Encevra Research Round 26

What Can the Watchers Ever Know? Minimum Observables for Detecting Omitted Change

Program: Encevra Research
Date: 2026-10-17
Research status: One finite-model identifiability analysis executed; one prospective shadow experiment proposed
Protocol status: Observation Sufficiency Certificate 0.1, proposed only
Architecture impact: None
Production impact: None

Result boundary

Round 26 asks when a collaborative AI system can know that a quiet evidence window really means no relevant change. It proves a general, model-independent limit: if an admissible changed world and an admissible unchanged world produce exactly the same observations, no decision rule over those observations can be correct in both worlds. The only remedies are to add an observable that separates the worlds, exclude one world with an explicit trust assumption, or weaken the claim.

The round then executes E26-A, a deterministic finite-state analysis over seven possible failure roots and eight observable types. The analysis enumerates every fault set up to budgets of zero, one, two, and three failed roots, tests every observable subset, lists all opposite-target worlds that remain observationally identical, and finds inclusion-minimal and minimum-cost separating sets. The declared state space contains 2, 16, 58, and 128 worlds at those budgets, respectively. Eight executable tests passed.

Within this declared model, one target-sensitive channel is sufficient with no faults, two root-disjoint target-sensitive channels are necessary and sufficient against one arbitrary silent root failure, three are necessary and sufficient against two, and four are necessary and sufficient against three when the fourth witness is genuinely independent. When the nominal witness is allowed to share the issuer's root, no available set identifies change against three failed roots. This is an exact result for the finite model and a direct instance of the root-disjointness theorem in Section 7. It is not an estimate of real incident frequency or AI performance.

The Round 25 SACP observable set still has eight opposite-target ambiguity pairs under the one-fault universe. Adding one independently rooted external commitment reduces that count to zero under the same universe. Under two faults, SACP plus one external commitment leaves 26 ambiguity pairs; adding an independently rooted behavior probe reduces the count to zero. These counts weight every enumerated fault set equally and must not be read as prevalence-weighted rates.

No production incident was sampled. No live repository webhook, deployment control plane, database, provider account, runtime, user traffic, or customer data was queried or changed. No language model was benchmarked. No new Encevra answer benchmark was run. The proposed certificate was not installed. No release gate, claim state, retention policy, or production behavior changed.

Abstract

Round 25 showed that liveness challenges, sequence commitments, and self-monitoring can detect many ways in which missing change evidence masquerades as stability. It also recorded five designed misses: selective non-emission without an external commitment and verifier common-mode failure without an independent challenge path. Its prescribed fallback, when a real multi-authority shadow trial with hidden adjudication is unavailable, was a formal indistinguishability analysis rather than a larger authored fixture suite.

Round 26 takes that fallback. Its research question is: For each omitted-change failure class, which unchanged and changed worlds remain observationally identical under a given observer set, and what is the minimum additional independently rooted observable or trust assumption needed to separate them?

The central object is an observation function. A world contains the target fact, such as whether a relevant authority change occurred, plus latent fault state. A selected observer set maps that world to an observable vector. Two worlds are indistinguishable when the vectors are equal. The target is identifiable only when every observational equivalence class contains one target value. This makes the boundary operational: more messages, agents, votes, or checks help only if they split an ambiguity class that crosses the target boundary.

Primary distributed-systems results establish why guarantees depend on timing, authentication, failure, and knowledge assumptions [S05-S10]. Diagnosability research asks whether faults can be inferred from observable event sequences and extends the idea to active diagnostic actions [S09, S11]. PeerReview, Certificate Transparency, and TUF show how signed records, witnesses, append-only commitments, versions, and expiry can expose some faults while retaining explicit trust limits [S12-S14]. Mixed multi-agent language-model evidence shows both benchmark gains from debate and systematic cases in which discussion propagates errors or verification fails [S15-S17]. Together these sources support the need to reason about information structure before spending more computation on collaboration. They do not establish the proposed Encevra mechanism.

Round 26 proposes the Observation Sufficiency Certificate (OSC). Before a system certifies NO_RELEVANT_CHANGE, the OSC names the target predicate, threat model, failure roots, selected observables, root-lineage assumptions, separated world pairs, residual indistinguishable pairs, verifier trusted computing base, cost, and expiry. If any residual equivalence class contains both changed and unchanged worlds, the only admissible result is UNKNOWN_WITH_RESIDUAL_AMBIGUITY, not stability.

The useful brain-inspired idea is limited to active sensing. Biological active sensing is described as a closed loop in which an observer estimates task-relevant state and a planner chooses actions expected to obtain informative observations [S18]. Encevra's testable analogue is to choose the next probe by expected ambiguity reduction per unit cost. This is not a claim that the protocol copies a brain. A fixed-budget experiment can reject the analogy if adaptive probes do not reduce residual ambiguity or harmful false-stability decisions relative to fixed probes.

1. Archive continuity and corrected round number

The automation brief retained the historical label Round 11. Before topic selection, the live Lovable project was inspected at commit f1197fa663970737c3da35bc8e98c63c60f59c78. The Research overview, rounds index, registry, full Round 25 and Round 24 manuscripts, and the relevant Round 10 and Round 9 routes show completed work through Round 25. The correct next sequential number is therefore Round 26 [S01-S04].

The carried program state is:

Round 25 preferred a preregistered E25-B shadow window over real multi-authority observations with hidden independent adjudication. That access was not available in this run. Round 25 explicitly prescribed the present fallback: enumerate omission classes that remain undetectable under each observer set and prove the minimum new trust assumption needed to separate them [S02].

2. Research question, novelty, scope, and exclusions

2.1 Primary question

For each omitted-change failure class, which changed and unchanged worlds produce the same observable trace under the selected observer set, and what is the minimum new observable, independence condition, timing assumption, or claim restriction needed to distinguish them?

The decision target is deliberately narrow:

T(w) = 1 if a relevant authority change occurred in the declared window and claim scope; otherwise T(w) = 0.

Round 26 does not try to identify every root cause. It first asks the simpler and logically prior question: can the target fact be identified at all from the evidence available to the system?

2.2 Novelty against the archive

Round 24 asks where received evidence originated. Round 25 asks whether configured observation paths were live, fresh, continuous, and reconciled. Round 26 asks whether those observations contain enough information to distinguish the target worlds even if every implemented check behaves exactly as specified.

This is not another evidence-exchange protocol, selective-revision policy, or procedural-learning method. The new object is the equivalence class induced by an observer set. It provides a stopping rule for verification effort: a check has zero value against a candidate failure when it produces the same output in both members of the unresolved world pair.

Three adjacent directions were considered and deferred:

2.3 In scope

• Binary target identifiability for a declared authority, claim, and time window.
• Silent issuer omission, transport loss, observer failure, stale or false-clean checkpoints, external-channel failure, probe blind spots, common-origin witnesses, and verifier failure.
• Exact observational equivalence, minimal separating sets, root-disjointness, and claim narrowing.
• A finite root-fault model with transparent assumptions and executable enumeration.
• A proposed OSC data structure and three-way decision rule.
• A prospective fixed-budget experiment with natural and sandbox strata kept separate.

2.4 Exclusions

• Estimating the prevalence or probability of any fault.
• Proving the correctness of an authority, observer, cryptographic primitive, clock, network, or verifier.
• Continuous-time behavior, arbitrary Byzantine strategies, or every possible interleaving.
• Product answer accuracy, calibration, safety, latency, user trust, or economic value.
• Deploying a monitor or changing a claim's production status.
• Literal brain copying.

3. Definitions and evidence semantics

Three levels of claim must remain separate:

1. NO_RECEIVED_CHANGE: the selected receipt stream contains no relevant change.
2. NO_RECORDED_CHANGE: a named authority commitment explicitly closes the window as clean.
3. NO_ACTUAL_RELEVANT_CHANGE: the target authority did not undergo a covered change.

The first can be true when transport fails. The second can be true when the issuer selectively omits. The third requires assumptions or independent observations that rule out both. OSC prevents a system from silently upgrading one level into another.

4. Literature map and competing positions

4.1 Impossibility results are conditional, not pessimistic slogans

Lamport, Shostak, and Pease show that Byzantine agreement depends on the number and behavior of faulty participants and on whether messages are authenticated [S05]. Fischer, Lynch, and Paterson show a nontermination possibility for deterministic consensus in a fully asynchronous model with one crash fault [S06]. Dwork, Lynch, and Stockmeyer recover solvability under declared partial-synchrony assumptions [S07].

The shared lesson is not that reliable collaboration is impossible. It is that a guarantee is meaningful only relative to a model. A Round 26 certificate must therefore name timing, fault, authentication, and independence assumptions rather than bury them in implementation prose.

4.2 Knowledge is defined by possible worlds that look the same

Halpern and Moses formalize knowledge in distributed systems using points that an agent or group cannot distinguish. They show that strong common knowledge is unattainable in many practical systems with temporal imprecision, while weaker variants can support useful coordination [S08].

Round 26 uses the indistinguishability construction, not the full epistemic logic. Its practical consequence is similar: when both a changed and unchanged world remain possible under the same view, the system does not know which target value holds. The appropriate response may be a weaker statement, such as no recorded change through checkpoint q, rather than an unsupported global stability claim.

4.3 Diagnosability asks whether faults leave distinguishable observations

Sampath and colleagues define diagnosability for discrete-event systems in terms of whether fault occurrences can be inferred from observable behavior within a bound [S09]. Their later active-diagnosis work studies control actions that make otherwise ambiguous faults distinguishable [S11].

This is the closest formal precedent for Round 26. The mapping is not exact: Encevra's target is a claim-relevant change, its observers can include AI agents and external services, and the finite model is not a general automaton over unbounded traces. Still, diagnosability supplies the right question: what information is present in the projection that the monitor can see?

4.4 Failure detectors add information with explicit accuracy and completeness properties

Chandra and Toueg model failure detectors as potentially incorrect information about failures and classify them by completeness and accuracy [S10]. The detector is an oracle abstraction with stated properties, not proof that a timeout reveals the underlying cause.

Round 26 similarly treats heartbeats, acknowledgments, and canaries as health information. They can turn false certainty into UNKNOWN, which is valuable. They do not by themselves make a hidden semantic change identifiable.

4.5 Accountability and transparency cover observed or committed actions

PeerReview maintains signed records and uses correct witnesses to expose deviations that are observed by a correct node [S12]. RFC 9162 uses append-only logs, signed timestamps, inclusion proofs, consistency proofs, and monitoring to make recorded certificate activity auditable [S13]. TUF uses signed roles, monotonic versions, expiry, and threshold rules to expose rollback and freeze under a declared threat model [S14].

These mechanisms support checkpoints, root separation, freshness, and post hoc accountability. Their stated limits matter just as much. A correct log of incomplete input is still incomplete. A monitor must exist and observe the commitment. A set of witnesses is only as independent as its actual failure roots.

4.6 More agent discussion can help or harm

Du and colleagues report benchmark improvements from multi-agent debate across selected reasoning and factuality tasks [S15]. Wynn, Satija, and Hadfield report settings in which debate decreases accuracy and causes stronger agents to revise correct answers toward incorrect peers [S16]. Cemri and colleagues identify multi-agent failures including information withholding and incomplete or incorrect verification in the systems they studied [S17].

The evidence is mixed, not contradictory. Interaction can expose useful information when agents possess or elicit different evidence. It can amplify error when agents share a blind source, accept persuasive mistakes, or fail to verify. Round 26 predicts that agent count alone should not improve an omission guarantee when all agents inhabit the same observational equivalence class.

4.7 Active sensing offers a limited, testable analogy

Yang, Wolpert, and Lengyel describe active sensing as a closed loop between an observer that estimates task-relevant state and a planner that chooses actions expected to obtain informative sensory input [S18].

Round 26 borrows one design hypothesis: choose probes based on their expected ability to split the current ambiguity class, not on a fixed schedule alone. The analogy is rejected if this rule fails under matched observation cost. No claim is made that an AI assurance system implements neural computation.

5. Claim and evidence map

6. Formal model

6.1 Worlds, targets, and observations

Let W be the admissible world set. A world includes the target fact, latent authority state, and fault state. Let T: W -> {0,1} be the target predicate. Let S be a selected set of observables. The observation function is:

O_S: W -> Y_S

where Y_S is the Cartesian product of each selected observable's categorical outputs.

Define an equivalence relation:

w ~_S w' if and only if O_S(w) = O_S(w').

The target is identifiable under S when every equivalence class is target-pure:

For all w,w' in W, if w ~_S w', then T(w) = T(w').

6.2 Separation matrix

For every admissible pair p = (w0,w1) with T(w0) != T(w1), define:

M[p,s] = 1 when observable s produces different values in the two worlds, and 0 otherwise.

An observer set identifies the target exactly when it hits every opposite-target pair:

For every p, at least one selected s has M[p,s] = 1.

With sensor costs c_s, choosing a least-cost identifying set is a weighted hitting-set problem. Round 26 solves the tiny instance by exhaustive enumeration so that no optimizer assumptions obscure the result.

6.3 Decision semantics

OSC uses three outputs:

• CHANGE: at least one valid selected channel reports a covered change.
• NO_RELEVANT_CHANGE: all target-crossing pairs are separated under the declared model and the observed vector lies only in target-zero worlds.
• UNKNOWN_WITH_RESIDUAL_AMBIGUITY: the current equivalence class contains both target values, a required observation is missing, or a certificate assumption is not met.

Abstention is not evidence of a change. It preserves the distinction between epistemic uncertainty and contradiction.

7. Propositions and proofs

Proposition 1: observational indistinguishability limit

Let w0 and w1 be admissible worlds such that T(w0) != T(w1) and O_S(w0) = O_S(w1). No deterministic decision function d over O_S can return the correct target value in both worlds.

Proof. Because the observable vectors are equal, the function receives the same input in both worlds. A function returns the same output for the same input. The target values differ, so that single output is wrong in at least one world. QED.

A randomized rule does not escape the guarantee boundary. It induces the same output distribution for the same observable vector, so it cannot assign probability one to different correct values in the two worlds. Randomization may trade error between worlds; it does not identify the hidden fact.

Proposition 2: root-disjoint lower bound

Assume each target-sensitive channel has one declared failure root. A failed root may force every channel under that root to emit exactly its quiet-world value. Under up to f arbitrary failed roots, zero-error identification of a change requires at least f + 1 target-sensitive root-disjoint channels.

Necessity. With at most f roots, an adversary can fail every available root and make all selected channels emit quiet values. The changed world is then indistinguishable from a quiet world. Proposition 1 applies.

Sufficiency under the premise. With f + 1 root-disjoint channels and at most f failed roots, at least one channel survives. If every surviving channel truthfully reports the target, a changed world produces at least one change observation, while a quiet world does not. QED.

The sufficiency statement is intentionally conditional. If a surviving channel lacks coverage, shares an undeclared dependency, or has an incorrect verifier, it is not a surviving truthful root under the premise.

Proposition 3: health-only non-identification

Suppose health observable h has the same value in a changed and unchanged world whenever their fault state is the same. Adding only h cannot split that same-fault opposite-target pair.

Proof. The pair has equal prior observations and equal h. Their extended vectors remain equal. QED.

Heartbeats, queue acknowledgments, and canaries remain useful. They can prevent a missing path from being interpreted as a clean close, turning NO_RELEVANT_CHANGE into UNKNOWN. They cannot show that a semantically hidden change occurred when their health state is normal.

Proposition 4: three exhaustive ways to repair an ambiguous claim

For an ambiguous opposite-target pair, a sound system must do at least one of the following:

1. add an observable with different outputs in the pair;
2. add a justified assumption that removes one member from the admissible world set;
3. weaken the target so both worlds have the same target value.

These are exhaustive because identifiability depends only on the world set, target function, and observation function. Changing none of them leaves the pair unchanged.

8. Diagnostic E26-A

8.1 Purpose

E26-A checks the internal consequences of the formal definitions and exposes the smallest failure cuts for named observable packages. It is a finite design-space analysis, not a production benchmark and not an authored sequence-fixture expansion.

8.2 Root model

Seven binary failure roots are enumerated:

This is a root model, not a component count. Receipt and checkpoint share several primary roots. External commitment, probe, and witness are independent only in the declared independent variant.

8.3 Observable semantics

Eight categorical observables are implemented:

Silent failure is intentionally adversarial: it produces the same value as a healthy quiet world. Transport, observer, and verifier health signals can cause UNKNOWN, but a silent issuer omission keeps health green.

8.4 Enumeration

For each maximum fault budget f, the script enumerates every subset of the seven roots with cardinality at most f, then crosses each set with both target values. The resulting world counts are:

For each of the 255 nonempty observable subsets, the analysis builds the exact observation vector for every world, lists every changed-quiet pair with equal vectors, finds inclusion-minimal identifying sets, and calculates minimum sets under a transparent illustrative cost model. Primary sensors cost one unit; separately rooted external, probe, and witness channels cost three units each.

The cost units are not dollars, tokens, or latency. They are a sensitivity device. The structural zero-versus-nonzero ambiguity result does not depend on them.

8.5 Named package results

The following table uses the independent-witness variant. Counts are exact for the enumerated worlds and unweighted by real-world probability.

SACP improves on receipt absence by converting transport, observer, and verifier failures into UNKNOWN. Its remaining one-fault false-clean case is issuer omission, exactly the residual class identified by Round 25. One external root closes that class under the one-fault assumption. A second external root is needed under two arbitrary failures.

8.6 Minimum identifying sets

The exhaustive output contains every inclusion-minimal set, not only the examples in this table.

8.7 Root-lineage ablation

With SACP plus a witness and a one-fault budget:

The code and message count are unchanged. Only the dependency graph changes. This is the clearest demonstration that nominal diversity is not epistemic diversity.

8.8 Minimal blinding cuts

• SACP has size-one structural cuts at issuer, transport, observer, and verifier. The latter three may produce UNKNOWN rather than false-clean when their health signals are included.
• SACP plus external commitment has size-two cuts pairing the external root with issuer, transport, observer, or verifier.
• SACP plus external commitment and behavior probe has analogous size-three cuts.

The distinction between a structural cut and a harmful false-clean decision matters. A visible health failure can leave the target unidentified while still preventing an unsafe certificate.

8.9 Verification and reproducibility

Eight deterministic unit tests passed. They cover the Round 25 issuer-omission pair, one- and two-root lower bounds, witness correlation, health-only insufficiency, named package guarantees, and output generation. The bundle includes the executable model, tests, complete ambiguity pairs, minimal sets, blinding cuts, package metrics, source inspection ledger, and SHA-256 manifests.

8.10 Evidence boundary

E26-A establishes logical consequences of its own semantics. It does not establish that seven roots are exhaustive, that each declared root is independent in a deployed system, that every failure produces the modeled output, or that any fault set is likely. Its exactness is internal, not ecological.

9. Minimum additional requirement by omission class

10. Observation Sufficiency Certificate

10.1 Purpose

The OSC is a proof obligation attached to a negative change claim. It does not prove the world is safe. It records why the selected observations are sufficient under a specific threat model and exposes what remains unknown.

10.2 Proposed schema

osc_version: 0.1-development
certificate_id: osc:<claim>:<authority>:<window>
target:
  predicate: relevant_change_occurred
  claim_id: C21-CUR-TERMINAL
  authority: provider_contract
  window: [start, end]
  relevance_rule_digest: sha256:...
threat_model:
  maximum_failed_roots: 1
  timing: checkpoint_due_within_15m
  excluded_worlds:
    - reason
roots:
  primary:
    dependencies: [issuer, transport, observer, verifier]
  external_policy_log:
    dependencies: [external_signer, external_transport, verifier_2]
observables:
  - id: primary_checkpoint
    root: primary
    outcome: clean
    freshness: pass
  - id: external_commitment
    root: external_policy_log
    outcome: clean
    freshness: pass
separation:
  candidate_pair_count: 8
  unresolved_pair_count: 0
  matrix_digest: sha256:...
verifier_tcb:
  evaluator: osc-checker-v0.1
  raw_evidence_digest: sha256:...
decision: NO_RELEVANT_CHANGE
expires_at: timestamp
limitations:
  - no guarantee beyond one failed root
  - no guarantee outside the declared relevance rule

10.3 Protocol

1. Define the target predicate, authority, relevance rule, and time window before observing the result.
2. Enumerate or reference the admissible opposite-target world pairs under the threat model.
3. Register each observable with outputs, coverage, failure dependencies, freshness, and root lineage.
4. Build the separation matrix and select a cost-bounded set that hits every target-crossing pair.
5. Collect signed or reproducibly addressable raw observations.
6. Evaluate health, freshness, root-lineage, and target-sensitive outcomes without silently treating missing as clean.
7. Independently verify the certificate when verifier common mode is in scope.
8. Emit NO_RELEVANT_CHANGE only if the realized observation class is target-pure. Otherwise emit CHANGE or UNKNOWN_WITH_RESIDUAL_AMBIGUITY.
9. Expire the certificate when the window, threat model, root map, relevance rule, or verifier changes.

10.4 What the certificate must not do

• It must not count agents, model families, or service names as independent roots without a dependency graph.
• It must not convert a heartbeat into evidence of semantic completeness.
• It must not treat a cryptographic signature as proof that the signed content is complete or true.
• It must not omit residual ambiguity to simplify the user interface.
• It must not reuse a certificate outside its claim, authority, window, or fault budget.

11. Worked example

Suppose Encevra maintains the claim: No relevant provider-policy change occurred during window W.

The primary observer has a signed checkpoint. Two research agents independently summarize the provider documentation, but both retrieve through the same cached feed. The observed vector is:

Worlds A and B are indistinguishable. The agents may use different prompts or models, but their inputs share the omission. Majority vote, debate, longer reasoning, and higher confidence cannot separate the worlds. The correct OSC decision is UNKNOWN_WITH_RESIDUAL_AMBIGUITY, with the residual pair recorded as quiet versus issuer_selective_omission.

Now add an external policy commitment signed from a separately operated release registry. In World A it closes clean. In World B it records the new policy version. The new channel splits the pair. Under the explicit assumption that at most one of the primary and external roots silently fails, two clean observations support NO_RELEVANT_CHANGE; a disagreement supports CHANGE or targeted revalidation.

If the external registry is populated from the same cached feed, it does not split the pair. The OSC root graph must then collapse the nominal external channel into the issuer root, and the decision returns to UNKNOWN.

The example shows why the mechanism is not merely "ask one more agent." The useful unit is an independently rooted observation that changes the equivalence class.

12. Brain-inspired hypothesis, kept narrow

The limited analogy is active sensing. A biological system can direct sensing actions toward information relevant to the current task [S18]. Encevra can similarly choose a probe that maximizes the expected number or weight of target-crossing ambiguity pairs it separates.

For unresolved pair set P and candidate probe a, define the model score:

value(a) = sum over p in P of weight(p) * separates(a,p) / cost(a).

This is a design heuristic, not a biological claim. The weights may be uniform, risk-based, or empirically estimated later. Under Round 26's evidence boundary, only uniform structural weights are used.

The testable hypothesis is:

At the same total observation cost, ambiguity-directed probes will leave fewer harmful target-crossing equivalence classes than fixed-cadence or randomly selected probes.

The necessary ablations are fixed probe order, random order, root labels removed, and a falsely independent witness. The analogy is rejected if the adaptive policy does not improve the preregistered structural and decision metrics.

13. Alternative explanations and rival designs

13.1 Conservative abstention may explain apparent safety

OSC may reduce false stability by returning UNKNOWN more often, not by learning more. E26-B must therefore report useful-current availability, false suspicion, revalidation load, and detection delay alongside harmful false-clean decisions.

13.2 An external channel may help because it is better, not because it is independent

The benefit could come from higher coverage or quality. E26-B should cross source quality with root independence where possible and include same-root high-quality and different-root matched-quality conditions.

13.3 Root metadata may be wrong

A declared witness can share infrastructure, credentials, prompts, retrieval caches, or upstream authorities. The correlation ablation shows that guarantees collapse when the graph is wrong. Operational root audits are part of the protocol, not optional documentation.

13.4 Active probes may perturb the system

Probes can create side effects, consume rate limits, reveal monitor strategy, or observe a path different from user traffic. The experiment needs synthetic, rate-limited probes, side-effect accounting, and separate coverage labels.

13.5 The finite model may omit the actual failure

An exact answer to an incomplete model can be precisely wrong. OSC must expose the model version and residual assumptions. Production decisions should include an escape path for unmodeled evidence and periodic threat-model review.

13.6 A probabilistic policy may be more useful than a zero-error certificate

Real systems may accept bounded risk. A later round can attach calibrated probabilities to equivalence classes, but probability does not remove identifiability limits. Without information or justified priors, a confident posterior can still be a restatement of assumptions.

13.7 End-to-end revalidation may dominate formal selection

Full revalidation can be simpler and safer for a small claim set. E26-B includes revalidate-all as a strong high-cost reference and compares every method at matched total verification budget.

14. Falsifiable hypotheses

No result for these hypotheses is claimed in this round.

15. Proposed experiment E26-B

15.1 Prerequisites

E26-B should run only when Encevra has:

• two or three frozen current claims with explicit relevance rules;
• at least three real authority observers with recorded dependency graphs;
• hidden independent adjudication unavailable to the tested policies until verdict lock;
• an immutable replay record;
• a sandbox capable of controlled omissions without touching production data;
• approval for non-destructive, synthetic end-to-end probes.

15.2 Strata

Natural shadow windows and sandbox fault windows must remain separate.

• Natural stratum: 120 consecutive eligible windows, analyzed as observed without injected faults. This estimates operational abstention and workload, not rare-fault recall.
• Sandbox stratum: 120 replayed or synthetic windows balanced across no fault, single-root omission, two-root omission, stale or rollback, common-origin witness, verifier common mode, and negative controls. This measures fault discrimination under known ground truth.

The sample sizes are planning targets, not power claims. A preregistration should calculate detectable effects from pilot-free conservative bounds or sequential stopping rules before collection.

15.3 Policies

1. Receipt-absence baseline.
2. Round 25 SACP.
3. SACP plus a fixed external channel.
4. SACP plus OSC-selected observables.
5. Same-root multi-agent ensemble with the same model-call budget.
6. Revalidate-all high-cost reference.
7. Oracle using hidden adjudication, reported only as an upper reference and never as a deployable policy.

15.4 Controlled total compute and observation budget

Each non-oracle policy receives the same per-window envelope:

• at most 12 observation units;
• at most 24 model-call units if language models are used;
• the same maximum wall-clock horizon;
• the same frozen candidate observation pool;
• the same evidence retention and privacy constraints.

One observation unit is preregistered by operation class and median cost, not retrofitted after results. Unused budget is recorded. Revalidate-all may exceed the envelope and is reported as a reference, not a matched competitor.

15.5 Baselines and ablations

• Remove root labels while preserving messages.
• Relabel the witness as independent while routing it through the primary cache.
• Remove verifier canaries.
• Remove explicit clean checkpoints.
• Replace adaptive probes with fixed order.
• Replace adaptive probes with random order under identical seeds.
• Collapse all LLM agents to one model and one retrieval root.
• Vary the assumed fault budget without changing observations.

15.6 Primary metrics

15.7 Regression and harmful-change metrics

• Privacy-bearing bytes collected by each observer.
• Probe side effects and rate-limit consumption.
• User-facing claim oscillation.
• Repeated alerts for the same unresolved equivalence class.
• Incorrect automatic restoration after a gap.
• Historical claim mutation.
• Cross-authority over-scoping.
• Decision reproducibility from the immutable evidence record.
• Verifier disagreement and unresolved verifier faults.

15.8 Analysis plan

Use paired comparisons on the same immutable windows. Lock policy outputs before hidden adjudication is revealed. Report every metric by natural versus sandbox stratum, authority family, fault root, claim, and observed root count. Do not pool injected frequency with natural frequency. Use exact paired tests or bootstrap intervals appropriate to the endpoint, publish all misses and false alerts, and report both absolute counts and denominators.

The primary superiority comparison is OSC versus SACP on harmful false-stability certification in the sandbox stratum. The co-primary noninferiority condition is useful-current availability in quiet natural windows with a preregistered margin no larger than five percentage points. Failure on either condition blocks adoption.

16. Expected failure modes of OSC

1. Threat-model laundering: an unsafe world is excluded without evidence.
2. Root inflation: multiple services are counted as independent despite shared infrastructure or upstream data.
3. Coverage overclaim: a behavior probe is treated as covering claims it cannot exercise.
4. Verifier recursion: the same component creates, evaluates, and certifies the evidence.
5. Stale matrix: system dependencies change while an old separation matrix remains valid on paper.
6. Assumption drift: a one-fault certificate is reused after the system accepts correlated failures.
7. Availability collapse: excessive UNKNOWN decisions trigger broad and costly revalidation.
8. Adversarial probe recognition: an issuer answers canaries but omits ordinary changes.
9. Clock and expiry failure: freshness checks share a faulty time source.
10. Evidence retention gap: the certificate survives but raw inputs needed for independent replay do not.
11. Model incompleteness: a real failure lies outside every enumerated world pair.
12. Metric gaming: policies optimize ambiguity counts while ignoring harm, privacy, or user-facing stability.

The protocol should fail closed on certificate integrity while failing scoped, not global, on target uncertainty.

17. Limitations

• The finite model is binary, atemporal within each window, and intentionally small.
• It represents worst-case silent failures but not forged positive change reports or every Byzantine interaction.
• The illustrative cost model is arbitrary and used only to demonstrate selection mechanics.
• The root-disjointness theorem depends on accurate root grouping and truthful surviving channels.
• A zero ambiguity count can reflect an incomplete world set.
• The analysis does not validate cryptography, signatures, clocks, network delivery, or source semantics.
• The multi-agent literature cited uses different tasks, models, and protocols; it motivates caution but does not validate OSC.
• Active sensing is a design analogy, not evidence of cognitive or neural equivalence.
• No independent human adjudicator reviewed E26-A because its ground truth is generated directly from the declared model.
• No live system, model, or benchmark result supports an effectiveness or deployment claim.

18. Decision and next-round handoff

18.1 Round 26 decision

Adopt OSC only as a research artifact and pre-deployment proof obligation. Do not install it as a production gate from E26-A. The operational rule carried forward is:

A quiet observation vector supports NO_RELEVANT_CHANGE only when its OSC equivalence class is target-pure under a declared, current, and independently reviewable threat model. Otherwise the result is UNKNOWN, and the public claim must be narrowed to what the evidence actually establishes.

18.2 Preferred Round 27

Run E26-B over a small real multi-authority shadow window with hidden independent adjudication, fixed budgets, and natural and sandbox strata reported separately. Publish every miss, false alert, abstention, restoration, and root-lineage correction. Do not deploy from the shadow trial.

18.3 Fallback if access remains unavailable

Do not build a larger synthetic world model. Instead, apply OSC manually to one real current claim, produce an audited root-dependency inventory, and measure whether each proposed observation would actually split a live ambiguity class. If even that access is unavailable, the next distinct theoretical question should be fixed-budget verification allocation using ambiguity reduction as the objective, with no effectiveness claim.

19. Reproducibility record

The bundle contains:

• encevra_research_round_26.md - manuscript source;
• round26_identifiability.py - exhaustive finite-model checker;
• test_round26_identifiability.py - deterministic tests;
• outputs/summary.json - model and minimum-set summary;
• outputs/package_metrics.csv - named policy outcomes;
• outputs/minimal_identifying_sets.csv - all inclusion-minimal sets;
• outputs/ambiguity_pairs.csv - all reported indistinguishable opposite-target pairs;
• outputs/minimal_blinding_cuts.csv - smallest structural cuts;
• source_inspection.csv - source, date, inspected material, support, and non-support ledger;
• outputs/sha256sums.txt - generated-output manifest;
• run.sh, README.md, and LICENSE.

Reproduction requires Python 3.11 or later and no third-party packages:

cd research/round26
./run.sh

The script prints minimum sets and rewrites the structured outputs deterministically. The PDF renderer is separate from the diagnostic and does not affect results.

20. Source inspection register

All web sources below were accessed on 2026-10-17. The bundle's source_inspection.csv records the full inspection ledger.

21. Full sources

[S01] Encevra Research. Research overview, rounds index, and registry. Lovable project 8483bc59-da36-492e-b8b8-6bf684070241, inspected at commit f1197fa663970737c3da35bc8e98c63c60f59c78, 17 October 2026.

[S02] Encevra Research. "Round 25: When Silence Looks Like Stability: Detecting Missing Change Evidence in Collaborative AI Systems." 16 October 2026. Full manuscript inspected in the same Lovable project.

[S03] Encevra Research. "Round 24." 15 October 2026. Full manuscript inspected in the same Lovable project.

[S04] Encevra Research. "Round 10: Evidence exchange, selective revision, and a testable path toward brain-inspired intelligence" and Round 9 on learning reusable procedures from verified experience. Full route texts inspected in the same Lovable project, 17 October 2026.

[S05] Lamport, Leslie, Robert Shostak, and Marshall Pease. "The Byzantine Generals Problem." ACM Transactions on Programming Languages and Systems 4, no. 3 (July 1982): 382-401. PDF. DOI: 10.1145/357172.357176.

[S06] Fischer, Michael J., Nancy A. Lynch, and Michael S. Paterson. "Impossibility of Distributed Consensus with One Faulty Process." Journal of the ACM 32, no. 2 (April 1985): 374-382. ACM record. DOI: 10.1145/3149.214121.

[S07] Dwork, Cynthia, Nancy Lynch, and Larry Stockmeyer. "Consensus in the Presence of Partial Synchrony." Journal of the ACM 35, no. 2 (April 1988): 288-323. PDF. DOI: 10.1145/42282.42283.

[S08] Halpern, Joseph Y., and Yoram Moses. "Knowledge and Common Knowledge in a Distributed Environment." Journal of the ACM 37, no. 3 (July 1990): 549-587. PDF. DOI: 10.1145/79147.79161.

[S09] Sampath, Meera, Raja Sengupta, Stephane Lafortune, Kasim Sinnamohideen, and Demosthenis Teneketzis. "Diagnosability of Discrete-Event Systems." IEEE Transactions on Automatic Control 40, no. 9 (September 1995): 1555-1575. IEEE record. DOI: 10.1109/9.412626.

[S10] Chandra, Tushar Deepak, and Sam Toueg. "Unreliable Failure Detectors for Reliable Distributed Systems." Journal of the ACM 43, no. 2 (March 1996): 225-267. PDF. DOI: 10.1145/226643.226647.

[S11] Sampath, Meera, Stephane Lafortune, and Demosthenis Teneketzis. "Active Diagnosis of Discrete-Event Systems." IEEE Transactions on Automatic Control 43, no. 7 (July 1998): 908-929. IEEE PDF. DOI: 10.1109/9.701089.

[S12] Haeberlen, Andreas, Petr Kouznetsov, and Peter Druschel. "PeerReview: Practical Accountability for Distributed Systems." Proceedings of the 21st ACM Symposium on Operating Systems Principles, 2007. PDF. ACM record: 10.1145/1323293.1294279.

[S13] Laurie, Ben, Eran Messeri, and Rob Stradling. "Certificate Transparency Version 2.0." RFC 9162, December 2021. RFC Editor. DOI: 10.17487/RFC9162.

[S14] The Update Framework. "The Update Framework Specification," current 1.x specification, retrieved 17 October 2026. Specification.

[S15] Du, Yilun, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch. "Improving Factuality and Reasoning in Language Models through Multiagent Debate." ICML 2024. Project and paper. OpenReview.

[S16] Wynn, Andrea, Harsh Satija, and Gillian Hadfield. "Talk Is Not Always Cheap: Understanding Failure Modes in Multi-Agent Debate." arXiv:2509.05396, 5 September 2025. arXiv.

[S17] Cemri, Mert, et al. "Why Do Multi-Agent LLM Systems Fail?" arXiv:2503.13657, 17 March 2025. arXiv. OpenReview.

[S18] Yang, Scott Cheng-Hsin, Daniel M. Wolpert, and Mate Lengyel. "Theoretical Perspectives on Active Sensing." Current Opinion in Behavioral Sciences 11 (October 2016): 100-108; author manuscript available in PMC in 2018. PMC. DOI: 10.1016/j.cobeha.2016.06.009.

22. Closing statement

Round 25 taught Encevra to distrust unexplained silence. Round 26 makes the stronger boundary explicit: some silence is not merely hard to detect; it is indistinguishable from stability under the available view. Reliability improves when the system records that fact, buys an observation that actually separates the worlds, or narrows its claim. It does not improve merely because more agents discuss the same missing evidence.