/--valid-at time-travel reads per ADR 0016, have since shipped); TTL/authority arbitration flagged as design-only everywhere they are stated; “AWS originated the P language” reworded (P: Microsoft/UC Berkeley); LOC corrected to ~470 non-test (a figure from that early snapshot — historical; the codebase is now ~33k src LOC).

Novelty risks (kept deliberately visible)

These are the reviewer objections the project must pre-empt, not hide:

1. Authority-weighted supersession is THE differentiator vs Graphiti’s recency-only arbitration — now implemented, enforced at the write boundary, and runnable through CLI/MCP over file, SQLite, and DB-verified Postgres stores (no longer “zero code”). The remaining honesty caveat is product hardening: survived challenges are recorded and consulted by the opt-in earned-supersession gate (ADR 0017: earned entrenchment = corroboration + survived challenges), and production deployments still need operated witness and source-key workflows.
2. Three of four “combination” ingredients are individually prior art (replay = event sourcing; hash-chain = transparency logs; bitemporal/provenance = SQL:2011/ PROV). Only typed authority-as-entrenchment arbitration is uncommon in this space.
3. Belief-base framing is vocabulary, not a novel mechanism — bitemporal DBs already give “history matters, retract doesn’t resurrect.” Fact it as principled grounding, not as a contribution.
4. On the temporal axis dent8 now has closed valid-time intervals (ADR 0016): valid_to via --valid-from/--valid-to, an elapsed valid_to bounds read-time freshness like a TTL (expires_at is the earliest bound), plus --as-of/--valid-at time-travel reads — so it is no longer behind Zep on the missing-valid_to axis. Frame the remaining difference precisely: dent8 treats valid_to as an asserted read-time freshness/validity bound, not a running lifecycle edge-invalidation — so state the exact bitemporal gap rather than asserting either a missing interval or full parity.
5. Deterministic replay alone is not unique — Zep/Graphiti reconstruct from episodes too. The precise differentiator is the typed, hash-verified, single-source-of-truth log.
6. Hash-chain tamper-evidence is table-stakes, not a contribution (RFC 6962, blockchains). It is built (canonicalization + SHA-256 chain, DB-verified); the resistance upgrade (an operated external witness) is the part still maturing.

Method note

Research harness: scope → 6 parallel search angles → URL-grounded fetch → 3-vote-style adversarial verification (refute-default) → parallel section synthesis → completeness/ novelty critic. ~28 agents, ~1.3M tokens. The verbatim section drafts, verdicts, and critic output are preserved outside the repo; this dossier and the topical docs are the curated, corrected result.