Encevra Research Round 18

When Evidence Changes: Versioned Support and Bounded Belief Repair

9 October 2026 · Research paper · Version 1.0

Evidence status: Primary-source literature review, read-only inspection of the current Encevra archive, and executed exact diagnostics over constructed records. Twelve primary sources were inspected at the depths recorded in Section 17: ten full papers and two authoritative repository or abstract records. Thirty-six deterministic checks passed, including twenty-four temporal-query checks across all six arrival orders of a three-record example and twelve checks of baseline failure, conflict handling, dependency repair, alternative support, and source integrity. Twelve developer fixtures were generated. No new Encevra language-model benchmark, experimental model call, human annotation, production inference change, or deployment was performed. E18-B and H18-1 through H18-3 are proposals, not results.

Abstract

An agent can keep the newest statement and still answer the wrong question. A late correction may describe an earlier interval rather than the present; two current sources may conflict without either superseding the other; a changed premise may invalidate only part of a conclusion; and a later audit may ask what was known at the time rather than what is believed now. Round 17 studied a fixed source when the question changes. Round 18 holds that boundary apart and asks what representation and repair policy are needed when evidence itself changes.

We propose a Versioned Support and Dependency Ledger (VSDL). It is append-only, distinguishes valid time from recorded time, records source identity and update relations, permits an explicit disputed state, and links derived claims to their supporting premises. When an admitted update changes support, a repair process revisits only the affected dependency cone and preserves independently supported or historically recorded material. The proposal combines established ideas from truth maintenance, bitemporal databases, and data provenance. It does not claim to invent those ideas or to have demonstrated an Encevra advantage.

An exact worked example shows why neither destructive overwrite nor a deterministic newest-recorded rule is sufficient for current, historical, and as-known queries. The accompanying program verifies the example under every arrival order and checks bounded invalidation behavior. A proposed five-arm experiment separates the value of bitemporal representation from dependency-directed repair under one declared total-work budget. The design should be rejected or simplified if a strong overwrite summary or chronological evidence history performs as well at lower cost, if dependency construction causes excessive false invalidation, or if explicit conflict states mainly produce unhelpful abstention.

1. Numbering, archive inspection, and continuity

The live archive contains completed research through Round 17, so this paper is Round 18, correcting the automation request's older Round 11 designation. The archive was inspected at project commit af1d3181f90143ab3515d30543409ddd79ac452b. The Research overview, rounds index, publications surfaces, registry, full texts of Rounds 9 and 10, complete Round 16 manuscript, and complete Round 17 manuscript were reviewed. The archive identifies Round 17, When the Question Changes: Recoverable Evidence Under a Context Budget, as the latest completed entry. [I1-I7]

Round 10 already proposed evidence exchange, selective revision, belief dependencies, and retraction. Round 18 therefore cannot present dependency tracking in general as a new Encevra idea. Round 17 fixed the source packet and changed the later question; it explicitly reserved source updates and contradictory new evidence for the next round. Round 15's roadmap also assigned contradictory evidence and dependency changes to Round 18. The narrower contribution here is a testable separation of valid time, recorded time, unresolved conflict, and downstream repair under a controlled budget.

The public archive still describes the relevant end-to-end studies as proposed. This round does not bypass the Round 13 collection gate, alter the roadmap's release-risk phases, or modify Noetis. A careful ledger design is useful only if it survives measurement against simpler systems.

2. Research question, novelty rationale, and exclusions

Research question: At fixed total work and source access, does an append-only record that distinguishes valid time from recorded time and tracks claim dependencies improve current, historical, as-known, and downstream-consequence answers over strong overwrite, chronological-history, and newest-recorded baselines, without increasing false invalidation, unhelpful dispute states, or cost beyond prespecified margins?

This question was selected over broad confidence calibration and general verification allocation for three reasons. First, it is the explicit handoff from Round 17. Second, the literature contains a strong recent result for deterministic newest-serial assembly, which creates a useful baseline but also exposes a precise gap: a total order over versions does not represent late correction, historical validity, unresolved overlap, or dependency repair. [S6] Third, changing evidence makes selective revision measurable. A system can be evaluated on what changed, what should have changed, and what should have remained stable.

The primary setting is an evidence-backed agent answering repeated questions as source records arrive. The source packet can include policies, specifications, reports, or task facts. Some records supersede earlier ones prospectively, some correct an earlier interval retrospectively, some conflict without a settled ordering, and some are irrelevant to the active conclusion.

The round excludes:

1. General truth discovery or automatic proof that a source is trustworthy.
2. A universal source-authority ranking. Any authority relation in the experiment is supplied by the task or independently labeled.
3. Model-weight editing as the primary mechanism. Parametric approaches remain a competing position.
4. Unlimited retention, global causal modeling, or a complete ontology of every possible dependency.
5. Claims that hashes establish truth, that citations establish entailment, or that an append-only log is harmless by default.
6. Live legal, medical, financial, or operational deployment.
7. Literal brain copying. The biological comparison in Section 14 supplies one repair-scheduling hypothesis only.

The task is not simply “use the latest information.” It must first specify what “latest” means for the question. A current-state question, a historical question evaluated with present evidence, an audit of what was known at an earlier time, and an explanation of an earlier action can legitimately require different records.

3. Four temporal questions that a single latest value cannot answer

Bitemporal database work distinguishes valid time, when a fact is treated as holding in the modeled domain, from transaction time, when it is stored in the database. Clifford and Isakowitz formalize both dimensions and operations including timeslice and rollback. Their paper does not evaluate language models, but its distinction directly prevents one timestamp from doing two jobs. [S2]

For this round:

• valid_at is the domain time about which the question asks.
• recorded_at is when Encevra admitted a record into the inspected ledger.
• source_asserted_at is when the source published or asserted it, if known.
• decision_at is when an answer or action was produced.

The first two are indispensable to the experiment. The latter two are retained for provenance and audit. They must not be silently collapsed.

An overwrite store can retain the current answer but loses at least some historical or as-known distinctions. A newest-recorded rule can answer the synthetic total-order task used by MemoryAgentBench but can select a late correction as if it applied to the present. A chronological transcript retains bytes but leaves relation classification and consequence repair to the reader. VSDL is intended to make those policies explicit, not to guarantee their correctness.

4. Literature map and competing positions

The focused review was conducted on 9 October 2026. Primary sources were inspected directly, with the exact scope and limitations recorded in Section 17 and the machine-readable source ledger. The review is intentionally narrower than a survey of all temporal reasoning, databases, memory agents, or model editing.

The literature supports three narrower conclusions. First, temporal conflict, continual updating, and consequence propagation are demonstrated failure surfaces in the evaluated settings. Second, explicit structured assembly can help on a synthetic total-order task, but its advantage is not universal and much of its gain may come from separating extraction from generation rather than the freshness operator alone. Third, database and truth-maintenance methods provide mature representational ideas, but transferring them to fallible evidence extraction creates new error channels.

The evidence is mixed on architectural prescriptions. OAKS reports that agentic memory methods remain limited and can underperform naive RAG in aggregate, while doing better in some higher-update subsets. HoH shows harm from outdated retrieval, but its Wikipedia snapshots and mock search engine do not supply multi-source authority resolution. LOKA offers a parametric external-memory route, but its task and measurements differ from auditable bitemporal retrieval. Reddy and Challaram find no overall structured-assembly advantage on their small LongMemEval check. [S6, S9-S11]

The open Encevra question is therefore not whether updates exist. It is whether explicit bitemporal and dependency structure adds reliable value beyond good extraction, chronology, and deterministic ordering once total work is controlled.

5. Claim and evidence register

No paper in the source set tests the complete VSDL intervention. No inspected source establishes literal neurological equivalence, production safety, or a Noetis performance benefit.

6. Proposed mechanism: Versioned Support and Dependency Ledger

VSDL is an external record with two separable layers: an append-only evidence ledger and a claim-dependency view derived from it. “Append-only” means an update creates a new record and relation rather than silently replacing the bytes of the old record. It does not mean every record remains active, visible to every query, or retained forever without governance.

6.1 Evidence and claim record

The ledger distinguishes source records, claims extracted from them, derived conclusions, and actions or answers. A source can remain immutable while its active interpretation changes. A derived claim can become invalidated while the historical answer receipt remains intact. A dispute can remain unresolved instead of being collapsed to whichever candidate was ingested last.

6.2 Resolution and repair protocol

1. Append and bind. Admit the source version, compute its content identity, and preserve the original locator and bytes where policy permits. Never mutate a prior record in place.
2. Extract with explicit unknowns. Propose claim key, value, temporal scope, update relation, authority basis, and supporting span. If the relation or interval is unclear, record unknown rather than fabricating precision.
3. Validate the update event. Check source identity, task applicability, and whether the new record actually concerns the same claim. A duplicate or irrelevant update should not trigger semantic repair.
4. Select the query contract. Resolve valid_at and as_of_recorded. For a decision audit, use the earlier receipt rather than reconstructing an idealized history.
5. Assemble candidates. Retrieve all applicable versions within the common access budget. Apply explicit corrections or supersessions only over their declared scope. Preserve equal-authority incompatible candidates as disputed.
6. Compute the impact cone. When active support changes, traverse recorded dependencies from the changed claim. Do not invalidate unrelated branches. An OR-supported conclusion remains supported if another valid route survives.
7. Re-evaluate affected conclusions. Re-run the smallest needed derivations using the newly selected premises. Mark invalidated, changed, unchanged-after-review, or indeterminate. Do not treat reachability alone as proof that the answer must flip.
8. Answer with a receipt. Return the strongest supported answer, relevant temporal interpretation, unresolved conflict when material, and the selected evidence versions. Store the terminal response and cost.
9. Retain the previous state. Preserve old evidence, prior conclusions, and action receipts for rollback, audit, and error analysis, subject to privacy and retention policy.

This protocol does not require multiple model identities. One model can perform extraction and answer generation around deterministic storage and repair components. Conversely, multiple agents do not help if they share the same stale record or reinforce a wrong update classification.

7. Worked example: a late correction is not the current policy

Consider a fictional purchasing policy. Approval is required when an order amount is at least the threshold. The following records arrive:

On 20 January, the system saw P-v1 and evaluated an order of 55 as requiring approval. On 1 March, the policy office reports that the January threshold should have been 60. The correction changes the retrospective policy evaluation for the order: 55 is below 60. It does not change the current March threshold, which remains 75. It also does not erase the historical fact that the system required approval on 20 January based on the evidence it then had.

A destructive overwrite with 75 can answer only the first row. A timestamp-only newest-recorded resolver selects the March correction, 60, for the current question and is wrong. A raw transcript retains all three statements but does not by itself classify their intervals or repair the derived order evaluation. VSDL makes those operations explicit and inspectable.

Now add a second record from an equal-authority policy source asserting 70 from 1 February onward, with no settled correction or supersession relation. The fixture returns disputed for the current threshold. If the same assertion instead comes from a vendor report and the task explicitly ranks the policy office above vendors for this question, VSDL retains the vendor record but selects 75. The ranking is part of the task contract; VSDL does not infer that institutional ordering as truth.

The derived-claim graph contains January threshold -> order-55 evaluation -> January audit summary. A corrected January threshold reopens that branch. A separate shipping promise remains untouched. If a license conclusion has two valid alternative supports, removing one does not invalidate it. These cases test repair precision as well as repair recall.

8. Executed exact diagnostics

E18-A is the supplied standard-library Python program. It implements only the constructed semantics in Section 7. It is deliberately small enough to inspect and rerun without network access, credentials, or a model provider.

The program enumerates every arrival order of P-v1, P-v2, and P-v1-correction. For each of the six permutations, it checks four query contracts: current after the correction, January evaluated now, January as known before the correction, and February as known before the correction. All twenty-four permutation-query checks return the stipulated values. The ordering result depends on the explicit valid_at and recorded_at fields; it is not evidence that a language model will extract those fields correctly.

Twelve additional checks verify:

• the timestamp-only newest-recorded baseline returns 60 for the March current query while the bitemporal resolver returns 75;
• two histories with the same current value can require different January answers;
• a late correction does not rewrite the earlier as-known state;
• direct and transitive dependents are included in the impact cone;
• an unrelated branch is excluded;
• alternative support prevents false invalidation;
• an equal-authority overlap remains disputed;
• a declared lower-authority record does not override the policy record; and
• mismatched or unavailable source bytes are not treated as resolved evidence.

All thirty-six named checks passed. The program also emits twelve developer fixtures spanning clean supersession, late correction, unresolved overlap, and irrelevant update, each with current, historical, and as-known query forms.

The current-only impossibility is exact but narrow. Let history A have January 50 and current 75, and history B have January 60 and current 75. A representation containing only the current value is identical for A and B. No decoder using only that representation can answer the January query correctly in both worlds. This is an indistinguishability argument over two constructed histories, not a measured accuracy ceiling for a real memory system.

The diagnostics do not infer claim identity, source authority, valid intervals, corrections, contradictions, dependencies, or entailment. All were hand specified. They do not measure model calibration, retrieval, natural-language understanding, latency, or database scale. Passing E18-A establishes that the reference semantics are internally consistent across the tested cases, nothing more.

9. Proposed experiment E18-B

E18-B is a controlled development study, not a production trial. It remains gated on verification of the existing Round 13 collector and independent evidence labels. If the collector cannot preserve source versions, actual receiver exposure, provider attempts, and usage, repair that infrastructure before running another model comparison.

9.1 Units and update patterns

Create twelve independent source families. Each family should use a different domain and surface form while preserving a machine-checkable or independently adjudicated temporal structure. Examples can include fictional policies, device specifications, project states, schedules, and versioned task facts. Avoid live high-stakes claims. Source family, not individual query, is the generalization unit.

For each family, author four update patterns:

1. Clean supersession: a new value starts at a new valid time.
2. Late correction: a later-recorded source corrects an earlier interval without changing the current value.
3. Unresolved overlap: two applicable sources conflict without a valid precedence relation.
4. Irrelevant update: a source changes outside the active claim's dependency cone.

Each pattern receives two query forms: one direct temporal question and one downstream consequence question. Direct questions are balanced across current, historical-now, and as-known contracts. Consequence questions require at least one labeled dependency, with AND, OR, and irrelevant-branch cases represented across families.

9.2 Five matched arms

All arms receive the same source bytes, source metadata available in the task, update schedule, base model, and maximum access. O0 is explicitly instructed to preserve qualifications, dates, uncertainty, and sources. A0 may reason over the whole admitted chronology. T0 receives the same structured extraction interface as B0 where fields overlap. B1's extra dependency construction, storage, and repair work is charged to its budget.

The primary mechanistic contrast is B1 versus B0 on downstream-consequence correctness after a support change. This isolates the dependency layer from bitemporal selection. The main secondary contrast is B0 versus T0 on historical and as-known questions, which isolates two-time representation from deterministic newest-recorded assembly. B1 versus O0 is a system-level comparison and cannot by itself identify why a difference occurs.

Add separately accounted diagnostic controls:

• oracle claim identity and time fields, to estimate the ceiling after extraction errors are removed;
• oracle dependency edges, to isolate repair execution from dependency discovery;
• complete-source delivery, with any relaxed context cap disclosed;
• random affected-cone and global-rewrite controls at matched work;
• removed valid-time, removed recorded-time, removed conflict state, and removed alternative-support semantics;
• noisy dependency edges at frozen corruption rates; and
• a parametric-memory competitor only if it can answer the same historical and audit contracts or clearly reports that it cannot.

The oracle controls are not part of the candidate's headline performance. They localize failure stages.

9.3 Assignment count and controlled total work

The core development design has 12 source families × 4 update patterns × 2 query forms × 5 arms = 480 assigned episodes. This is a feasibility design, not a powered confirmatory sample. Repeated patterns and query forms within a family are correlated. Report family-level intervals or a hierarchical analysis rather than treating 480 rows as independent.

Use one provisional maximum of 24,000 reported model tokens per episode, including all extraction, update classification, summarization, lookup planning, answer generation, repair, retry, and failed-call input and output. The resulting planning ceiling is 11.52 million reported model tokens for the core assignments. It is a ceiling, not a target. Stop early when an arm completes, do not pad output, and report actual tokens, calls, latency, storage operations, and missing usage separately. A simpler arm may spend unused work on its best feasible reasoning or retrieval within the same cap.

For each episode, cap the initial source packet at 6,000 tokens, each update packet at 1,000 tokens, the retained handoff or ledger view supplied to the answerer at 1,024 tokens including metadata, and discretionary source reopening at two reads of at most 800 returned tokens each. These numbers are provisional development settings. Freeze them, the actual model versions, prompts, retry policy, and overflow behavior before held-out collection. Tokens are an imperfect proxy for compute and do not make architectures equal in hardware cost.

Preserve hashes for source packets, every versioned record, extracted fields, dependency graph, initial answer, update event, repaired state, actual answerer prompt, terminal answer, and usage record. A citation key or database row that was not exposed to the receiver is not evidence that the receiver used it.

10. Metrics, regression tests, and harmful changes

The primary endpoint is derived-consequence correctness after a support change, over all assigned B1 and B0 episodes. A terminal response counts as correct only if it satisfies the frozen temporal contract and is supported by evidence actually available to that arm. Score unresolved reference labels separately rather than forcing agreement.

Report the complete transition matrix: correct-to-correct, correct-to-wrong, wrong-to-correct, wrong-to-wrong, supported-to-disputed, and disputed-to-supported. Measure answer change and support change separately. A changed answer is not harmful when the underlying valid fact changed; an unchanged answer is not stable when it rests on stale support.

Use negative controls in every family: duplicate evidence, semantically irrelevant updates, later records with earlier valid intervals, and removal of one route where an alternative support survives. A repair system that improves changed cases but damages these controls has not demonstrated selective revision.

Calibration is secondary in this round. Collect confidence only after the temporal status is resolved, and score Brier loss or selective risk by query contract. Do not let well-calibrated uncertainty substitute for factual and support correctness.

11. Falsifiable hypotheses and decision rules

H18-1: Two times beat one order where their predictions differ. B0 improves the combined historical-now and as-known accuracy over T0 by at least five percentage points at the frozen budget, while current-query accuracy is not more than two points worse. The hypothesis is rejected as an architectural requirement if T0 falls within the frozen equivalence interval or wins on utility-adjusted cost. Analyze the late-correction subset separately because clean supersession may not distinguish the methods.

H18-2: Dependencies add value beyond bitemporal selection. B1 improves downstream-consequence correctness over B0 by at least five points, with no more than two points of added harmful revision or false invalidation on negative controls. It is rejected if direct re-answering over the selected evidence matches B1, if dependency errors erase the benefit, or if global rewrite performs as well at lower implementation cost.

H18-3: Explicit disputes reduce unsupported certainty without collapsing coverage. On unresolved-overlap episodes, B1 reduces unsupported commitment relative to O0 by at least five points while losing no more than two points of supported completion on answerable non-conflict episodes. It is rejected if the ledger mainly produces unnecessary disputes, if a strong prose baseline calibrates equally well, or if task-supplied authority resolves nearly all cases.

The margins are provisional planning values, not predicted effects or safety guarantees. Use development-family variability to set the confirmatory sample size and freeze one primary contrast, one budget point, the joint benefit-harm rule, and multiplicity treatment before collecting held-out families. Require confidence bounds to satisfy both benefit and damage margins; a favorable point estimate is insufficient.

An architectural null is useful. If O0 matches B1 within a narrow prespecified interval and costs less, prefer O0. If B0 matches B1, keep bitemporal representation but drop explicit dependency repair. If T0 matches B0 on every task that matters operationally, do not install unused temporal complexity.

12. Ablations and alternative explanations

Several explanations can imitate a VSDL benefit without validating its mechanism:

1. Extra structure rather than temporal semantics. T0 uses structured candidates too. B0 minus T0 localizes valid-time and recorded-time policy beyond formatting.
2. Extra context or retrieval. Match source access and charge actual receiver exposure. Re-run with exact decisive evidence delivered to all arms.
3. Oracle metadata leakage. Keep oracle time, relation, and dependency controls outside headline results. In the end-to-end arms, metadata must be produced from the same available sources.
4. Judge preference for schemas or caveats. Use executable labels when possible and blinded independent review otherwise. A judge should not see arm names or mechanism prose.
5. Task simplicity. Include direct, multi-hop, AND, OR, irrelevant-branch, and genuinely unresolved cases across independent families.
6. Parametric prior suppression. Use counterfactual fictional domains and separately report natural facts. TRACK and When Facts Change show that context-parametric conflict is a distinct issue. [S7, S8]
7. Deterministic code alone. Compare B1 with deterministic newest-recorded T0 and with B0 using the same extraction. If code policy explains the gain, the dependency layer has not earned credit.
8. More calls or slower reasoning. Report actual total work and an equal-work performance frontier, not accuracy at an unreported resource difference.
9. Source quality rather than repair. Cross the representation with clean, conflicting, and missing evidence. A ledger cannot recover an update that collection never observed.

Important ablations remove one field or operation at a time: valid time, recorded time, explicit correction relation, disputed state, dependency edges, OR groups, targeted repair, immutable decision receipts, and source hashes. Hash removal tests identity control only; it should not be interpreted as a truth ablation.

13. Expected failure modes and safeguards

Claim identity failure. Two sources may use different names for the same entity or the same phrase for different entities. A stable key can encode a wrong merge. Retain original text, confidence, and a reversible link; test alias and homonym cases.

Temporal-scope failure. “Effective next quarter,” retroactive amendments, local time zones, open intervals, and missing dates can be misread. Preserve source wording and mark unknown intervals. Require review before a high-consequence record becomes authoritative.

Relation failure. A new statement may supplement rather than supersede, or describe a different jurisdiction. unknown and coexists are necessary states. Newest-wins is not a safe default.

Authority failure. Provenance records who said something, not whether they are correct. Authority may vary by question and can itself be disputed. Never derive trust solely from frequency, recency, or document rank.

Incomplete dependency graph. Missing edges leave stale consequences. Spurious edges trigger collateral repair. Measure both repair recall and precision, and retain a global audit control for critical domains.

Cycles and alternative support. Derived claims can support one another or have multiple independent justifications. A naive traversal may loop or invalidate too much. Use strongly connected component handling, bounded traversal, and explicit AND/OR support groups in implementation studies.

History rewriting. A correction can tempt the system to rewrite an earlier action as if it had used the corrected fact. Store immutable decision receipts and distinguish retrospective evaluation from what was known then.

Overcautious conflict. An explicit disputed state can become a blanket refusal. Require the strongest supported conditional answer and score unhelpful disputes.

Update poisoning. A malicious or low-quality source can create repair churn. Validate access, applicability, and task-specified authority before activating an update. Append-only retention also increases privacy and poisoning exposure, so access control, deletion policy, and retention limits are mandatory.

Scale and cost growth. An append-only ledger and dependency graph can grow without bound. Compaction must preserve the temporal distinctions needed by declared query contracts, and deletion must leave a tombstone or auditable policy record where appropriate. This round does not solve long-term storage governance.

Explanation illusion. A neat dependency path can be wrong or incomplete. Treat it as a testable support record, not a faithful account of a model's internal causal process.

14. Brain-inspired hypothesis, narrowly stated

Sevenster, Beckers, and Kindt report in a human fear-conditioning experiment that prediction error was necessary for the tested pharmacological disruption associated with reconsolidation. Only the primary abstract was inspected for this round. The result concerns a specific associative fear-memory paradigm and does not establish a general algorithm for agent memory. [S12]

The useful analogy is a repair trigger: do not globally rewrite memory whenever a record is retrieved. Reopen a bounded support neighborhood when new evidence violates an expected value, temporal scope, or support relation. This yields a testable software comparison:

• no explicit change trigger, with ordinary answering;
• prediction-error-style trigger plus targeted dependency-cone repair; and
• global re-evaluation after every admitted update.

Hold total work constant and measure missed consequences, false invalidation, latency, and cost. If the targeted trigger misses more material consequences than it saves in collateral work, reject the analogy. A conventional database explanation may fully account for any benefit. That outcome would not weaken the engineering result; it would weaken only the brain-inspired framing.

15. Limitations

The archive inspection establishes continuity and numbering, not the correctness of prior conclusions. Earlier rounds contain proposed hypotheses that remain untested.

The source review is focused and not exhaustive. Ten full papers were inspected, but only selected methods, results, and limitations were read closely. Doyle's full paper was not retrieved in this run; the MIT repository record and detailed abstract were inspected. The Sevenster paper was represented by the PubMed record and abstract. Publication venues, tasks, backbones, and dates differ, so reported scores are not directly comparable.

The exact diagnostics are constructed. Their fields, authority ranks, dependencies, and answers are hand specified. The six arrival permutations demonstrate order invariance of the small resolver when semantic times are available. They do not demonstrate extraction accuracy, database completeness, natural-language reasoning, scalability, or robustness to adversarial sources.

VSDL borrows heavily from truth maintenance, temporal databases, and provenance systems. Its contribution here is an experimental decomposition for Encevra, not a new database theory. Production-grade bitemporal systems include semantics and operational concerns far beyond this paper's reference code.

The proposed 480-episode design is a development study. The planning ceiling is large but does not imply statistical power, equal FLOPs, or an obligation to consume the ceiling. Model APIs may omit usage, cache accounting, or reasoning details. All such gaps must remain explicit.

Authority labels can encode institutional or cultural assumptions. A benchmark with tidy total orders may reward the wrong social policy. Preserve unresolved cases and involve domain experts before any high-consequence use.

No new Encevra model study, human evaluation, or production change was run. The hypotheses should not be described as validated, and the exact reference outputs should not be used as benchmark performance claims.

16. Conclusion and next-round handoff

Round 18 separates four operations that are often compressed into “update memory”: record when a statement is valid, record when it became known, decide how conflicting records relate, and repair only the conclusions whose support changed. A newest-recorded rule is strong for a total-order benchmark but fails the constructed late-correction case. A current-only record cannot recover every historical distinction. A dependency graph can target repair, but its own extraction errors may create new harms.

VSDL is therefore a hypothesis-bearing interface, not a result. Its decisive tests are B0 versus T0 for bitemporal selection and B1 versus B0 for dependency-directed repair, with strong overwrite and chronological-history baselines, actual total-work accounting, and regression controls. The architecture earns its cost only if it improves derived consequences and temporal queries while preserving unrelated claims, alternative support, useful coverage, and immutable action history.

Next handoff: first validate the existing collector against the twelve development fixtures and confirm that source versions, admitted records, dependency edges, actual receiver exposure, terminal answers, and usage can be reconstructed. Then run the smallest development slice that distinguishes T0, B0, and B1 on late correction and irrelevant update. Do not run all 480 episodes until the transition records and independent labels are reliable. If collection fails, Round 19 should fix the measurement path. If the pilot is valid, Round 19 can test source-authority and conflict-calibration policies without changing the bitemporal core. A later round may study retention and compaction under long-lived ledgers.

17. Sources and inspection record

The links below were checked on 9 October 2026. “Inspected” states what was actually read in this round. External papers are not redistributed in the reproducibility package.

[S1] Jon Doyle. “A Truth Maintenance System.” MIT Artificial Intelligence Laboratory Memo AIM-521, issued 1 June 1979; journal version in Artificial Intelligence 12(3), 1979. MIT DSpace record and persistent handle. Inspected: repository metadata and complete detailed abstract, including recorded reasons, belief revision, dependency-directed backtracking, explanation summarization, and dialectical modules. The linked full PDF was not retrieved in this run. Supports: truth-maintenance precedent. Does not support: accuracy of dependencies extracted by language models or the proposed VSDL experiment.

[S2] James Clifford and Tomas Isakowitz. “On the Semantics of Transaction Time and Valid Time in Bitemporal Databases.” NYU Working Paper IS-92-42, 23 December 1992. Full paper. Inspected: full 40-page PDF, with close reading of temporal definitions, lifespan, timeslice, rollback, bitemporal combination, until-changed semantics, and conclusion. Supports: separate valid and transaction time and time-specific views. Does not support: LLM efficacy or the simplified resolver's completeness.

[S3] Todd J. Green, Grigoris Karvounarakis, Zachary G. Ives, and Val Tannen. “Update Exchange with Mappings and Provenance.” VLDB 2007, 23-28 September 2007, pp. 675-686. Full paper. Inspected: full 12-page PDF in the primary proceedings, especially system model, edit logs, provenance and trust filtering, alternative derivations, implementation experiments, assumptions, and future work. Supports: update provenance, trust-conditioned import, and derivation-aware incremental deletion. Does not support: transferring prototype scale or correctness to agent memory; the authors note restrictions and unresolved conflict/update-dependency work.

[S4] Zexuan Zhong, Zhengxuan Wu, Christopher D. Manning, Christopher Potts, and Danqi Chen. “MQuAKE: Assessing Knowledge Editing in Language Models via Multi-Hop Questions.” EMNLP 2023, 6-10 December 2023, pp. 15686-15702. ACL Anthology and PDF. Inspected: full 17-page PDF, with close reading of task construction, two- to four-hop evaluation, editing results, MeLLo, and limitations. Supports: direct edit recall does not guarantee propagation to multi-hop consequences in the tested systems. Does not support: temporal provenance or production behavior.

[S5] Yuanzhe Hu, Yu Wang, and Julian McAuley. “Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions.” ICLR 2026; arXiv:2507.05257v4, 28 June 2026. arXiv. Inspected: full 33-page conference paper, including framework, FactConsolidation construction, tested memory systems, context lengths, results, cost material, and limitations. Supports: incremental memory evaluation and a synthetic newest-serial conflict task. Does not support: late corrections, partial orders, source authority, or historical/as-known queries.

[S6] Vikas Reddy and Sumanth Reddy Challaram. “Reliable Post-Retrieval Assembly for Agent Memory: Separating Evidence Extraction from Policy Execution.” arXiv:2606.01435v2, 2 August 2026; accepted at the Lifelong Agent Workshop at COLM 2026. arXiv. Inspected: full 11-page v2 PDF, including deterministic highest-serial policy, controlled and whole-pipeline comparisons, MAB v3 results, LongMemEval check, and limitations. Supports: structured post-retrieval assembly as a serious baseline and exact execution of a declared total-order policy. The isolated policy executor adds 2.0 points on average and 0 at 262K in its targeted comparison, while the whole pipeline gains more; the 45-item LongMemEval comparison is null. Does not support: a universal “determinism wins” claim, bitemporal correction, or dependency repair.

[S7] Jonas Wallat, Wolfgang Nejdl, and Sandipan Sikdar. “When Facts Change: Temporal Knowledge Conflict Resolution in LLMs.” Findings of ACL 2026, July 2026, pp. 2154-2184. ACL Anthology and PDF. Inspected: full 31-page PDF, including WIKIRECENTCHANGES construction, temporal-conflict setup, prompted mutability reasoning, results, and limitations. Supports: models in the study often distinguish changed from stable facts in reasoning without reliably carrying the distinction into final predictions. Does not support: multi-source authority, bitemporal storage, or a VSDL effect.

[S8] Yiyang Feng, Zeming Chen, Haotian Wu, Jiawei Zhou, and Antoine Bosselut. “Tracking the Limits of Knowledge Propagation: How LLMs Fail at Multi-Step Reasoning with Conflicting Knowledge.” EACL 2026, March 2026, pp. 5813-5847. ACL Anthology and PDF. Inspected: full 35-page PDF, including WIKI, CODE, and MATH tasks, direct and propagation errors, aggregation-scope experiments, failure analysis, and limitations. Supports: updated facts can be integrated inconsistently and still fail to propagate through multi-step reasoning. Does not support: streaming updates, source authority, or a tested repair graph.

[S9] Jie Ouyang, Tingyue Pan, Mingyue Cheng, Ruiran Yan, Yucong Luo, Jiaying Lin, and Qi Liu. “HoH: A Dynamic Benchmark for Evaluating the Impact of Outdated Information on Retrieval-Augmented Generation.” ACL 2025, 27 July-1 August 2025, pp. 6036-6063. ACL Anthology and PDF. Inspected: full 28-page PDF, with close reading of dataset construction, retrieval and generation experiments, timeliness-awareness analysis, conclusion, and limitations. Supports: outdated passages can reduce accuracy and increase harmful outputs even alongside current material in the tested RAG setups. Does not support: real-time coverage, multi-source synthesis, or the VSDL mechanism; the paper notes monthly snapshot and same-article-source limitations.

[S10] Jiyeon Kim, Hyunji Lee, Dylan Zhou, Sue Hyun Park, Seunghyun Yoon, Trung Bui, Franck Dernoncourt, Sungmin Cha, and Minjoon Seo. “Can Large Language Models Keep Up? Benchmarking Online Adaptation to Continual Knowledge Streams.” ACL 2026, 2-7 July 2026, pp. 42240-42272. ACL Anthology and PDF. Inspected: full 33-page PDF, with close reading of OAKS construction, fourteen-model evaluation, RAG and agent-memory comparisons, transition analysis, acquisition latency, distraction, phase misses, and limitations. Supports: frequent updates produce over-update, under-update, distraction, and later-interval degradation in the tested streams. Does not support: late correction versus current validity, multi-source authority, or decision-audit semantics.

[S11] Binchi Zhang, Zhengzhang Chen, Zaiyi Zheng, Jundong Li, and Haifeng Chen. “LOKA: Conflict-Aware LLM Knowledge Update with Adaptive Knowledge Memory.” ACL 2026, July 2026, pp. 16689-16715. ACL Anthology and PDF. Inspected: full 27-page v2 PDF, including motivation, conflict-aware memory units, mapping and routing, experiments, ablations, conclusion, and limitations. Supports: modular parametric memory as a serious alternative for joint learning and unlearning. Does not support: auditable historical/as-known retrieval; threshold selection and lifelong scaling remain open in the paper.

[S12] Dieuwke Sevenster, Tom Beckers, and Merel Kindt. “Prediction Error Governs Pharmacologically Induced Amnesia for Learned Fear.” Science 339(6121), 15 February 2013, pp. 830-833. PubMed record and DOI. Inspected: authoritative bibliographic record and full abstract only. Supports: in the reported human fear-conditioning paradigm, prediction error was a boundary condition for the tested reconsolidation effect. Does not support: a general memory theory, dependency graphs, agent architectures, or literal brain copying.

Internal continuity. [I1] Round 9, 30 September 2026, full procedural-learning paper. [I2] Round 10, 1 October 2026, full evidence-exchange and selective-revision paper including H10-B. [I3] Round 13, 4 October 2026, and Round 14, 5 October 2026, collection and evaluation commitments. [I4] Round 15 public summary, 6 October 2026, plus the previously inspected private roadmap. [I5] Round 16, 7 October 2026, complete missing-condition paper. [I6] Round 17, 8 October 2026, complete fixed-source, changed-question paper. [I7] Pinned project snapshot identified in Section 1: current overview, rounds index, publications and registry surfaces, full Rounds 9, 10, 16, and 17, and project guidance. Canonical links identify intended archive pages; repository source was inspected because preview edits may precede deployment. Private notebooks and implementation files are not reproduced.

Reproducibility note

The companion package contains this manuscript, a rendered PDF, the standard-library diagnostic program, exact results, twelve developer fixtures, a machine-readable source ledger, the proposed E18-B protocol, delivery status, and a SHA-256 manifest. Running python3 diagnostics.py rewrites only the diagnostic results and fixtures. The expected summary is:

{"checks_passed": 36, "arrival_permutations": 6, "permutation_query_checks": 24, "semantic_checks": 12, "fixtures": 12, "experimental_model_calls": 0}

The fixtures and reference outputs are development aids, not hidden benchmark data or measured model results. Keep their labels out of future candidate prompts.