Encevra Research Round 16
Before the Answer Closes: Detecting Missing Decision Conditions
7 October 2026 · Research paper · Version 1.0
Evidence status: Focused primary-source review, a read-only inspection of the current research archive and one inference implementation file, and executed finite-world diagnostics. The diagnostics cover 6,912 Boolean function–observation pairs and 37 named checks; twelve small evaluator fixtures were generated. No new Encevra language-model benchmark, model inference, human annotation, training, production change, or deployment was performed. The model studies below are proposed, not preregistered or run.
Abstract
An answer can cite accurate evidence and still depend on a fact nobody checked. Encevra's recent work asks whether messages help a receiver and whether complementary evidence is worth retaining. This round moves one step upstream: how does a system discover that a decision condition is missing before it knows which message to request? We propose a bounded contrastive sufficiency probe. It searches for two scenarios consistent with the admitted evidence that require different answers, validates the difference, and turns the unresolved condition into a specific acquisition request or conditional answer. Failure to find such a pair is recorded as unresolved search, not proof of completeness. A separate source-coverage probe tests whether the system omitted a relevant rule or source from its representation. The inspected implementation supplies a concrete target: a broad evidence-sufficiency field reflects the existence of at least one supported finding, which is weaker than decision sufficiency. Exact diagnostics demonstrate that distinction, the danger of treating bounded search as certification, and the reopening of a decision when the modeled scope expands. A controlled-budget experiment compares the proposal with direct reasoning, an explicit checklist, and simpler evidence delivery. The contribution is a falsifiable measurement and interface proposal, not a new general theory of information acquisition or evidence that collaboration improves Noetis.
1. Archive correction, continuity, and novelty
The current archive contains completed Rounds 11 through 15. The next sequential entry is therefore Round 16, correcting the original Round 11 designation. The archive was inspected at project commit 3527c9e6563ac8730b64dbed49a2cfebb85e2b1e, including the Research overview, rounds index, publications and research registry, full texts of Rounds 9 and 10, and relevant earlier work. Round 15's public entry is a summary; its complete internal manuscript and handoff were inspected separately. The expanded internal Round 14, edition 1.1, was also consulted. No private notebook text from those rounds is reproduced here. [I1–I7]
| Earlier commitment | Carry forward | What this round changes |
|---|---|---|
| Round 9: learn procedures from verified experience | A procedure needs conditions of applicability. | Investigate discovering an unrepresented condition before procedure reuse or learning. |
| Round 10: evidence exchange and selective revision | Preserve evidence, origin, and independent initial outputs. | Test the detection of an evidence gap before a useful message has been identified. |
| Round 11: allocate verification effort | Benefit, damage, and total cost matter. | Hold allocation simple; investigate the candidate gap that a controller would otherwise never see. |
| Round 12: terminal release risk | The evaluated object is the exact final artifact. | Separate evidence sufficiency from probability of correctness or permission to release. |
| Round 13: observable transitions and independent endpoints | Capture initial and terminal outputs, exposure, and all attempts. | Add sufficiency and missing-condition labels to the existing collection gate. |
| Round 14: fallible evaluation and targeted audit | Preserve unresolved labels and audit the claimed contrast. | Distinguish uncertain annotations from genuinely underdetermined tasks. |
| Round 15: context-dependent message value | Useful prerequisites may need to be combined. | Ask how a prerequisite enters the workspace in the first place. |
The choice follows Round 15's explicit Round 16 handoff on noticing missing information. Confidence calibration would substantially overlap Round 12. Review allocation overlaps Round 11 and the expanded Round 14. Another treatment of message complementarity would duplicate Round 15. Missing-condition discovery is adjacent but distinct: a system cannot request, retain, or price a dependency that its current task representation omits.
This question already has substantial prior art. Context-sufficiency analysis, clarification benchmarks, constraint-based information seeking, and structured evidence acquisition cover much of the territory. [S1–S6] The specific Encevra proposal combines a validated decision-changing witness, an explicit scope limit, and a separate source-discovery intervention. Its novelty is an experimental distinction and a concrete system target. It does not claim to invent possible-world reasoning, counterexamples, checklists, or clarification.
The public registry and the internal milestone document use different roadmap phase boundaries. This round does not silently rewrite either. It follows their shared operational priority: complete the existing collection gate, investigate a bounded mechanism, and require actual observations before architectural promotion.
2. Research question and boundaries
Research question: At a fixed total inference and source-access budget, does requiring a validated pair of evidence-compatible scenarios with different decisions reduce unsupported commitment more effectively than a strong missing-information checklist, without reducing useful completion or increasing harmful revision?
A secondary, mechanism-specific question asks whether a small source-coverage probe helps when the missing condition is absent from the system's representation, rather than merely an unfilled value in a known schema. These are parts of one discovery problem, not separate proposals for a new memory system or verification scheduler.
The primary scope is a bounded document decision: an explicit task, a declared evidence hierarchy, a time or version, and a specified output requirement. Examples include fictional policy application, structured calculations, and software configuration constraints. We exclude clinical deployment, legal advice, general open-world truth certification, model-weight training, and claims about consciousness. Medical and legal datasets discussed in the literature are evidence about those studies, not validation for consequential use.
Three distinctions define the scope. Correctness asks whether an answer matches the relevant outcome. Support asks whether the permitted evidence justifies the answer under the task's rules. Sufficiency asks whether that evidence determines the requested answer at the specified granularity. A lucky guess can be correct but unsupported. A well-supported amount can be insufficient to decide eligibility. A complete source packet can still be misread.
The permitted output matters. If the task allows a bounded conditional answer, the system can say what follows in each admissible case. If the task requires a single executable action, the same evidence may be insufficient. The protocol must not reward refusal when a useful conditional answer satisfies the user's actual request.
3. Literature map: competing positions
| Research line | What the inspected evidence supports | What remains unsettled |
|---|---|---|
| Context sufficiency: Joren et al. [S1] | Separating inadequate context from failure to use it changes interpretation of RAG errors. Their sufficiency autorater is itself fallible. | A context classifier does not prove that a source set is complete or that a decision is correct. |
| Constraint-based acquisition: QuestBench and MT-InfoSeek [S2, S3] | Missing-variable selection and final information sufficiency can be evaluated separately from answer accuracy. | A supplied variable universe simplifies the discovery of an entirely omitted condition. |
| Structured evidential acquisition: InfoGatherer [S4] | A document-grounded evidential network supplies an explicit alternative to unstructured confidence-based questioning. | The learned graph and elicited masses can be wrong; dialogue turns are not total compute. |
| Clarification taxonomies and checkpoints: CLAMBER and AskBench [S5, S6] | Missing intent, ambiguity, false premises, and unnecessary questioning require different labels. | A benchmark's hidden checklist is not automatically a complete model of a real user's needs. |
| Recognition versus action: Su and Cardie [S7] | Explicit ambiguity judgments and spontaneous clarification behavior can diverge. | Prompt framing and model-judge errors limit causal interpretation of that gap. |
| Discovery versus use: Engländer et al. [S8] | An agent may encounter a relevant resource without investigating it. | Obvious injected solutions differ from naturally occurring, fallible evidence. |
| Adversarial unanswerability: SQuAD 2.0 [S9] | Plausible answer cues should remain in negative examples. | Extractive unanswerability alone does not evaluate multi-source decision prerequisites. |
These positions should not be collapsed into “more questions help.” A classifier may identify a gap while the answering policy ignores it. Retrieval may make a query answerable while also encouraging premature commitment. A structured model can improve acquisition within its represented variables while omitting an important variable entirely. Conversely, a strong model may answer a general-knowledge question correctly using permitted prior knowledge even when the retrieved text is insufficient. Our evidence-restricted task contract must therefore be declared, not retroactively chosen to make one method look safer.
The closest formal predecessors are [S2, S3]. Round 16 adopts their separation between acquiring sufficient information and computing the answer. It does not present that separation as a new invention. The added stress condition is representation incompleteness: relevant rules, source classes, or decision conditions are not handed to the detector as a complete list.
The strongest rival is simple delivery. If giving one capable agent the complete, well-organized admissible source packet removes the failure, the useful change may be retrieval or presentation. A specialist team and a complex gap detector would still need to earn their costs.
4. A current system target, with a narrow audit claim
The inspected inference file computes a broad evidence-sufficiency field from whether at least one approved, verified, supported non-question finding exists. Its readiness wording also describes the presence of a supported finding. That is an existence condition, not an exhaustive check that all decision-relevant prerequisites have been established. [I7]
This is a static observation about one pinned code snapshot. It is not a measured failure rate, a complete audit of every release safeguard, or proof that a user received a wrong answer. The same file includes review, final checking, repair, and additional caution mechanisms. Those can catch particular omissions. The narrow conclusion is that the meaning of this field cannot by itself support a completeness claim.
The inspected path also retains initial checks internally but does not directly return an immutable initial answer alongside the final answer in the returned object. Other storage paths were not comprehensively audited. Consequently, this inspection does not mark Round 13's end-to-end collection gate as passed or failed across the entire application. It identifies a concrete binding to verify before live measurement.
An immediate research interface should distinguish “some supported evidence is present,” “a decision-changing gap has been witnessed,” and “the decision is determined within a declared, checked scope.” Renaming or expanding a field is not a performance result. The proposed detector must be evaluated separately from the precision of its labels.
5. What kind of information is missing?
| State | Example | Appropriate test or response |
|---|---|---|
| A known variable is unfilled | The exception rule is visible; qualifying status is absent. | Seek the value or retain a conditional answer. |
| A relevant source is available but unexposed | The catalog lists an amendment that never entered the receiver's input. | Inspect that source and preserve the exposure record. |
| A condition is omitted from the representation | The detector models only the threshold, not the exception. | Search beyond its current checklist; challenge the modeled scope. |
| The evidence is inconsistent | Two equally admissible records give incompatible effective dates. | Preserve conflict; do not treat an empty feasible set as certainty. |
| The request is ambiguous | “Current policy” has no specified date or entity. | Resolve intent or answer the explicit alternatives when permitted. |
| An unknown is irrelevant to the decision | Item color is missing but cannot change review requirements. | Answer without a needless question. |
| Information is sufficient but unused | All conditions are exposed; the model applies the wrong rule. | Investigate reasoning or integration, not acquisition alone. |
These are operational labels, not mutually exclusive diagnoses of an internal mental state. One episode can contain an unexposed amendment, an omitted exception variable, and a mistaken calculation. A stage label needs the evidence that supports it: corpus membership, actual prompt exposure, a validated alternative, and terminal behavior.
“Unknown unknown” is relative to the representation. A condition can be unknown to the receiver while obvious to a source catalog or a human reviewer. If the entire evidence-generating process hides a condition with no observable clue, the detector has no universal way to discover it. We should measure identifiable opportunities for discovery rather than promise omniscience.
6. A scope-bound sufficiency contract
For a closed, finite fixture, let M specify the variable domains, admissible rules, source precedence, target, and output contract. Let E be the observations actually exposed to the receiver. Let Ω(M,E) be the assignments consistent with M and E. Let g(w) be the required decision in assignment w. This is standard constraint-based reasoning adapted to this task; [S2, S3] are direct prior art.
G(M,E) = { g(w) : w is in Ω(M,E) }
Conflict: Ω(M,E) is empty
Determined in scope: Ω(M,E) is nonempty and |G(M,E)| = 1
Underdetermined: |G(M,E)| > 1
The nonempty condition is essential. Contradictory premises do not earn a successful answer merely because a universal statement over an empty set is vacuously true.
A gap witness is a pair w₀,w₁ that both satisfy the exposed evidence and admitted constraints, but have different decisions. It establishes that the modeled evidence does not determine that decision. It does not establish which scenario is actual, that either scenario is likely, or that every missing condition has been found.
In natural language, a model-generated pair is only a candidate witness. It must not change an observed amount, invent a new source authority, silently change the question, or disregard a stated exception. A formal checker can validate small structured cases. For semantic cases, separate review must preserve unresolved validity. A second model's agreement is not an infallible validator.
The asymmetry matters: one valid counterexample pair can establish a gap; a bounded search that finds none does not establish completeness. Exhaustive enumeration can establish sufficiency only within its declared finite model. A model-generated universe that omits the relevant exception supplies a conditional result about the wrong universe.
Use four externally meaningful statuses: witnessed_gap, determined_within_scope, conflict, and not_established. Attach the model/scope version, evidence digest, output requirement, and checking method. An empty list of proposed missing facts maps to not_established unless a justified completeness check exists. Neither a confidence score nor a unanimous panel vote changes that rule.
For a fixed M, adding consistent observations can only shrink Ω and cannot introduce a new decision value. Changing M is different. Discovering an amendment or revising source authority can reopen the decision. Any stored sufficiency result therefore needs the same kind of version binding that Round 12 required for terminal evaluation, without borrowing its separate risk-calibration claim.
7. Proposed mechanism: challenge the decision and the scope
The proposal uses the existing evidence workflow and a bounded diagnostic step. It does not require an extra agent; one model with structured state is a first-class implementation.
- Freeze the boundary. Save the task, initial answer, source manifest, actually exposed text, output contract, and remaining budget. Separate admissible prior knowledge from source-restricted facts.
- Propose a decision-changing pair. Ask for two concise scenarios that retain all observed facts and differ only in unresolved conditions. Require the affected decision and a source or task-based reason each alternative is admissible. Hypothetical scenarios are tagged as hypotheses, never inserted into evidence as observations.
- Validate the pair. Check consistency, source applicability, units, time, and actual decision divergence. Reject pairs that achieve disagreement by rewriting the givens. Record unresolved semantic checks rather than manufacturing validity.
- Probe the scope separately. Within a fixed acquisition allowance, inspect a source catalog, exception index, schema reference, or other task-appropriate overview for relevant material absent from the current representation. The probe must not be restricted to the detector's already proposed variables. Its cost comes from the same budget.
- Acquire or qualify. A valid pair identifies a useful distinction to investigate. Ask the authorized source or inspect the relevant artifact. If no answer is available, provide the strongest supported conditional result allowed by the task. Do not keep asking the user for a fact already present in accessible evidence.
- Re-evaluate and stop honestly. Update the observed evidence, recompute the scope-bound status, and label the exact terminal answer. Budget exhaustion is an execution reason, not a statement that the evidence is sufficient.
The first implementation should use a small predeclared limit, such as two proposed pairs and one catalog probe per revision boundary. These numbers are development choices to test, not a theorem or production recommendation. A checklist receives the same total budget and may use saved work for direct reasoning or retrieval.
The source probe targets a specific weakness of the pair generator: both proposed scenarios may inherit the same omitted condition. It cannot guarantee open-world completeness either. Catalogs can be incomplete; a relevant source can be mislabeled; the generator can fail to recognize a newly exposed rule. Report the source universe actually searched.
A request generated from a witness need not be minimal or optimal. Determining the best sequence of queries is already part of the Round 11 and Round 15 agenda. This round holds query selection simple enough to isolate whether the necessary distinction was discovered at all.
8. Worked example: a true amount, an unsupported decision
Consider a fictional policy. Purchases above 10,000 units require review, except when the purchase is eligible for a research exemption and an authorization has been issued. The exposed purchase record gives an amount of 12,000. Eligibility and authorization are absent. The rule and amount are genuine within this invented fixture; this is not advice about an actual policy.
| Evidence-compatible scenario | Unobserved conditions | Required answer |
|---|---|---|
| w₀ | Eligible = false; authorized = true | Review required. |
| w₁ | Eligible = true; authorized = true | Review not required. |
Both scenarios preserve the same amount, threshold, rule, and question. Therefore a checked quote establishing the amount cannot establish the review decision. A system can correctly report that supported evidence exists while the decision remains underdetermined.
If an authorized record establishes eligibility is false, review is required regardless of authorization. Asking about authorization then adds no decision value. If eligibility is true, authorization still matters. Under a stipulated uniform distribution over the two binary conditions, a policy that queries eligibility first and authorization only when needed uses 1.5 queries on average; a fixed set guaranteeing an answer across all cases contains both variables. These are exact properties of the fixture, not estimates of user behavior or Noetis efficiency.
Now make the case harder: the initial packet contains only the general threshold paragraph, while the source catalog lists a separate exception supplement. A detector whose model excludes exceptions may find no alternative and appear certain. A catalog probe can reveal the additional rule and reopen the decision. The distinction is between a missing value in a known rule and a missing rule in the represented scope.
Matched controls are necessary. An unrelated travel supplement should not change the purchase decision. A duplicated threshold paragraph should not count as a new independent observation. A complete record establishing both exemption conditions should permit an answer. An unknown decorative field should not force a question. A plausible unofficial annotation should be assessed under the declared source hierarchy, not treated as an amendment merely because it offers a convenient answer.
9. Exact diagnostics and what they establish
The supplied standard-library implementation was executed. It enumerates every Boolean decision function on three binary variables and every partial observation state. There are 2⁸ = 256 functions and 3³ = 27 states per function: 6,912 function–state pairs. Exhaustive restriction of each truth table yields 3,778 determined pairs and 3,134 underdetermined pairs. Every underdetermined pair has a valid two-world witness.
An independent counting identity checks the determined total. With k unobserved variables, there are C(3,k)·2^(3−k) partial states. A Boolean function is constant on the corresponding 2^k rows in 2·2^(8−2^k) ways. Summing their product for k=0,…,3 gives 3,778. This checks the finite enumerator; it is not a result about the distribution of real decisions.
| Executed diagnostic | Result | Interpretation |
|---|---|---|
| Exhaustive three-variable domain | 6,912 pairs; 3,134 validated gap witnesses | The reference checker matches the stated finite semantics. |
| Supported amount with unresolved exemption | Both review and no-review remain possible | At least one supported fact does not imply decision sufficiency. |
| Four bounded probes of a three-input AND rule | All four return 0; the full eight-world domain also permits 1 | Agreement among checked alternatives does not certify completeness. |
| Expansion from a no-authorization scope to the full policy scope | Outcome set changes from {review} to {review, no review} | A result can be correct in a narrow model and inadequate for an expanded one. |
| Irrelevant unknown color | Two worlds remain, with the same required decision | State uncertainty need not require a clarification. |
| Invalid or contradictory witness cases | Changed givens and same-decision pairs rejected; empty feasible set marked conflict | Disagreement must be admissible and decision-relevant. |
| Balanced hidden binary twins | Any unchanged-observation binary answering policy averages 1/2 accuracy | More output confidence cannot reveal an unobserved bit. |
The twin calculation concerns a fixed, balanced pair of worlds with identical observations and no access to another distinguishing signal. If a randomized policy answers one with probability p, its average accuracy is (p + 1 − p)/2. It may correctly identify that the task is unresolved. The result does not forbid better performance when prior probabilities are unequal, new information is acquired, or the task allows a conditional response.
Thirty-seven named checks passed, including the exhaustive count and witness checks. Twelve evaluator fixtures were generated: six simple rule families, each with an incomplete and a complete observation state. They are developer fixtures, not an independently authored benchmark, natural-language corpus, or estimate of real-world prevalence. Their oracle tables must remain hidden from any future candidate model.
No model was asked to discover these witnesses. The checker receives an explicit finite universe; that is precisely the privilege a natural-language detector does not automatically have. The difficult empirical question remains whether candidate generation and source discovery recover useful distinctions at acceptable cost.
10. Claim and evidence table
| Claim | Evidence and status | Permitted conclusion |
|---|---|---|
| Supported evidence and sufficient evidence differ. | Current static field semantics [I7]; exact policy counterexample. | The field alone cannot certify decision completeness. No failure frequency follows. |
| Information acquisition and final answer accuracy are separable. | Constraint-based benchmarks [S2, S3]; RAG analysis [S1]. | Measure both. The distinction is established prior art. |
| Detecting ambiguity need not produce clarification. | Preprint's explicit-judgment versus QA comparison [S7]. | A detector must be evaluated together with the policy that uses it; mechanism remains uncertain. |
| More structured questioning can help in studied settings. | InfoGatherer [S4] and AskBench [S6], under their own designs. | Include structured baselines. No Noetis or equal-compute effect is established here. |
| Failed bounded search is not a completeness certificate. | Exact AND-rule counterexample and elementary logic. | Preserve not_established; do not convert zero found alternatives into zero risk. |
| A source-coverage probe improves missing-rule discovery. | Encevra hypothesis H16-2, untested. | Run the ablation; do not ship on analogy alone. |
| Brain-inspired state distinctions motivate a useful representation. | Human fMRI task-state study [S10] plus an engineering inference. | Test representation ablations; no anatomical correspondence or brain replication. |
“Established” here refers to a logical distinction or a result within the cited study's scope. It does not mean all detectors fail, all structured methods win, or a literature effect transfers to current models. “Mixed” describes the coexistence of accuracy gains, over-questioning, recognition–action gaps, and integration failures across different setups. Encevra's mechanism and effect sizes remain open hypotheses.
11. Experiment: from a valid collector to a useful detector
E16-A, executed: the finite-world diagnostics above. E16-B, proposed: a bounded missing-condition study after the existing Round 13 collection gate. Do not rename the offline diagnostics as completion of that gate.
Before model collection, verify preservation of immutable initial and terminal artifacts, source/rubric versions, every provider attempt, actual receiver exposure, and complete or explicitly partial usage. Validate deterministic labels on small closed tasks. For semantic tasks, retain independent first-pass reviews and adjudication history under Round 14. The current static inspection is insufficient to establish a working end-to-end collector.
For development, propose twelve independently authored source families with two matched instances per family, giving 24 instances. Include complete cases, missing values in visible rules, missing rules discoverable through a catalog, irrelevant unknowns, source conflicts, and fully exposed evidence that still requires nontrivial integration. Cross these with consequential exceptions and benign additions where possible. Do not manufacture every incomplete case by deleting a sentence while leaving an obvious blank.
Each instance has one frozen initial answer for the local comparison. Use two independent continuations per condition, randomized execution order, and isolated branch state. Four core conditions give 192 assigned continuations at one budget. This is a feasibility plan, not a powered superiority study or a commitment to purchase inference.
| Condition | Behavior | What it tests |
|---|---|---|
| B0: strong direct reasoning | Tuned single agent can inspect sources, reconsider, ask, or answer. | Whether any extra gap structure is needed. |
| B1: explicit checklist | Lists decision prerequisites and judges what is missing; no required scenario pair. | A simple structured baseline with the same capabilities. |
| B2: validated contrastive probe | Proposes and validates bounded decision-changing pairs, then uses the ordinary acquisition interface. | Incremental value of the witness requirement over B1. |
| B3: probe plus source coverage | B2 plus a bounded catalog/exception scan charged to its budget. | Additional value when the relevant condition is absent from the representation. |
The primary comparison is B2 versus B1. B3 versus B2 is a prespecified secondary contrast, especially on missing-rule cases. A catalog-only version of B1 is the corresponding ablation: without it, B3's advantage could be explained entirely by the extra source scan. Give all methods the same admissible corpus and tools; the candidate receives no evaluator-only list of missing conditions.
Add an information-delivery control on a separate diagnostic subset: provide the complete admissible packet to the same answerer, with no gap detector. If this rescues performance, acquisition or exposure is implicated; if it does not, integration remains a candidate bottleneck. This is a controlled diagnostic, not a deployment policy with free oracle retrieval. Source selection and context costs belong in a later deployable comparison.
A subsequent complete-system comparison must include the instrumented current Noetis workflow, a strong single agent, and a single agent using the same proposed state. Only that comparison can support a claim about team architecture. A local detector effect cannot establish that another specialist should be added.
12. Control total work, not just the number of turns
Use one pinned backbone and reasoning configuration for the primary mechanism comparison. Freeze prompts, context limits, source tools, and the evaluator before the held-out phase. Equal maximum turns alone is inadequate: a graph builder, validator, or critic may spend several calls inside one turn.
For planning, use one initial ceiling of 30,000 total reported model tokens per assigned episode, counting input and output across all calls, with reported reasoning tokens included exactly once under the provider's accounting convention. Permit at most four source-access operations under a common response-size limit. Reserve enough remaining budget for terminal composition before starting another diagnostic call. Every detector, invalid pair, validator, retrieval query, retry, and failed attempt counts. These ceilings are design choices to validate in the smoke test, not measured requirements.
Charge the frozen initial-answer cost to every hypothetical policy that requires it, while reporting the actual research bill separately when the initial artifact is physically reused. At 192 assigned continuations, a 30,000-token per-policy ceiling implies at most 5.76 million charged policy tokens under complete accounting; it is not a forecast or a provider price quote. Shared-prefix reuse changes the actual invoice, not the counterfactual system cost being compared.
Input plus output tokens are an observable budget proxy, not an exact FLOP identity. Keep the backbone and configuration fixed, record latency and actual provider charges, and disclose cached input accounting and hidden or missing usage. If material usage is unavailable, a strong matched-total-compute claim is blocked. Equal caps do not imply equal realized spending.
Permit simpler baselines to spend the saved detector budget on their own best feasible reasoning or source access. Tune all methods on separate development families with the same tuning allowance. Do not pad outputs to consume unused budget. After feasibility, compare quality–cost curves at a second predeclared ceiling, rather than choosing the budget where the candidate happens to win.
Independent evaluation has its own fixed effort budget. Count witness validation, whole-answer review, disagreements, and re-audits. A method that needs substantially more expert interpretation must report that cost even if its generation budget matches.
13. Endpoints that expose regressions
The primary endpoint is supported task completion over all assigned instances: the exact terminal response satisfies the frozen output contract, is correct under the reference, and does not commit beyond the admitted evidence. For a single-action contract, an unqualified answer while multiple actions remain possible is unsupported even if it guesses the hidden outcome. For a conditional-answer contract, an accurate, complete conditional response can count as completion. Report these task types separately.
| Metric | Definition | Failure it catches |
|---|---|---|
| Unsupported commitment | Definitive answers issued while the reference still permits materially different required answers | Lucky guesses and confident closure. |
| Useful completion | Supported completion divided by all assigned instances | Refusal or endless questioning that merely avoids errors. |
| Gap detection sensitivity and false alarms | Valid detected gaps among reference gaps; gap flags among sufficient controls | Missing the decisive condition versus inventing a gap. |
| Witness validity | Consistent, decision-changing pairs among all proposed pairs | Altered givens, fabricated authority, and theatrical disagreement. |
| Acquisition and exposure | Needed evidence found, delivered, and present in the actual receiver input | Confusing successful retrieval with successful communication. |
| Integration after sufficient exposure | Correct terminal responses on reference-sufficient trajectories | Continuing to blame missing context for reasoning errors. |
| Repair and damage | Wrong-to-correct and correct-to-wrong changes, with their eligible initial denominators | More corrections masking collateral harm. |
| Needless acquisition | Questions or reads after the decision is determined, under the fixed contract | Asking for every unknown instead of decision-relevant information. |
| Cost, delay, and failures | Total work, median and tail latency, timeouts, malformed outputs, and missing usage | Gains purchased through unreported overhead or exclusions. |
Track the full sequence from corpus availability to terminal retention. A pair can be valid and the resulting question useless if it asks the wrong source. A source can be read without its decisive passage reaching the final answer. Success at one boundary is not success at every boundary.
Do not score witness presence as correctness. A candidate could maximize that metric by declaring increasingly implausible alternatives. Also do not score every request for clarification as beneficial. The complete and irrelevant-unknown controls are essential to detect a detector that avoids committing by becoming unhelpful.
For unresolved semantic labels, preserve the allowed joint outcomes for the initial and terminal artifacts. Distinguish an underdetermined task, a failed generation, and a missing evaluation. Report identification bounds where needed, and do not drop inconvenient cases. The initial artifact is shared across conditions; copying its label does not create independent observations.
14. Hypotheses, inference, and rejection criteria
H16-1 — Validated alternatives improve discovery. B2 improves supported task completion and reduces unsupported commitment relative to B1 on held-out source families at the same budget. The hypothesis is weakened if witness generation rarely finds valid distinctions, adds harm, or loses to simple checklist review after its cost is counted.
H16-2 — Scope discovery addresses a different bottleneck. B3's incremental benefit over B2 is larger when a relevant rule is absent from the represented schema than when only a named value is missing. Compare this interaction with the catalog-only checklist ablation. If the simple catalog scan accounts for the effect, credit that scan and simplify the proposed mechanism.
H16-3 — Decision relevance prevents needless questioning. The probe preserves useful completion on sufficient and irrelevant-unknown controls while reducing unsupported closure on genuinely underdetermined cases. It fails if lower unsupported commitment comes mainly from broad refusal, invented conditions, or repeated requests for already available facts.
The 24-instance pilot should report raw paired outcomes and measurement feasibility, not claim small-effect confirmation. Average repeats within task-condition before family aggregation. Sample and resample source families for a future-family claim; repeated continuations, closely related variants, and the 6,912 mathematical cases are not additional independent families.
For later confirmation, provisional practical targets are a five-percentage-point gain in supported completion over B1, at least a five-point reduction in unsupported commitment on the incomplete stratum, and no more than two points of additional damage on initially correct answers. These are planning margins, not literature predictions or safety guarantees. Freeze one primary budget, a joint decision rule, and sample size from developmental family-level variability before using new families. Require confidence bounds to support both benefit and the chosen harm margin; a favorable point estimate alone is insufficient. Prespecify treatment of the secondary interaction and multiplicity.
A small inconclusive study does not show equivalence. A precisely estimated null effect, or a simpler method that wins at lower cost, is useful evidence against added complexity. Preserve these results in the hypothesis register instead of replacing the benchmark with a friendlier one.
15. Brain-inspired hypothesis and its alternative
Schuck and colleagues studied human task states requiring information beyond the current visual stimulus. Their fMRI analyses found decodable hidden, task-relevant state information in orbitofrontal cortex. The study motivates representing decision context, but it does not demonstrate autonomous discovery of unknown rules or a software implementation. Its representational findings are not an intervention proving that our proposed record causes better reasoning. [S10]
The engineering analogy is state separation: two situations that look alike in the currently attended evidence can require different decisions because a contextual condition differs. Test an explicit record of the unresolved condition and its source/version against an equal-budget flat summary containing the same facts. Then remove only the condition-to-decision link while preserving the text. Measure missing-condition recovery and collateral errors.
A benefit could come from ordinary retrieval cues or better formatting. A single agent with the same structured record could match a team. Either result should narrow the claim to the useful computational function. Naming a component after a brain area adds no evidence, and no result here supports literal brain copying.
16. Failure modes, limitations, and next handoff
The generator shares the original blind spot. Both scenarios can omit the same exception. A source probe supplies another route but can share the same retrieval bias. Held-out rule families and a catalog-only baseline are more informative than merely increasing the number of critics.
The generator invents a gap. An alternative may contradict the record, introduce an unauthorized exception, or broaden the task beyond the user's intent. Preserve the source of each assumption; validate before treating it as an acquisition obligation.
The validator becomes the bottleneck. Natural-language compatibility and material decision change can be ambiguous. A model judge may prefer elaborate witnesses. Use executable anchors, blinded review, and the unresolved-label discipline from Round 14.
A limited universe looks complete. Exhaustive search through a wrong model can be perfectly implemented and scientifically irrelevant. Keep scope completeness separate from within-scope determinacy. Neither a hash nor a large search count certifies an omitted rule.
A good probe is too expensive. Candidate generation, checking, and context expansion can consume the budget needed to solve the task. The strong direct-reasoning baseline must be allowed to use that budget productively.
The task rewards guessing. A benchmark judged only against a hidden answer can reward an unsupported correct guess. Conversely, a source-restricted metric can unfairly penalize legitimate prior knowledge unless the restriction was part of the task. Freeze both the evidence contract and the required output before collection.
The case design favors the mechanism. Repeated threshold-and-exception templates can turn discovery into pattern matching. The twelve generated fixtures are development tools, not a held-out evaluation set. New families must include benign unknowns, alternative valid solutions, conflicts, and integration failures.
This focused review is current through 7 October 2026. It is not a systematic survey; primary sources were inspected at the depths listed below, and their full experiment suites were not reproduced. Several important sources are preprints. The current-code finding is confined to the inspected snapshot and path. No new Encevra model benchmark was run, and no calibrated release guarantee or measured multi-agent advantage follows.
Next handoff: verify the existing collector on a small closed task, preserve a source-to-exposure record, and establish independent sufficiency labels before comparing detectors. Use the twelve development fixtures to check the mechanics, then author fresh families for the pilot. Run the smallest valid study, report failures and actual costs, and decide whether a checklist or direct source delivery already solves the problem. Round 17 can investigate preservation of decisive conditions under compression, as Round 15 proposed, but an observed collection or discovery failure should take priority over another speculative architecture layer.
17. Full sources and inspection record
All linked primary PDFs below were retrieved and inspected on 7 October 2026. Dates identify the publication or inspected edition, not a crawl timestamp. “Inspected” specifies the sections actually read; it does not claim that all appendices, proofs, or reported experiments were independently reproduced. Machine-readable records include retrieved-file hashes. External papers and private project source files are not redistributed in the research package.
[S1] Hailey Joren, Jianyi Zhang, Chun-Sung Ferng, Da-Cheng Juan, Ankur Taly, and Cyrus Rashtchian. Sufficient Context: A New Lens on Retrieval Augmented Generation Systems. ICLR 2025; inspected arXiv:2411.06037v3, 23 April 2025; first submitted 9 November 2024. Read definition and autorater in Section 3, dataset setup and error analysis in Section 4, and opening selective-generation discussion. Supports separating context sufficiency, correctness, and model use; its operational definition is not identical to our fixed-action determinacy criterion. Record · Inspected PDF.
[S2] Belinda Z. Li, Been Kim, and Zi Wang. QuestBench: Can LLMs Ask the Right Question to Acquire Information in Reasoning Tasks? NeurIPS 2025, Datasets and Benchmarks Track; first preprint 28 March 2025. Inspected conference PDF pages 1–3, including the distinction between parsing ambiguity and missing assignments and the sufficient-set formulation. Supports the formal predecessor and benchmark-design comparison; no reported model score is imported as an Encevra result. Conference record · Inspected PDF.
[S3] Yepeng Huang, Jiawen Zhang, Michelle Dai, Xiaorui Su, Shanghua Gao, Zi Wang, and Marinka Zitnik. Do LLMs Know What to Ask and When? Evaluating Multi-Turn Information Seeking. arXiv:2608.14808v1, 14 August 2026, preprint. Inspected PDF pages 2–9: framework, problem-versus-instance distinctions, dataset/oracle setup, and main results, including fixed-trace answer transfer. Supports separate acquisition and integration measurement; some domains use model-based oracles and the full appendices were not audited. Record · Inspected PDF.
[S4] Maksym Taranukhin, Shuyue Stella Li, Evangelos Milios, Geoff Pleiss, Yulia Tsvetkov, and Vered Shwartz. InfoGatherer: Principled Information Seeking via Evidence Retrieval and Strategic Questioning. arXiv:2603.05909v1, 6 March 2026, under-review preprint. Inspected Section 2 in HTML and PDF pages 4–8: elicited belief masses, question and stopping rules, simulation, baselines, and ablations. Supports a substantive structured competitor. The reviewed comparisons emphasize success and dialogue turns; they do not establish equal-total-inference superiority. Record · Inspected PDF.
[S5] Tong Zhang and colleagues. CLAMBER: A Benchmark of Identifying and Clarifying Ambiguous Information Needs in Large Language Models. ACL 11–16 August 2024, pages 10746–10766. Inspected PDF pages 1, 3, and 5: taxonomy, data construction/validation, and task setup. Used for separating kinds of ambiguity and evaluating question generation, not for claiming a current-model error rate. Record · Inspected PDF.
[S6] Jiale Zhao, Ke Fang, and Lu Cheng. When and What to Ask: AskBench and Rubric-Guided RLVR for LLM Clarification. Findings of ACL, 2–7 July 2026, pages 17120–17140. Inspected PDF pages 1–4: missing-information and false-premise settings, checkpoint construction, judge/user-simulator loop, and coverage/redundant-question metrics. Supports explicit endpoints; the combined simulator/judge requires independent scrutiny. Training results were not reproduced. Record · Inspected PDF.
[S7] Jinyan Su and Claire Cardie. Knowing but Not Showing: LLMs Recognize Ambiguity but Rarely Ask Clarifying Questions. arXiv:2605.25284v1, 24 May 2026, preprint. Inspected study setup and PDF pages 4–6: sampled AmbigNQ items, explicit judgment versus QA framing, and model-judged behavior. Supports distinguishing recognition from action; the QA prompt requests answers and the behavioral labels are fallible. Record · Inspected PDF.
[S8] Leon Engländer, Sophia Althammer, Ahmet Üstün, Matthias Gallé, and Tom Sherborne. Agents Explore but Agents Ignore: LLMs Lack Environmental Curiosity. arXiv:2604.17609v1, 19 April 2026, preprint. Inspected PDF pages 2–5: solution-injection design, discovery/interaction metrics, and experimental setup. Supports observing encounter and use separately. Complete injected solutions are a capability probe, not a model of trustworthy natural evidence; no recommendation to execute untrusted material follows. Record · Inspected PDF.
[S9] Pranav Rajpurkar, Robin Jia, and Percy Liang. Know What You Don't Know: Unanswerable Questions for SQuAD. ACL 15–20 July 2018, pages 784–789. Inspected PDF pages 1–2, including relevance and plausible-answer construction criteria. Supports difficult negative controls rather than obvious missing-text markers; this is extractive reading comprehension. Record · Inspected PDF.
[S10] Nicolas W. Schuck, Ming Bo Cai, Robert C. Wilson, and Yael Niv. Human Orbitofrontal Cortex Represents a Cognitive Map of State Space. Neuron 91:1402–1412, 21 September 2016, DOI 10.1016/j.neuron.2016.08.019. Inspected author-hosted PDF pages 2–4 and 8: hidden-state task, decoding analysis, and limitations/discussion. Supports a functional state-representation analogy, not software efficacy or an anatomical copying claim. Primary record · Inspected PDF.
Internal continuity and audit records. [I1] Round 9, 30 September 2026, full agenda. [I2] Round 10, 1 October 2026, full working paper. [I3] Round 11 and Round 12, allocation and terminal calibration; carried-forward claims, not newly replicated results. [I4] Round 13, 4 October 2026, audit and collection handoff. [I5] Round 14, 5 October 2026, public paper plus separately inspected internal expanded edition 1.1. [I6] Round 15 public summary, 6 October 2026, plus separately inspected complete internal manuscript and handoff. [I7] Current project snapshot identified in Section 1: overview, index, registry, relevant Rounds 3, 4–6 and 8, and the inference file's evidence-presence and returned-artifact paths. Archive canonical links identify the intended pages; project source was the authority inspected because preview changes can precede deployment. Internal notebooks and implementation files remain private.