Publications

A deterministic repository-readiness audit found 34 controls: 8 present, 3 partial, 23 absent; 26 critical controls partial or absent; fail-closed verdict BLOCKED. RISE and E30-B are proposed. No live intervention trial, language-model benchmark, production experiment, or Noetis answer-quality benchmark was run.

• Download reproducibility bundle (ZIP)

Active content probes can identify observable response behavior but cannot certify root provenance when provenance twins are interventionally equivalent. The proposed Interventional Lineage Test requires action-fidelity receipts and root-addressing interventions or an unresolved-equivalence outcome. E29-A is exact only for its declared finite deterministic diagnostic; E29-B is proposed and has no results. No live benchmark ran.

• Download PDF
• Download reproducibility bundle (ZIP)

Model-set intervals were conditionally valid in-model, but passive adequacy signals failed on the hidden-common-origin alias. TSAR did not universally dominate a scalar. E28-A is an exact constructed trace replay of 9,744 frozen histories and 29,232 weighted rows, not a production, language-model, or Encevra answer benchmark. Round 29 remains a proposed study of active dependency interventions that can reveal omitted dependencies.

• Download PDF
• Download reproducibility bundle (ZIP)

A model-set unanimity gate separates verdict safety from scheduling efficiency under fixed verification budgets, while an out-of-model common-origin stress shows that extra evidence can increase false confidence. E27-A is exact only for the declared scenario set, deterministic observation rules, authored cost tables, priors, and policies: 58 worlds per dependency variant, a 116-scenario planning set, 9,744 policy-world runs, and 504 aggregate evaluation cells. Four model-set gate policies produced no wrong categorical verdict in 3,712 in-model runs; in the out-of-model hidden_common_origin stress the exact policy's false-clean rate reached 33.68 percent at budget 9. A myopic adaptive heuristic sometimes underperformed a fixed target-first schedule. No production trial, no language-model reasoning benchmark, and no live Encevra benchmark was run. E27-B and H27.1 to H27.7 are proposed. Architecture impact: None. Production impact: None.

• Download PDF
• Download reproducibility bundle (ZIP)

E26-A is exact only for its declared finite model: 2, 16, 58, and 128 worlds at fault budgets 0 through 3, respectively. Seven failure roots and eight observable types were enumerated, with eight executable tests passed. Root independence is an assumption. The proposed Observation Sufficiency Certificate (OSC) is research-only, not installed or production-validated. No production incident was sampled, no language model was benchmarked, and no new Encevra answer benchmark was run. E26-B and H26.1 to H26.7 are proposed. Architecture impact: None. Production impact: None. The automation brief named Round 11, but the inspected archive ended at Round 25, making Round 26 sequential.

• Download PDF
• Download reproducibility bundle (ZIP)

Silence-Aware Coverage Protocol for missing change evidence. D25-A is an authored-fixture conformance diagnostic, not production evidence. On 80 authored traces (60 positive gaps, 20 negative controls), SACP alerted on 55 of 60 gaps and 0 of 20 controls: TP=55/60, FP=0/20, trace recall 0.9167, claim-weighted recall 0.8944. The five intentional misses were two selective non-emission cases with no independent external commitment and three verifier common-mode failures with no independent challenge path. No production incidents were sampled, no independent reviewer labelled the traces, no language model or provider API was called, and no new Encevra answer benchmark was run. D25-A: Executed (bounded diagnostic). E25-B and H25-1 to H25-10: Proposed. Architecture impact: None. Production impact: None.

• Download PDF
• Download reproducibility bundle (ZIP)
• Download full source ledger (JSON)

Authority lineage for claim revalidation: two analyses of one repository are not two independent watchers. D24-A is a deterministic authored-fixture conformance diagnostic, not an estimate of production precision, recall, reliability, prevalence, or cost. No Encevra answer benchmark, provider experiment, production observer, or deployment was run. D24-A: Executed (bounded diagnostic). E24-B and H24-1 to H24-8: Proposed. Architecture impact: None. Production impact: None.

• Download PDF
• Download reproducibility bundle (ZIP)

A shadow claim-revalidation gate tested prospectively on one real commit interval after the Round 22 manifest was frozen. D23-A: Executed (bounded shadow diagnostic). E23-B and H23-1 to H23-8: Proposed. Architecture impact: None. Production impact: None. Result boundary: one post-freeze real commit interval; same-process post-hoc impact labels; eight authored conformance fixtures; zero model, provider, production-database, and production-write calls; no product reliability, answer-quality, security, latency, or cost claim; no production CTAG or ABMG implementation.

• Download PDF
• Download reproducibility bundle (ZIP)

The Change-Triggered Assurance Graph (CTAG): bind each current claim to the code, configuration, external contracts, tests and evidence it depends on, move it to stale_needs_revalidation on relevant change, keep historical results frozen, and fail closed on unknown paths. Proposed, not implemented in production. D22-A is deterministic routing against a manually audited actual diff plus authored mutation fixtures. It made zero model/provider/production-database calls. It is not a model-quality benchmark, automatic dependency-discovery result, production integration test, security audit, product reliability claim, production change, or deployment.

• Download PDF
• Download reproducibility package (ZIP)

The Attempt-Bound Measurement Gate (ABMG): capture each provider attempt at the boundary, require provider-specific stream finality, persist completeness atomically, and reserve the audit-qualified label for complete traces. Executed deterministic development diagnostic only: 34 of 34 named checks; static audit found 10 of 31 requirements present, 9 partial, 12 absent; zero model, provider or production-database calls. E21-B and six hypotheses are proposed and untested. No model benchmark, production change, or deployment.

Minimal-Disclosure Invocation Receipts (MDIR) that bind application-observable evidence delivery for authorized audit while keeping raw prompts and source text out of routine telemetry. Executed deterministic development harness only: 76 of 76 named checks across 96 constructed episodes, eight synthetic task families, twelve corruption families, and 6,912 mutated episode-arm evaluations; 480 valid reveals verified and 480 altered reveals rejected; zero model calls. E20-B and four hypotheses are proposed and untested. No new model benchmark, production change, or deployment.

A measurement contract for reconstructing which source version reached each agent and how the final answer is bound to it. Executed deterministic diagnostics only: 44 named checks across 15 constructed fixtures; a static audit found 4 of 18 collector requirements present, 3 partial, 11 absent. E19-B and three hypotheses are proposed and untested. No new model benchmark, model call, production change, or deployment.

A versioned support and dependency ledger that separates valid time from recorded time, keeps disputes explicit, and repairs only affected conclusions. Executed deterministic constructed diagnostic only: 36 passing checks, twelve developer fixtures. The model study and three hypotheses are proposed and untested. No language-model benchmark, model call, human annotation, production change, or deployment.

A bounded handoff with declared scope and source locators, checked when the question changes. Executed finite constructed diagnostics only: 28 named checks, 15 partitions, 60 recompressions, twelve developer fixtures. The model study is proposed, not run or preregistered, and three hypotheses are untested. No new model benchmark, Noetis performance gain, or deployment.

A bounded contrastive sufficiency probe and separate source-coverage probe. Executed finite-world diagnostics only: 6,912 Boolean pairs, 37 named checks, twelve developer fixtures. The model study and three hypotheses are proposed and untested with language models. No new model benchmark, deployment, or calibrated release guarantee.

Distinguishes messages that change an answer from messages that improve it, and how separate evidence can be valuable only when combined. Exact mathematical diagnostics and proposed experiments; not new Noetis benchmark results.

A measurement protocol for revision gains when evaluation labels are missing, ambiguous, or differentially biased: joint label sets, bounds on repair and damage, and a protected random audit. Primary-source synthesis and 30 exact offline diagnostic checks; no new model benchmark, inference, or expert annotation.

A reproducible audit of saved Noetis records, an offline trace validator, and a pilot protocol for measuring repair, damage, and matched-cost advantage. Audit of existing artifacts and offline checks; no new model inference runs or calibrated release guarantee.

Calibrate the final revised answer under a defined selective-risk target; test whether terminal evidence and revision features improve useful release coverage at identical generation compute. Primary-source synthesis, proposed experiment, and executed exact synthetic diagnostics; no new language-model benchmark.

Predicting repair and damage with claim-specific checks, short bundles, and controlled end-to-end cost. No new Encevra or LLM benchmark; only exact illustrative diagnostic calculations were run.

A critical review and proposed experimental program: what communication and revision protocol makes a correction trustworthy enough to reuse? Introduces hypothesis H10, six comparison arms, and a falsification plan.

A testable question for the lab: whether systems can acquire useful procedures from verified experience, reuse them on unfamiliar problems, and revise them when conditions change, with two experiments specified for the next 90 days.

A failure mode in belief revision: models can acknowledge counterevidence in their response while retaining an incompatible conclusion. We examine what structured adversarial review can and cannot recover.

A research note on eligibility traces, delayed credit assignment and consolidation: how later feedback might modify a temporary trace without destabilising previously learned behaviour.

An examination of hippocampal replay as a hypothesis for recombining evidence into new inferences, with careful limits on its analogy to multi-agent analysis.

The cognitive-science hypotheses informing Encevra's approach to specialist coordination, evidence integration and adaptive reasoning.