Rounds four, five and six
Encevra studies published work on human cognition to develop testable ideas for evidence use, uncertainty, planning and memory. We distinguish the findings of those studies from our engineering hypotheses. Proposed components are evaluated on defined tasks before we make claims about their performance. The authors below have not evaluated or endorsed Encevra, none of these papers tested Encevra or Noetis or Velri, and nothing here reproduces a human brain. Where only an abstract and publication record were reviewed, the entry says so.
Round four
How people use evidence that argues against their own conclusion, how uncertainty is represented before an outcome, and the evidence-use audit those questions produced.
Information arguing against a held answer can be present in the brain's activity and still be selectively under-used when the answer is committed to (Park and colleagues, 2025). Confidence in an initial judgement predicts reduced processing of later disconfirming evidence (Rollwage and colleagues, 2020). Seeing the evidence and using it are separate events.
Model accuracy on long inputs depended on where the decisive passage sat, with material in the middle used least reliably (Liu and colleagues, 2024). This is a property of the systems, measured directly, and is the reason position is a test variable rather than an assumption.
Predicted uncertainty appears to tune how incoming information is processed, ahead of any result (Frank and colleagues, 2026). A system that only reacts after an error is different from one that records what it does not yet know.
Orbitofrontal activity represents a map of the current state, including parts not directly observable (Schuck and colleagues, 2016), and the value of the rejected alternative is tracked separately from the chosen one (Boorman and colleagues, 2011). Options, constraints and the switching condition are part of the decision, not commentary on it.
Metacognitive accuracy can fail in one domain while remaining intact in another (Fleming and colleagues, 2014), and additional model-based deliberation only pays off under certain task structures (Kool and colleagues, 2016). Both argue for measuring each added check separately rather than adding checks by default.
Evidence against a held answer can be encoded and still be discounted at the point of use. That separation motivates auditing how a final answer uses counterevidence, not only whether it saw it.
Limitation: A human neuroimaging study of perceptual and value judgements. It says nothing about language models.
Read the sourceHigh confidence in a first answer suppressed later processing of disconfirming evidence. That motivates letting a reviewer read the sources before it sees the draft conclusion.
Limitation: A laboratory perceptual task. The mechanism is not established for document analysis.
Read the sourceModel accuracy varied with where the decisive passage sat in the input. That is the direct motivation for testing the same exception at the start, middle and end of a packet.
Limitation: The paper evaluated the models available to its authors at the time. It did not test the models Encevra uses now.
Read the sourceUncertainty anticipated before an outcome appears to be handled differently from surprise after it. That motivates recording what is uncertain, and why, as its own record.
Limitation: Intracranial data in a small patient sample. It is not a design for software.
Read the sourceDecisions are made against a represented state, including parts that are not directly observed. That motivates an explicit decision state: options, constraints, versions, dates.
Limitation: A controlled fMRI task. Business decision state is far larger and less defined.
Read the sourceThe value of the option not taken is tracked separately. That motivates naming the strongest alternative and the condition that would switch to it.
Limitation: A reward-learning task, not a procurement comparison.
Read the sourceOrdered planning steps motivate declaring the inputs a step requires before it runs.
Limitation: Navigation, not analysis. The analogy is a hypothesis, not a finding.
Read the sourceA compact cue appears to trigger recovery of a more detailed record. That motivates storing retrieval cues that point back to original sources rather than reusing summaries.
Limitation: Human electrophysiology. It does not validate a retrieval design.
Read the sourceSelf-monitoring accuracy can be impaired in one domain while intact in another. That motivates measuring monitoring per task rather than as one global score.
Limitation: A small lesion patient sample.
Read the sourceMore deliberation is not always worth its cost. That motivates measuring whether each added check produced a validated correction.
Limitation: Two-step laboratory tasks with small effect sizes.
Read the sourceRound five
Twenty-one further empirical studies, two background works and two published correction notices, with the literature search updated through 11 September 2026. The strongest conclusion is that selecting information, keeping it available, relating it to previous experience, judging uncertainty and learning from outcomes have distinguishable mechanisms. They interact; they are not a simple sequence of independent specialists. Full-text methods and results were consulted for 16 of the 21 additional studies; five entries are restricted to abstracts and publication records.
Mouse work implicated a prefrontal and thalamic pathway in choosing which sensory stream drives behaviour (Wimmer and colleagues, 2015), and a thalamic structure sustained a rule in cortex without carrying a readable copy of the rule itself (Schmitt and colleagues, 2017). Timed thalamic microstimulation in macaques raised detection by a reported 4.9 percentage points under one aligned condition (Boshra and colleagues, 2026). Representing a task and regulating how other circuits process it are different jobs.
Task-irrelevant information was still present in prefrontal populations; context changed how activity contributed to the choice (Mante and colleagues, 2013). Selecting an item already held in memory used shared control machinery with selecting an incoming one (Panichello and Buschman, 2021). Decoding an item does not show it is driving the answer.
An unattended item could still be probed from human EEG, while instructed forgetting reduced that trace (Wolff and colleagues, 2017), and prefrontal recordings showed brief gamma and beta events rather than uniform sustained oscillation (Lundqvist and colleagues, 2016). Retained, prioritised and prepared-for-use are three different states.
The same objects were reorganised by the current goal (Mack, Love and Preston, 2016), an abstract two-dimensional concept space showed a grid-like directional signature (Constantinescu and colleagues, 2016), and rest-period hippocampal ripples preceded later schema-based inference (Xiao and colleagues, 2025). Similarity, sequence and cause remain separate properties an experiment must distinguish.
Suppressing dorsal CA1 in mice impaired inferring an outcome never directly paired, while sparing the conditioned response (Barron and colleagues, 2020), and expert mathematical reflection recruited networks distinct from the core language network (Amalric and Dehaene, 2016). The engineering inference is modest: evaluate the operations an answer needed, not only how fluent its explanation is.
Timed dopamine activation enabled learning about an otherwise blocked cue (Steinberg and colleagues, 2013) and was reported to support learning between neutral events (Sharpe and colleagues, 2017, read with both correction notices). A system can learn efficiently from a biased signal and get better at the wrong objective.
Prefrontal and subcortical circuits were associated with learning how controllable outcomes are (Liu and colleagues, 2026), and the expected-value-of-control account frames whether further effort is worth its cost (Shenhav and colleagues, 2013). Outcome quality, decision quality and the value of another check are three different quantities.
Sensory reliability and distance from the boundary were combined in medial prefrontal activity (Bang and Fleming, 2018); macaque opt-out decisions drew on remembered content, uncertainty, trial history and arousal (Ning and colleagues, 2026); and choice and confidence developed together rather than strictly in sequence (Vivar-Lazo and Fetsch, 2026). A confident sentence, a self-rating and a validated error probability are not interchangeable.
Phase-aligned stimulation during sleep improved recognition in a six-person subgroup, mainly through fewer false alarms (Geva-Sagiv and colleagues, 2023), while striatal replay of a learned sequence persisted after hippocampal lesions (Thompson and colleagues, 2026). Reprocessing generated briefs is not automatically analogous to consolidation and could reinforce an incorrect conclusion.
Patterned cortical on and off periods in awake mice fulfilled some local sleep functions without replacing sleep (Driessen and colleagues, 2026), and an adversarial collaboration across 256 participants challenged important claims of two leading consciousness theories without naming a winner (COGITATE Consortium, 2025). A mechanism existing in biology, a software design capturing some abstraction of it, and that product beating a baseline are three separate claims.
Which input reaches a decision is actively gated, not simply amplified. That motivates treating selection as its own auditable step.
Limitation: Fifteen mice in the main behavioural cohort, with several intervention comparisons on three or four animals. It does not show that all attention uses this gate.
Read the sourceA structure sustained the rule without containing a readable copy of it, and more activity was not the same as better computation. That separates holding a goal from restating it.
Limitation: Small animal cohorts and indirect measures of functional connectivity.
Read the sourceA timed, spatially aligned intervention raised detection, giving causal leverage on attentional control rather than a correlation.
Limitation: Two macaques, a 4.9 percentage point mean gain in one condition, and electrical stimulation recruits a whole local cascade.
Read the sourceIrrelevant information stayed represented while contributing differently to the choice. That is the availability-versus-use distinction in its cleanest form.
Limitation: Trained perceptual choices in two macaques, mostly non-simultaneous recordings, with some differences between animals.
Read the sourceSelecting from memory used the same control representation as selecting from input. That motivates treating retrieval from a case record as a selection step with its own failure modes.
Limitation: Principally recording evidence in two macaques. Population subspaces are mathematical descriptions, not anatomical boxes.
Read the sourceAn item can be temporarily unattended and still recoverable. Not currently in use is not the same as discarded.
Limitation: Human EEG with 30 and 19 participants in the main experiments. Activity-silent storage is a hypothesis, not a proven synaptic mechanism.
Read the sourceAveraging differently timed events created a false impression of continuous activity. That is a direct warning about averaged evaluation metrics hiding intermittent failures.
Limitation: Four macaques across two tasks. It does not rule out persistent firing or establish one universal pattern.
Read the sourceThe organisation of the same objects changed with the current goal. That motivates re-deriving relationships when the decision question changes, instead of reusing a stored structure.
Limitation: Abstract and publication record reviewed only. Correlated activity does not establish that one structure creates concepts.
Read the sourceAbstract relationships showed a map-like signature, supporting judgements about positions never directly trained.
Limitation: Twenty-eight participants on a deliberately constructed two-dimensional task. No individual grid cells were recorded.
Read the sourceActivity during pauses predicted later inference about unseen relations rather than direct retrieval.
Limitation: Abstract-level review, 42 patients with epilepsy. Association, not proof that inducing ripples improves reasoning.
Read the sourceSuppressing one region impaired inferring an untrained outcome while sparing the directly learned response. Combining separately learned links is its own operation.
Limitation: The human and mouse tasks differed in complexity and instruction. This is convergence on a constrained computation, not identical reasoning.
Read the sourceExpert mathematical reflection was not reducible to ordinary sentence processing. Evaluate the operation, not the fluency of the explanation.
Limitation: Fifteen mathematicians and fifteen matched controls, cross-sectional. It is not an argument that a language model cannot reason.
Read the sourceA timed teaching signal changed what was learned about an otherwise blocked cue. Timing and identity of feedback matter, not just its presence.
Limitation: Deliberately artificial optogenetic conditions in rats. It does not make all dopamine activity a reward error.
Read the sourceLearning can concern which event follows which, not only how rewarding a cue is. That widens what an outcome should update.
Limitation: Abstract-level review, and it must be read with both correction notices below. Claims stay confined to the tested manipulations.
Read the sourceThe notice reports duplicated or mis-scaled figure histograms and a scatterplot point whose negative value had been transformed to its absolute value. It travels with the original paper.
Limitation: A correction notice, not new evidence. Its existence does not by itself invalidate the central result.
Read the sourceThe notice states that one reported stimulation period was one second rather than two. Publishing the record in full is part of citing it honestly.
Limitation: A correction notice, not new evidence.
Read the sourceLearning how controllable an outcome is differs from learning whether it was good. That motivates checking attribution before updating from feedback.
Limitation: Abstract-level review. A signal near a transmitter-associated nucleus is not a measurement of transmitter release.
Read the sourceA framework for deciding whether more effort is worth its cost, which is the question behind every optional extra check in a run.
Limitation: An integrative theory revisited as background, not a new experiment identifying a universal controller.
Read the sourceEvidence reliability and expected correctness were separable. Anything correlated with easy cases is not automatically a confidence signal.
Limitation: Thirty-two participants, fMRI associations in one perceptual task. It does not guarantee reported confidence is calibrated.
Read the sourceDeclining a risky retrieval drew on remembered content, uncertainty, trial history and arousal together. Monitoring has several inputs, including ones unrelated to the evidence.
Limitation: Abstract-level review. Opting out is a behavioural measure of monitoring, not evidence of introspection.
Read the sourceChoice and confidence developed together rather than strictly in sequence, so a confidence step bolted on after an answer is a design choice, not a law.
Limitation: 407 neurons in two macaques on one task. First online 18 November 2025, 2026 issue. Later revision of confidence is not excluded.
Read the sourceThe timing of communication between regions affected a specific memory outcome, which sharpens what an offline process would have to demonstrate.
Limitation: Eighteen participants, twelve tested, six in the synchronised subgroup. Gains were mainly fewer false alarms; associative accuracy did not reliably improve.
Read the sourceOffline replay persisted after extensive lesions of the structure usually assumed to drive it. A coordinated process need not have one indispensable source.
Limitation: Mice on a five-step sequence. It does not show every procedural memory is hippocampus-independent.
Read the sourceSome functions of a process can be isolated and tested separately rather than treated as one bundle.
Limitation: Local physiological and memory measures in mice. It does not show sleep is unnecessary or that all its functions were replaced.
Read the sourcePreregistering which observation would count against a favoured theory is the methodological standard we hold our own experiments to.
Limitation: Revisited as background, not counted among the round's additional studies. It named no winner and tested nothing about AI.
Read the sourceRound six
Eleven empirical studies and one correction notice, three of them published in 2026, selected only where they add a distinct mechanism, qualify the type of evidence, or change a concrete experiment. None of the selected work appears among the 118 distinct records of the earlier rounds. This round closes the bounded review: future literature searches should answer specific failures observed in Encevra rather than survey the field again.
Single-neuron recordings during awake neurosurgery found human prefrontal cells that tracked what another person believed, distinct from what the participant believed (Jamali and colleagues, 2021). For a system reading a vendor packet, that maps onto keeping each party's knowledge, claims and obligations separate rather than merging them into one view of the facts.
Brain-computer interface work found that patterns outside the existing repertoire were hard to produce quickly, and that the boundary moved only with long-term learning (Oby and colleagues, 2019), with later work relating learnability to the geometry of the available activity space (Busch and colleagues, 2026). The engineering parallel is blunt: prompting can surface a capability a system already has, and it is not the same as the system acquiring one.
Default mode network dynamics were related to creative thinking, with a reported causal component (Bartoli and colleagues, 2024). That motivates scoring originality of an analysis separately from how much it produces, since a longer brief is not a better one.
People do not use external reminders optimally, and their choices depend on metacognitive beliefs rather than only on actual memory (Gilbert and colleagues, 2020); training those beliefs improved offloading decisions (Ngai and Gilbert, 2026). Encevra's real unit of performance is a reviewer plus the tool, not the tool alone.
A surprising event engaged a fronto-basal ganglia suppressive mechanism that disrupted ongoing cognition, not just movement (Wessel and colleagues, 2016), and stimulating that stop circuit degraded cognitive performance in rodents (Heston and colleagues, 2020). Recovery after an interruption, such as a cancelled or failed step mid-run, is worth measuring as its own outcome.
Human layer 2/3 neurons showed dendritic action potentials supporting computations a single point-like unit cannot perform (Gidon and colleagues, 2020). This sets a realistic boundary on the biological analogy: parameter counts and neuron counts are not comparable quantities.
Cells tracked another person's belief distinctly from the participant's own. That motivates keeping each party's asserted knowledge separate in a multi-party document packet.
Limitation: 212 single units recorded in 11 people during awake neurosurgery. A clinical sample and a constrained belief task.
Read the sourceNew activity patterns appeared only after extended learning, separating what a system can already be coaxed into from what it must learn.
Limitation: A brain-computer interface task in monkeys. It does not transfer directly to model training.
Read the sourceRelates how learnable a control pattern is to the geometry of the available activity space, which is the human counterpart to the monkey result above.
Limitation: Abstract and publication record reviewed; full methods not reviewed.
Read the sourceAn intervention affected creative production, supporting originality as a separately manipulable quantity rather than a by-product of effort.
Limitation: Intracranial work in a clinical sample on laboratory creativity measures, not business analysis.
Read the sourcePeople offload according to beliefs about their memory rather than its actual capacity. A tool can be accurate and still be used badly.
Limitation: Abstract-level review. Laboratory memory tasks; first online 26 August 2019, 2020 issue.
Read the sourceTraining changed how well people used an external aid, which suggests measuring reviewer onboarding as part of product performance.
Limitation: Laboratory offloading tasks. It does not establish an effect for professional document review.
Read the sourceAn unexpected event suppressed ongoing cognition, not only motor output. Interruption and recovery are separate measurable costs.
Limitation: Human laboratory tasks with a specific stopping manipulation.
Read the sourceA causal rodent test of the same stop mechanism, strengthening the claim beyond a human correlation.
Limitation: Rodent circuit manipulation; the cognitive tasks are simple by design.
Read the sourceSingle human neurons performed computations a simple summing unit cannot, which bounds how far any brain-to-software comparison can be pushed.
Limitation: Abstract and publication record reviewed; full methods not reviewed. Slice physiology from surgical tissue.
Read the sourceThese are engineering experiments inspired by the work above. The studies do not validate them. Status labels reflect the actual project record: the audit runs on request and stays off in normal analysis until it has been evaluated.
After the brief is drafted, the run records how each material finding affects the recommendation: which findings support it, which challenge it, and one disposition for every material challenge. A challenge either changes the answer, narrows it, is settled by named evidence, or stays unresolved. References to findings that do not exist in the run are rejected, and evidence claimed as resolving must exist and must itself have passed claim support review.
A reviewer reads the decision question and the original sources without the draft conclusion or other specialists' output, then its observations are compared with the draft. The model and provider that actually ran each step, including any fallback, is recorded. Several models agreeing is not treated as independent evidence.
Each open issue attaches to the specific claim it affects, with a type, what it could change, one bounded next action, and its resolution status. A check that failed to run is recorded as an execution failure, never as low confidence and never as a completed review. No percentage confidence score is shown anywhere.
Long packets are split into contiguous, paragraph-aligned segments labelled with their true document position, and each specialist reads them in a different order. No passage sits in the middle of the prompt for every reader. Segments are exact slices of the submitted text, so quotes are still located character for character, and deterministic tests check that no segment is dropped, duplicated or altered under rotation.
An unresolved material challenge, or a failed structural check on the record, makes the brief conditional rather than settled, and the reason is shown with the brief.
A fictional two-vendor packet with a known answer, in three variants that move the decisive cancellation limit to the start, middle and end. Deterministic tests cover the record checks and the failure paths with mocked model responses.
Options, objectives, entities, constraints, source versions, assumptions and dates are extracted deterministically from the approved findings and the material, with the strongest alternative identified. Fields that cannot be established from evidence are left empty, not invented.
Stored findings keep a pointer to the original passage and its document version, with a material digest that detects when the document changed. A re-read always goes back to the source, so a generated summary cannot become authoritative through repetition. Scoped to a single case; cross-case learning remains a later, separately evaluated feature.
A deterministic check flags findings whose quote concerns a different party, period, object or superseded version than the question implies. The notes are advisory only; they never change a verdict or a release status.
Observed failure patterns and per-stage cost are logged from the execution trace. In this phase the router does not change model selection; it only records what a change would have done, so a future router can be evaluated against measured outcomes.
Each queued run records its stage (queued, running, completed, failed, requeued). When a worker is cancelled or its lease expires, the stage shows how far the run got and what it re-does on resume, making stuck-run debugging faster without changing the existing lease, cutoff or retry rules.
Derivations already declare their required inputs and sourced operands through the restricted arithmetic AST. A dependent calculation that lacks a required input returns an explicitly conditional result rather than a fabricated number.
These features are opt-in. The default API path is unchanged: a caller who does not request them gets the same behaviour, cost and output as before. None can turn an incomplete answer into a released one.
No performance result is claimed for these rounds. Any future number will be published with the task, the baseline, the sample and the measurement method. Routing between providers, adding memory and coordinating specialists changes the product system; it does not train a new foundation model.
