Encevra Research Round 17

When the Question Changes: Recoverable Evidence Under a Context Budget

8 October 2026 · Research paper · Version 1.0

Evidence status: Primary-source literature review, a read-only inspection of the current archive and one inference implementation file, and executed finite diagnostics. Twenty-eight checks passed, covering all 15 partitions of a four-world example, 60 deterministic recompressions, and a small source-resolution protocol. Twelve developer fixtures were generated. No new Encevra language-model benchmark, experimental model inference, human annotation, or production change was performed. The proposed model study has not been run or preregistered. Nothing in this paper establishes a measured Noetis advantage.

Abstract

A compressed handoff can be sufficient for the question that produced it and insufficient for the next question. Encevra's previous round examined how to notice a missing decision condition. This round asks what happens when the condition was available earlier but was omitted because a different question appeared relevant. We study a bounded representation containing supported facts, their dependencies, a declared scope of use, and references to preserved source material. At a question change, a receiver checks what the representation supports and can reopen evidence within a common budget. This is not a new invention of source retrieval: ReadAgent, WhenLoss, and the recent RIME framework already supply strong predecessors. The Encevra contribution is a specific test of question changes after a handoff is frozen, including confidently wrong answers that may never trigger ordinary uncertainty-based recovery. An exact example shows that preserving the initial decision perfectly can leave a 25% irreducible error on a later question under stipulated assumptions. A factorial study separates the contribution of structured representation from the contribution of an explicit scope check, while charging writing, retrieval, validation, and failed attempts. The proposal should be rejected or simplified if a strong prose summary with the same source access performs as well.

1. Numbering, archive inspection, and continuity

The current archive contains completed research through Round 16, so this paper is Round 17, correcting the original Round 11 designation. The archive was inspected at project commit 6698f4d26eac8c0df0911a7d0bdc313b705a61ea. The Research overview and rounds index include Round 16. The completed Lovable handoff reports that its full manuscript is preserved and its preview checks passed. This is evidence about a completed research entry, not completion of its proposed model experiments.

The full texts of Rounds 9 and 10 were retrieved again and confirmed unchanged from the versions previously inspected in full. Round 16 was checked against the complete authored manuscript. The Round 15 internal manuscript and the previously inspected relevant work in Rounds 3, 4–6, 8, and 13–14 inform the carry-forward below. Private notebook text and implementation source are not reproduced. [I1–I6]

Round 15 explicitly assigned message representation under a bandwidth limit to Round 17; Round 16 carried that question forward. This paper follows that sequence. It does not treat the earlier broad H10-B claim as a new discovery. The narrower intervention is an explicit check of a representation's scope when the question changes, with the future question withheld from the compressor and source access held comparable across methods.

The archive still describes the relevant model studies as proposed. This paper does not bypass the Round 13 collection gate or rewrite the research roadmap's phase boundaries. Another polished mechanism description is not a substitute for a valid collector and independent endpoints.

2. Research question and exclusions

Research question: When a follow-up question changes which facts are decisive, does a compressed handoff that records dependencies and a declared scope improve supported completion over a strong contextualized prose summary, when both can reopen the same sources and all work is charged to the same budget?

The mechanism question is whether an explicit scope check improves recovery beyond the receiver's ordinary judgment that it needs more information. A receiver can be confidently wrong without saying that its memory is insufficient. A scope check could help, but it could also add redundant calls, overstate its own coverage, or delay a correct answer.

The primary setting is a sequence with an immutable source packet, an initial question q₀, a compressed handoff, and a later q₁. The source facts do not change. Only the information needed by the question changes. Examples include moving from overall release readiness to the status of a specific prerequisite, or from a total amount to its components. Source updates and contradictory new evidence are reserved for a later round. Snapshot-mismatch checks below are interface diagnostics, not a study of update resolution.

We exclude model-weight or KV-cache compression, training a new language model, clinical or legal deployment, general truth certification, and claims that an architectural diagram reproduces the brain. The proposed intervention can be implemented by one model with external state. It does not require an additional specialist or establish that a team is preferable.

The target is supported task completion, not preservation of an earlier model's wording or decision at any cost. If the full-source answer is wrong, compression-induced disagreement may be an improvement. If the original task permits a conditional answer, maintaining that uncertainty may be the correct behavior. An evaluator must distinguish source-relative agreement, correctness under the task, and support from actually available evidence.

3. Literature map and competing explanations

The focused search was conducted on 8 October 2026. It covered task-agnostic compression, question-aware retention, source reopening, memory evaluation, decision fidelity, and a limited biological analogy. Primary papers were inspected directly at the depths recorded in Section 16. The review is not exhaustive, and several recent papers remain preprints.

The evidence supports several viable explanations. Compression can remove important information. It can also improve a fallible reader by excluding distractions or changing placement. Retrieval can repair a missing detail, yet fail when the reference is stale, the query is poor, or the receiver overlooks the retrieved passage. A structured record can work because of formatting rather than better information retention.

The strongest alternative to the proposal is therefore not a careless generic summary. It is a carefully prompted prose record that preserves qualifiers and source references, paired with a capable reader free to request the original evidence. ReadAgent is an important precedent for that baseline. EPC is an important retention competitor, and RIME is a recent recovery competitor. [S2–S4]

We do not import their reported scores into an Encevra comparison. The setups differ in writer models, budgets, memory construction, target tasks, and judges. RIME reports query-time token costs; total-work superiority here remains untested. WhenLoss treats its performance gaps as operational diagnostics with format and reader effects, rather than unique causal decompositions. [S3, S4]

4. What the current implementation establishes

The inspected inference path first limits incoming material to a profile-dependent prefix. It constructs findings and performs claim review; that review receives both a local quote window and the full material admitted after the cap. Synthesis then receives approved findings, and the final check compares the brief with that approved-finding representation. [I6]

This is a useful set of research boundaries: admission of source material, transformation into findings, and use of those findings in an answer. A detail could be absent at any one of them. The audit does not show that the claim reviewer sees only a short quote window; the complete admitted material is also supplied. Nor does it establish that the production system has no other storage or retrieval path.

The narrow audit target is preservation across representations. Before evaluating a follow-up workflow, verify which original source bytes are available, which version and spans a finding identifies, and whether a later receiver can actually recover them. A citation identifier is not proof that the cited information reached the next model call. A source locator is not a free, always-working memory channel.

No inference was run against this path, and no failure rate is inferred from static code. Other application files and persistence paths were not comprehensively audited. The proposed protocol has not been installed in Noetis. Its value must be measured before changing production behavior.

5. A formal distinction: sufficient now versus sufficient later

Consider a finite set of possible source records W. A question q has a task-defined answer g_q(w). A deterministic compressor C maps a record to a message m. For a declared question family Q, call the message representation sufficient if every pair of records mapped to the same message has the same required answers for all questions in Q:

C(w) = C(w')  implies  g_q(w) = g_q(w') for every q in Q.

This is an elementary indistinguishability condition, not a new information-theoretic result. It concerns the information represented, assuming a correct decoder and the declared task model. A language model can fail even when the condition holds; a lucky answer can succeed when it does not.

Two records are equivalent for Q when their answer vectors (g_q(w): q in Q) agree. A representation that always answers every question in Q correctly must distinguish these equivalence classes. If there are K classes and no extra information channel, at least K message states are required, or at least ceiling(log₂ K) bits for a fixed-length binary code. This is a counting statement, not a token-budget forecast.

Expanding Q can split an equivalence class. A record that was sufficient for the old question family can become insufficient for the larger one even though neither the source nor the compression changed. That is the key difference from Round 16's discovery of a fact missing from an initially posed decision: here the fact may have been observed and discarded under an earlier task.

Source reopening changes the problem. A reader now uses a message plus a transcript of permitted source accesses. It can recover distinctions absent from the message if those accesses supply them. The message-only bound does not apply to that enlarged channel as though retrieval were free. Count the source store, index, locator metadata, query construction, returned text, validation, and failure behavior in the system description.

A hash can bind a response to source bytes. It does not show that the source is true, that a passage entails a claim, or that a compressed record is sufficient for all future questions. If identifiers themselves encode information about the answer, they are an information channel too. Randomize neutral identifiers and prevent evaluator-only labels from entering either the message or the retrieval interface.

6. Worked example: release readiness is not a diagnosis

Take a fictional software artifact whose release requires both a passing test suite and a valid signature. Let t and s denote those two binary conditions. The initial question asks whether release is permitted; its answer is t AND s.

A one-bit message containing only the correct release answer is perfect for q₀. It merges records A, B, and C. Now the user asks, “Did the tests pass?” Records A and C require different answers but produced the same message. Under a stipulated uniform distribution over the four records, even the best decoder using that message alone achieves only 3/4 accuracy on the follow-up. No additional deliberation over the same bit can distinguish the merged cases.

This does not mean a real summarizer necessarily loses a quarter of its accuracy. The distribution, decision rule, and encoding are deliberately constructed. The example establishes a possibility and a diagnostic target: zero initial-task loss does not imply zero later-task loss.

For one concrete record, the source says: “Artifact Orion. Snapshot v7. Tests: PASS. Signature: FAIL. Release requires both tests and signature to pass.” A better retained record might preserve the two conditions explicitly. If the budget cannot accommodate them, a record can identify the source and state that it was compressed for release readiness, not for identifying which prerequisite failed.

When the later question asks about tests, the receiver can reopen the pinned source and answer that the tests passed. If access fails, it should state what the available record establishes: release was not permitted, but the compressed conclusion alone does not identify the failing prerequisite. It should not convert lack of retained detail into a claim that tests failed.

The example also exposes the trigger problem. A receiver might confidently reason, incorrectly, that “not ready” implies “tests failed.” A recovery mechanism activated only by an explicit insufficient-answer signal would never run. This is a hypothesis about a possible behavioral failure, not an observed RIME or Noetis error. The empirical study must compare it with an ordinary reader allowed to inspect sources proactively.

7. Proposed protocol: retain scope and reopen evidence

The proposed handoff is a bounded record, not an authoritative replacement for the source. The writer and receiver may be the same model in separate stages.

1. Freeze the writing boundary. Save the admitted source packet, version manifest, q₀, output contract, and actual writer input. The future q₁ remains hidden. Record excluded or unprocessed material rather than describing the packet as exhaustive.
2. Write under the message budget. Preserve supported facts and material relations such as conjunctions, exceptions, units, temporal scope, and unresolved conditions. Attach a compact source locator and the scope for which the record was constructed. All fields count toward the same cap.
3. Bind claims to evidence. Store a stable document/version identifier and a resolvable span or section reference. The external source store preserves original material. A locator is a request path, not evidence that a reader has inspected the source.
4. Check the new question. At q₁, ask whether the retained facts and relations support the requested answer. An unfamiliar target, entity, granularity, or condition triggers reassessment even if an initial answer sounds confident. A changed wording alone need not trigger retrieval.
5. Reopen within budget. Request the smallest useful source portion, including enough surrounding context to interpret qualifications. Check identity and applicability. Keep a trace of returned text and what was actually included in the receiver's prompt. Charge failed requests as attempts.
6. Answer or preserve the limit. Evaluate the exact terminal response under q₁'s contract. If the source is unavailable, the budget is exhausted, or support remains uncertain, give the strongest supported answer rather than guessing a lost condition. Avoid a blanket refusal when a useful conditional answer is possible.

The scope field is a claim about the writer's representation, not a formal certificate in natural language. A schema can omit the same condition as the writer. A scope check can incorrectly classify a follow-up as covered. We therefore measure scope-check errors directly and retain a source-retrieval baseline that does not depend on that field.

For the first study, use a simple record with five elements: the current task, supported facts, their required relations, unresolved or excluded coverage, and source locators. This is a proposed interface, not a claim that five fields are optimal. If the fields crowd out decisive facts, the design fails its own budget constraint.

A concise receiver instruction suitable for development is: “Use only the supplied record and permitted source reads. Check whether this question asks for a distinction the record retained. A previous correct conclusion does not establish all of its prerequisites. If needed, request the pinned evidence and reassess; report unresolved support if access fails.” This instruction is a test condition, not a proven safeguard.

8. Executed diagnostics and their limits

The supplied Python program enumerates every partition of the four source records in Section 6. A partition represents which records a deterministic message encoding makes indistinguishable. It checks three growing question families: release only; release plus test status; and release plus both prerequisite statuses.

The program also enumerates 60 deterministic recompressions of those partitions. None restores a question distinction already lost. The reason is simple: a function of an existing message cannot split records that already produced the same message without new information. Some recompressions preserve all tested distinctions, so this result does not imply that every compaction must cause additional loss or that degradation is geometric.

The exact diagnostic verifies the 3/4 follow-up accuracy ceiling for the initial-answer-only encoding. It also checks that a stipulated fallible reader can perform better on a useful compact message than on a full record. Information preservation and a particular reader's accuracy are different objects. Compression can improve the latter without creating new information.

A toy resolver checks successful source access, missing sources, exhausted access budgets, and mismatched snapshots. It charges failed lookups and demonstrates that a matching content hash does not validate the truth of source statements. A deletion example removes “not” from a sentence while leaving every remaining word in its original order. This shows why an extractive operation alone cannot guarantee semantic preservation; it is not a reproduced failure of LLMLingua-2.

Twenty-eight named checks passed. Twelve small developer/evaluator fixtures were generated. Their source-state tables and answers must remain hidden from future tested models. They are not a held-out benchmark, independently sampled tasks, or estimates of real-world error frequency.

No language model generated the encodings, detected question changes, selected a passage, or interpreted a source in these diagnostics. The finite task IDs and reference answers were supplied explicitly. The difficult empirical work concerns those missing capabilities, not whether a four-world truth table can be enumerated.

9. Claims, evidence, and counterclaims

The established part is the logical distinction and the existence of tested compression/retrieval approaches. The mixed evidence concerns which representation and access policy works under which tasks and costs. The open Encevra question is incremental utility under question changes and a fully accounted budget. None of the sources establishes this local effect without a new experiment.

10. Proposed experiment: freeze the message before the question changes

E17-A denotes the executed exact diagnostics. E17-B is a proposed model study, gated on the existing collection requirements. Preserve the initial and final artifacts, source versions, all provider attempts, actual receiver exposure, and complete or explicitly partial usage. Do not classify the local reference script as passing the end-to-end collection gate.

Author twelve independent source families for a development pilot. For each family, construct an initial q₀ and three follow-ups: a paraphrase requiring the same distinction, a related question requiring a different retained condition, and a question requiring a detail absent from the initially useful answer. The latter detail remains in the original source. Match answerability, source length, distractors, and evidence position where feasible. Include benign unknowns and questions that are genuinely unsupported even by the full source.

The writer sees the source and q₀, plus a generic instruction to support later use. It never sees the actual follow-up question or hidden rubric. Freeze and hash the message before q₁ is revealed. Reuse that message across the three follow-up branches; isolate branches so one branch's retrieved fact cannot leak into another. A declared query family may be supplied in a separate condition, but do not silently give the candidate a better forecast than the baseline.

Use a two-by-two comparison:

Both writing prompts request preservation of qualifications, uncertainty, and source identity. Do not reserve those desirable instructions for the candidate. The primary contrast is S1 versus P1, isolating representation given the same explicit check. P1 versus P0 and S1 versus S0 estimate the incremental effect of the check. The interaction with question change tests whether gains concentrate where the proposal predicts.

Two independent writer realizations per family and representation, each evaluated under both recovery policies and three follow-ups, yield 288 assigned answer episodes at one budget. The repeated answers and branches are not independent source families. Within each representation, reuse the same frozen writer output across its two recovery-policy branches; charge writing cost to each hypothetical deployed policy while recording the actual research bill separately.

Additional controls should include query-agnostic extractive retention, an EPC-style predictive-retention method, and a faithful ReadAgent-style gist-and-lookup adaptation. Compare a policy that reopens evidence only after an explicit insufficient answer as an ablation, not as the sole recovery baseline. Adaptations must be labeled and documented; they are not replications of published systems. These extra conditions require additional assignments beyond the 288 core episodes.

11. Equal resources and diagnostic interventions

The pilot planning point is a 768-token handoff cap, counting metadata and source locators, and a 24,000-token total model-work ceiling per hypothetical episode, counting all input and output across writing, checking, lookup-query generation, answering, and repair. Reported reasoning tokens count exactly once under the provider's convention. Use one pinned backbone and reasoning setting. These are provisional development limits, not observed requirements or a spending commitment.

Use source packets of at most approximately 6,000 tokens for this pilot, so a one-pass writer can inspect the same admitted material in every condition. Give all core conditions the same immutable source store, the same retrieval interface, and at most two source reads with a common 800-token response cap. Fix query and result formatting in advance. All prompts and returned evidence count toward inference usage when supplied to a model. Source operations, indexing work, storage, latency, and validation overhead are separately logged.

Reserve room for a final response before issuing another diagnostic call. Malformed messages, over-budget attempts, retries, failed reads, and timeouts remain assigned episodes. A hard message limit needs an honest overflow status or declared truncation policy; do not silently cut a dependency and then call the record complete. Allow simpler methods to spend unused work on their best feasible reasoning, not padded output.

At the stated ceiling, 288 core episodes permit at most 6.912 million charged policy tokens under complete accounting. Shared writing reduces the actual bill, not the deployed-policy cost attributed to each branch. Neither token totals nor identical caps imply exact FLOPs or identical realized spend. If material usage is unavailable, a strong matched-total-compute claim is blocked. Record prices and cached-input treatment at collection time rather than importing current price estimates.

The handoff cap and the inference-work ceiling are different constraints. Report them separately. After feasibility, predeclare another handoff cap and work ceiling and compare quality–cost curves. A method that wins only with more source reads or substantially more expert checking has not earned an equal-resource claim.

For failure localization, run three controlled interventions on a small diagnostic subset. First, supply the already stored handoff directly, bypassing retrieval of that handoff. Second, append independently identified decisive source passages, bypassing passage search. Third, let the same reader inspect the complete admitted source. These are access and presentation interventions, not infallible oracles or uniquely identified causal effects. They can change formatting, distraction, and available context as well as information.

The complete-source control may exceed the handoff or per-read cap; label that resource relaxation explicitly. It answers whether a simpler full-source reader can solve the task, not whether the candidate beat an equally constrained baseline. If a deployed direct-read policy is feasible at comparable total cost, compare it on the same quality–cost frontier. A complex compression workflow must justify itself against that option.

12. Metrics that reveal harmful compression

The primary endpoint is supported completion on changed-question episodes, over all assigned instances. The terminal answer must satisfy the frozen q₁ contract, be correct under the independent reference, and be supported by the evidence actually exposed to the receiver. Report unchanged-question controls separately.

For repair and damage, compare an initial and terminal answer to the same follow-up question. Comparing q₀'s answer with q₁'s answer would mistake a legitimate question change for a revision error. Store an initial q₁ draft for a prespecified diagnostic subset if the deployed policy does not normally produce one; account for that added cost and keep the subset's estimand separate.

Factual consistency, lexical overlap, and semantic similarity are useful diagnostics, but they do not replace the task endpoint. A summary may copy every included fact and still omit the relation that makes those facts decisive. Conversely, a shorter message can improve a model's use of the retained evidence. The study must allow that outcome rather than treating compression as necessarily harmful.

Report recovery failure conditional on whether the source was available, as well as completion across all assignments. Preserve inaccessible-source and budget-exhaustion cases in the overall denominator. Include deliberately irrelevant source reads and already sufficient messages so an always-reopen policy cannot win by exploiting an evaluation that rewards retrieval itself.

Use executable reference tasks where possible. On semantic tasks, preserve independent first-pass labels, adjudication history, and unresolved status under Round 14. A model judge that sees its own preferred representation may be biased. Blind the representation label and use independent evidence-based review on the consequential contrast. Do not use the full-source model's answer as the only ground truth.

13. Falsifiable hypotheses and decision rules

H17-1: Relations survive useful compression. Under the same explicit scope check, S1 improves changed-question supported completion over P1 without unacceptable damage or cost. It is weakened if the apparent benefit comes from longer messages, extra retrieval, easier prompts, or a favorable judge. It is rejected as an architectural requirement if strong prose performs equivalently within a sufficiently narrow interval or wins at lower cost.

H17-2: Scope checks recover confident omissions. The explicit check improves recovery on changed-question cases, including cases where an ordinary receiver would answer rather than signal insufficiency. Test it against unrestricted discretionary lookup, not only a weak “retrieve after refusal” rule. A check that mainly increases unnecessary reads or misses the same omitted conditions is not a useful mechanism.

H17-3: Reopening matters only when it supplies usable evidence. Benefits decline when decisive evidence is unavailable or cannot be found within the allowed accesses. Correct behavior should then shift toward bounded, supported responses rather than fabricated recovery. If gains persist without exposing the needed evidence, investigate prior knowledge, label leakage, judge preferences, or another mechanism before crediting retrieval.

The 288-episode pilot is a feasibility design. It does not establish power for small effects. Average repeats and branches appropriately within their source family, and use family-level resampling or a prespecified clustered analysis for future-family claims. Generate held-out families independently; reworded fixtures and repeated seeds are not new independent evidence.

For a later confirmatory study, provisional practical criteria are at least five percentage points of improved changed-question supported completion, no more than two points of additional same-question damage, and no hidden material cost increase outside the declared budget. These are planning margins, not safety guarantees or predicted effects. Freeze one primary comparison, one budget point, the joint benefit/harm decision rule, and sample size using development-family variability before examining held-out results. Confidence bounds must support the chosen margins; a favorable point estimate is insufficient. Prespecify treatment of secondary contrasts and multiplicity.

The simplest winning policy should determine the implementation. If adding a scope check to prose captures the gain, keep that and drop the extra schema. If always reading a small original packet wins at comparable cost, use the packet. If no approach can preserve broad future utility at the chosen cap, increase the cap, narrow the promised question family, or retain a better source-access path rather than rename the failure.

14. A limited brain-inspired design idea

Horner and colleagues studied people learning overlapping event associations and used behavioral results, fMRI, and a computational model to examine holistic recollection. Their results support the possibility that a cue reinstates other elements of a bound event, including elements not directly requested. This provides a functional analogy for recovering related context, not evidence that a source database is a hippocampus. [S8]

The testable engineering idea is to reopen a related evidence bundle when one retained fact depends on a qualification elsewhere. Compare a single-span return with a linked fact-and-qualification return under the same returned-text and total-work budget. Include a random-neighbor bundle control. Measure recovery of decisive relations and contamination by irrelevant or misleading associated material.

A bundle can help through ordinary context restoration, with no biological explanation needed. It can also hurt by pulling in distractors. Test actual retrieved source bytes, not imaginative reconstruction of plausible missing details. The analogy motivates the ablation; it establishes neither a new mechanism of intelligence nor literal brain copying.

15. Limitations and next-round handoff

The scope description may be wrong. A writer can describe a narrow summary as broadly sufficient. Natural-language coverage judgments are not finite proofs. Evaluate missed scope changes and prohibit a metadata field from serving as its own ground truth.

The question distribution may change. Predicted future questions can favor common requests and sacrifice a rare decisive detail. Report results by declared and unanticipated question classes. No finite compression scheme should promise lossless answers to an unrestricted universe of future questions.

Pointers may conceal costs or failure. An opaque source reference can be cheap to store but expensive to resolve. Authorization, availability, version checks, search quality, and returned context are part of the system being evaluated. Preserve these failures rather than treating recoverability as a property of the message alone.

A better reader may make the mechanism unnecessary. Formatting, evidence placement, or direct source delivery can explain gains. Reassess with a strong current backbone and a single-agent implementation before adding team complexity. Older context-use studies motivate tests, not a claim about present-model deficits.

Our sources have limited transfer. Conversational memory differs from bounded operational decisions. The financial fidelity paper uses a source-reading model as its reference, not verified investment truth. WhenLoss uses task proxies and adapted baselines; RIME uses model-based judging and reports query-time efficiency. The human memory study is not a software evaluation. These limits prevent a pooled estimate of Encevra's expected gain.

Our new evidence is finite and constructed. The 15 partitions, 60 recompressions, 28 checks, and twelve fixtures test a reference definition and interface behaviors. They are not a language-model benchmark or proof of successful natural-language scope detection. No new Noetis performance result, learned representation, calibrated risk estimate, or deployed architecture follows.

Next handoff: complete the existing collection gate; preserve the source-to-handoff-to-retrieval-to-answer chain; validate independent labels for relation retention and changed-question support; then run the smallest useful factorial pilot. The main decision is whether a strong summary, a scope check, or a structured representation earns its cost. Round 18 may address what happens when the underlying evidence itself changes or conflicts, but should use the same recorded boundaries and prioritize any collection failure exposed here.

16. Full sources and inspection record

All external PDFs below were retrieved on 8 October 2026. Dates refer to the publication or inspected edition. The source ledger records working links, file hashes, and the same inspection scope. Selected sections were inspected; no cited experiment suite was independently reproduced. External PDFs and private earlier notebooks are not redistributed in the package.

[S1] Zhuoshi Pan, Qianhui Wu, Huiqiang Jiang, Menglin Xia, Xufang Luo, Jue Zhang, Qingwei Lin, Victor Rühle, Yuqing Yang, Chin-Yew Lin, H. Vicky Zhao, Lili Qiu, and Dongmei Zhang. LLMLingua-2: Data Distillation for Efficient and Faithful Task-Agnostic Prompt Compression. Findings of ACL, August 2024, pp. 963–981. Inspected publication record and PDF pages 2–4 and 6: task-aware versus task-agnostic framing, distillation, extractive constraints, classifier strategy, and evaluation setup. Supports an extractive baseline; no Encevra preservation guarantee follows. Record · Inspected PDF.

[S2] Kuang-Huei Lee, Xinyun Chen, Hiroki Furuta, John Canny, and Ian Fischer. A Human-Inspired Reading Agent with Gist Memory of Very Long Contexts. ICML, 21–27 July 2024, PMLR 235:26396–26415. Inspected PDF pages 2–5: pagination, gisting, parallel/sequential lookup, amortized processing costs, reader and judge setup. Supports the recovery baseline; its historical models and word-count accounting are not a current total-compute result. Record · Inspected PDF.

[S3] Jiangnan Yu, Kisson Songqi Lin, and Jilong Wu. WhenLoss: Diagnosing Write and Retrieval Bottlenecks in Long-Context Memory Systems. arXiv:2605.24579v1, 23 May 2026, preprint. Inspected PDF pages 1–6: diagnostic protocol, caveats, EPC, configurations, and main results. Supports stage comparisons and predictive retention. Scores use Contains Match/token-F1; original LongMemEval uses model judging. Adaptations and experimental code were not independently audited. Record · Inspected PDF.

[S4] Wanqi Zhou, Jiawei Lu, Yang Wang, Zhaolong Xing, Zhen Chen, Ai Han, and Haoyue Shi. Remember by Asking: Retrieval-Induced Memory Evolution for LLM Agents. arXiv:2609.34438v1, 28 September 2026, preprint. Inspected PDF pages 1–4: formation, recovery trigger, setup, and ablations. Supports a recovery competitor. The same backbone judges answers; adversarial questions are excluded; reported efficiency is query-time. Record · Inspected PDF.

[S5] Hoyoung Lee, Suhwan Park, Seunghan Lee, Jun Seo, Jaehoon Lee, Sungdong Yoo, Minjae Kim, CheolWon Na, Zhangyang Wang, Zach Golkhou, Minkyu Kim, Sotirios Sabanis, Alejandro Lopez-Lira, Dhagash Mehta, Soonyoung Lee, Chanyeol Choi, Wonbin Ahn, and Yongjae Lee. When Summaries Distort Decisions: Information Fidelity in LLM-Compressed Financial Analysis. arXiv:2606.29251v1, 28 June 2026, preprint. Inspected PDF pages 1–7: source-relative decision fidelity, rereading-noise comparison, candidate auditing, qualifier add-back, costs, and limitations. Supports evaluating consequential omissions. The fixed budget is expressed in bullets, add-back is a diagnostic that exceeds the budget, and source-model agreement is not verified decision correctness. Proprietary forecasting claims were not adopted. Record · Inspected PDF.

[S6] Di Wu, Hongwei Wang, Wenhao Yu, Yuwei Zhang, Kai-Wei Chang, and Dong Yu. LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory. ICLR 2025. Inspected conference PDF pages 1 and 3–6: capabilities, evidence/session construction, annotations, question-answer judging, and retrieval metrics. Supports observation boundaries and task diversity. Chat histories are constructed and curated; reported historical performance is not a current Encevra baseline. Record · Inspected PDF.

[S7] Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics 12:157–173, February 2024, DOI 10.1162/tacl_a_00638. Inspected PDF pages 1 and 3: controlled position changes, task setup, and source/distractor design. Supports position controls and the distinction between exposure and successful use. No historical error magnitude is asserted for current models. Record · Inspected PDF.

[S8] Aidan J. Horner, James A. Bisby, Daniel Bush, Wen-Jing Lin, and Neil Burgess. Evidence for Holistic Episodic Recollection via Hippocampal Pattern Completion. Nature Communications 6:7462, 2 July 2015, DOI 10.1038/ncomms8462. Inspected PDF pages 1, 7, and 8: abstract, interpretation, associative-structure limits, and participant/procedure methods. Supports a limited functional analogy and linked-context ablation, not software efficacy, neural equivalence, or a causal claim about Encevra. Record · Inspected PDF.

Internal continuity. [I1] Round 9, 30 September 2026, full procedural-learning agenda. [I2] Round 10, 1 October 2026, full paper including H10-B and the later communication agenda. [I3] Round 13, 4 October 2026, and Round 14, 5 October 2026, collection and evaluation commitments carried forward. [I4] Round 15 public summary, 6 October 2026, and separately inspected full internal roadmap. [I5] Round 16, 7 October 2026, complete missing-condition paper. [I6] Pinned project snapshot identified in Section 1: overview, index, registry, unchanged prior full texts, and the inference path's material admission, review, synthesis, and final-check inputs. Canonical links identify intended archive pages; repository source was inspected because preview edits may precede deployment. Private notebooks and implementation files remain private.