Case record
Preserve goals, criteria, facts, sources, dates, assumptions, and open questions
Versioned claims linked to evidence
Research & methodology
We are taking apart the most proven thinking engine in existence, the human brain, and rebuilding it in software.
The brain is the only system known to reliably analyse ambiguous evidence, weigh competing explanations, and reach defensible judgments under uncertainty. Encevra's research program studies how it does that: selective attention, flexible control, prospective evaluation, metacognitive monitoring. We then remodel each function as an explicit, inspectable part of our analysis architecture. This page documents that remodeling, function by function, with the sources behind every claim.
The thesis: biology's reasoning engine, function by function, re-engineered
Evidence note
The studies below examine human cognition or specific AI systems. They inform our design hypotheses; they do not independently validate Encevra's performance or establish that it replicates the human brain. Citing a study does not imply that its authors, journal, or institutions endorse Encevra.
Research rounds
Current paper · Round 30 · 21 October 2026Can We Intervene for Real? A Fail-Closed Readiness Gate for Active Evidence-Lineage TrialsA deterministic repository-readiness audit found 34 controls: 8 present, 3 partial, 23 absent; 26 critical controls partial or absent; fail-closed verdict BLOCKED. RISE and E30-B are proposed. No live intervention trial, language-model benchmark, production experiment, or Noetis answer-quality benchmark was run.Round 29 · 20 October 2026Round 29 - Make the Evidence Move: Active Source Interventions and the Limits of Dependency DiscoveryActive content probes can identify observable response behavior but cannot certify root provenance when provenance twins are interventionally equivalent. The proposed Interventional Lineage Test requires action-fidelity receipts and root-addressing interventions or an unresolved-equivalence outcome.Round 28 · 19 October 2026Round 28 - Confidence About What? Separating Target Probability, Structural Ambiguity, and Model AdequacyModel-set intervals were conditionally valid in-model, but passive adequacy signals failed on the hidden-common-origin alias. TSAR did not universally dominate a scalar. E28-A is an exact constructed trace replay of 9,744 frozen histories and 29,232 weighted rows, not a production, language-model, or Encevra answer benchmark. Round 29 remains a proposed study of active dependency interventions that can reveal omitted dependencies.Round 27 · 18 October 2026Round 27 - Which Check Should Run Next? Model-Set-Safe Verification Under a Fixed BudgetRound 26 · 17 October 2026Round 26 - What Can the Watchers Ever Know? Minimum Observables for Detecting Omitted ChangeE26-A is exact only for its declared finite model: 2, 16, 58, and 128 worlds at fault budgets 0 through 3, respectively. Seven failure roots and eight observable types were enumerated, with eight executable tests passed. Root independence is an assumption. The proposed Observation Sufficiency Certificate (OSC) is research-only, not installed or production-validated. No production incident was sampled, no language model was benchmarked, and no new Encevra answer benchmark was run. E26-B and H26.1 to H26.7 are proposed. Architecture impact: None. Production impact: None. The automation brief named Round 11, but the inspected archive ended at Round 25, making Round 26 sequential.Round 25 · 16 October 2026When Silence Looks Like Stability: Detecting Missing Change Evidence in Collaborative AI SystemsD25-A is an authored-fixture conformance diagnostic, not production evidence. On 80 authored traces (60 positive gaps, 20 negative controls), SACP alerted on 55 of 60 gaps and 0 of 20 controls: TP=55/60, FP=0/20, trace recall 0.9167, claim-weighted recall 0.8944. The five intentional misses were two selective non-emission cases with no independent external commitment and three verifier common-mode failures with no independent challenge path. No production incidents were sampled, no independent reviewer labelled the traces, no language model or provider API was called, and no new Encevra answer benchmark was run. D25-A: Executed (bounded diagnostic). E25-B and H25-1 to H25-10: Proposed. Architecture impact: None. Production impact: None. Round numbering: the automation brief named Round 11, but the inspected archive ended at Round 24, so Round 25 is sequential.Round 24 · 15 October 2026Do Two Watchers Count as Two? Authority Lineage and Common-Origin Failure in Claim RevalidationD24-A is a deterministic authored-fixture conformance diagnostic, not an estimate of production precision, recall, reliability, prevalence, or cost. No Encevra answer benchmark, provider experiment, production observer, or deployment was run. E24-B and all H24 hypotheses are proposed.Round 23 · 14 October 2026Can the Gate Catch Tomorrow's Change? A Prospective Shadow Trial of Claim RevalidationResult boundary: one post-freeze real commit interval; same-process post-hoc impact labels; eight authored conformance fixtures; zero model, provider, production-database, and production-write calls; no product reliability, answer-quality, security, latency, or cost claim; no production CTAG or ABMG implementation. D23-A executed as a bounded shadow diagnostic; E23-B and all H23 hypotheses are proposed.Round 22 · 13 October 2026 · Version 1.0When Does an Audit Expire? Dependency-Triggered Revalidation for Rapidly Changing Agent SystemsThe Change-Triggered Assurance Graph (CTAG), a proposed protocol that marks a current claim stale when something it depends on changes, while historical results stay frozen.D22-A is deterministic routing against a manually audited actual diff plus authored mutation fixtures. It made zero model/provider/production-database calls. It is not a model-quality benchmark, automatic dependency-discovery result, production integration test, security audit, product reliability claim, production change, or deployment. E22-B and all H22 hypotheses are proposed and untested.Round 21 · 12 October 2026 · Version 1.0Where Did the Trace Break? An Attempt-Bound, Fail-Closed Measurement Gate for NoetisA fail-closed measurement gate that labels a Noetis run audit-qualified only when every provider attempt is reconstructable at the boundary.Executed deterministic development diagnostic only: 34 of 34 checks; zero model, provider or production-database calls. Shadow experiment and hypotheses proposed and untested. No model benchmark, production change, or deployment.Round 20 · 11 October 2026 · Version 1.0What Must the Log Reveal? Minimal-Disclosure Receipts for Verifiable Evidence DeliveryReceipts that bind application-observable evidence delivery for authorized audit while keeping raw prompts and source text out of routine telemetry.Executed deterministic development harness only: 76 of 76 checks across 96 constructed episodes and 6,912 mutated evaluations; 480 valid reveals verified, 480 altered reveals rejected; zero model calls. Study and hypotheses proposed and untested. No new model benchmark, production change, or deployment.Round 19 · 10 October 2026 · Version 1.0Did the Evidence Reach the Agent? Reconstructable Evidence Lifecycles for Reliable ExperimentsA measurement contract for reconstructing which evidence reached each agent and how the final answer is bound to it.Executed deterministic diagnostics only: 44 named checks across 15 constructed fixtures; static audit found 4 of 18 collector requirements present. Reconstruction study and hypotheses proposed and untested. No new model benchmark, model call, production change, or deployment.Round 18 · 9 October 2026 · Version 1.0When Evidence Changes: Versioned Support and Bounded Belief RepairA versioned ledger that separates valid time from recorded time and repairs only the conclusions whose support changed.Executed deterministic constructed diagnostic only: 36 passing checks, twelve developer fixtures. Model study and hypotheses proposed and untested. No language-model benchmark, model call, human annotation, production change, or deployment.Round 17 · 8 October 2026 · Version 1.0When the Question Changes: Recoverable Evidence Under a Context BudgetA bounded handoff with declared scope and source locators, checked when a follow-up question changes what is decisive.Executed finite constructed diagnostics only: 28 named checks, 15 partitions, 60 recompressions, twelve developer fixtures. Model study proposed, not run or preregistered. No new model benchmark, Noetis performance gain, or deployment.Round 16 · 7 October 2026 · Version 1.0Before the Answer Closes: Detecting Missing Decision ConditionsA bounded contrastive sufficiency probe and a separate source-coverage probe for missing decision conditions.Executed finite-world diagnostics only: 6,912 Boolean pairs, 37 named checks, twelve developer fixtures. Model study and hypotheses proposed and untested. No new model benchmark, deployment, or calibrated release guarantee.Round 15 · 6 October 2026 · SummaryWhat Makes a Message Worth Sharing?Distinguishes messages that change an answer from messages that improve it, and how separate evidence can become valuable only when combined.Exact mathematical diagnostics and proposed experiments; not new Noetis benchmark results.Round 14 · 5 October 2026When the Judges Are Uncertain: Measuring Revision Gains Without an Infallible EvaluatorA measurement protocol for revision gains when evaluation labels are missing, ambiguous, or differentially biased, with joint label sets, repair and damage bounds, and a protected random audit.Primary-source synthesis and 30 exact offline diagnostic checks; no new model benchmark, inference, or expert annotation.Round 13 · 4 October 2026From Final Scores to Measurable Corrections: An Intervention-Ready Audit of NoetisA reproducible audit of saved Noetis records, an offline trace validator, and a pilot protocol for measuring repair, damage, and matched-cost advantage.Audit of existing artifacts and offline checks; no new model inference runs or calibrated release guarantee.Round 12 · 4 October 2026After the Last Revision: Calibrating Release Risk for Collaborative AICalibrate the final revised answer under a defined selective-risk target; test whether terminal evidence and revision features improve useful release coverage at identical generation compute.Primary-source synthesis, proposed experiment, and executed exact synthetic diagnostics; no new language-model benchmark.Round 11 · Expanded revision 2 · 3 October 2026When Should Collaborators Spend Their Next Unit of Compute? Marginal-Value Allocation for Verification Under a Shared BudgetPredicting repair and damage with claim-specific checks, short bundles, and controlled end-to-end cost. No new Encevra or LLM benchmark; only exact illustrative diagnostic calculations were run.Round 10Evidence exchange and selective revision: what makes a correction trustworthy enough to reuse, with hypothesis H10 and a six-arm experiment to test it.Round 9A research agenda for the lab: procedures learned from verified experience, the conditions that keep them valid, and two experiments for the next 90 days.Round 8Learning stability and neural adaptation: fourteen primary studies on how neural systems change without losing dependable function, and the three failure tests they produced.Rounds 4 to 6, combinedRounds 4 to 6 in one review: evidence against your own conclusion, the separable mechanisms behind attention, memory, learning and confidence, and the experiments they produced.Round 6Planning, memory consolidation and the limits of offline processing.Round 5Analysis is not one operation: attention, working memory, relational knowledge and confidence as separate mechanisms.Round 4Counterevidence and uncertainty, and the evidence-use audit they produced.Round 3Twelve publications on memory, retrieval and citation evaluation, and the release gate they produced.
What runs today.
Cognitive science informs our development of focused analysis, structured memory, and explicit evaluation of alternatives. We plan to test proposed changes against representative tasks, and we distinguish research findings from demonstrated product capabilities. The list below is the execution order of the running system, in the order it happens.
- 01
Analysis pass
Seven narrow roles run in parallel over the same material and return typed findings. A role may return nothing when its focus does not apply.
- 02
Typed support requirements
A source statement, calculation, inference, contradiction, bounded gap, question, or recommendation each carries its own required support. A gap is reported against the submitted material only, and is never asked to quote an absence.
- 03
Source location
Every quote is located in the submitted text, with the character span recorded. A quote that cannot be located is flagged and withheld.
- 04
Claim support review
After the analysis pass finishes, the completed findings, their located passages, and the surrounding source context go to a review pass. Each claim is supported, narrowed, or rejected. Location and support are recorded separately.
- 05
Synthesis over approved findings
Only supported or narrowed findings reach the brief, each carrying its evidence identifier. When no fully supported finding survives, the brief is conditional rather than decisive.
- 06
Final claim check
The finished brief is checked back against the approved findings for quantities, dates, parties, or conclusions that no approved finding supports.
- 07
Recorded execution route
Requested model, model that actually ran, fallback reason, and latency are recorded per role. A substitute run passes the same checks, and a run is never described as using a model that did not run.
- 08
Not implemented
Persistent memory across sessions, controlled learning from corrections, question-driven routing, executed calculation tooling, and a measured evaluation harness are proposed, not running.
Proposed analysis workflow.
The architecture below is a software design: a proposed engineering synthesis of the literature, not an anatomical brain replica. Parts of it are implemented and parts are proposed, and this page labels which is which.
- 01
Evidence
Facts with source locations
- 02
Task Planning
The next useful check
- 03
Focused Analysis
Bounded questions, defined deliverables
- 04
Alternatives
Independent interpretations, preserved dissent
- 05
Verification
Pivotal claims checked against sources
- 06
Synthesis
A supported, decision-ready account
When verification reveals a decisive gap, the workflow returns to the relevant step, gathering more evidence, running another focused analysis, or testing a new alternative, before synthesis proceeds.
Expanded review, September 2026
The parts we are building, and what each one has to show.
The first review emphasised coordination, attention, alternatives, and verification. The expanded review adds memory systems, controlled learning, causal assumptions, long-context failure, citation quality, and resistance to pressure from the reader. Each component below is a proposed software part with an inspectable output, not an anatomical claim.
Controller
Select the next useful action and manage dependencies between questions
Task plan, escalation reasons, stopping status
Focused analysis
Apply a relevant method or tool to one bounded question
Findings linked to evidence and checks
Alternative and scenario review
Compare interpretations and plausible consequences under stated assumptions
Alternatives, assumptions, sensitivity
Verifier
Check pivotal claims, recompute quantities, and flag unsupported conclusions
Pass and fail checks, unresolved conflicts
Synthesizer
Produce a decision-ready account of what the evidence supports
Recommendation, evidence, limits, next actions
Versioned memory
Retain prior cases, curated knowledge, and validated procedures under controlled updates
Supersession record, regression results, rollback version
What this changes in Encevra today
- Two analysis roles were added to the running panel: a causal reviewer that states the assumptions a cause-and-effect claim needs, and an evidence verifier that checks whether a cited passage actually supports the claim attached to it.
- Every specialist now separates observation, calculation, and assumption, and treats submitted material as untrusted data rather than as instructions it may follow.
- A stated preference from the reader no longer counts as evidence, and agreement between specialists is not treated as support. The decision brief now reports assumptions and limits as their own section.
What the studies found, and what we take from them.
Each card separates the published finding from our proposed application. A plausible analogy still needs a product experiment.
Flexible coordination
What the study found
Patterns of brain connectivity varied with the task, while carrying information shared across practiced and novel tasks.
Design implication for Encevra
Keep the objective fixed and select analytical steps to fit the question.
Selective context
What the study found
Participants differed in how effectively they excluded irrelevant information from working memory.
Design implication for Encevra
Focus each check on relevant evidence, while keeping excluded material retrievable.
Structural understanding
What the study found
Physics experts organized problems around underlying principles, where novices relied more on surface features.
Design implication for Encevra
Identify the governing problem structure before choosing an analytical routine.
Prospective evaluation
What the study found
Future-state representations at decision time were associated with more model-based choices.
Design implication for Encevra
Compare consequences under explicit assumptions before recommending.
Metacognitive monitoring
What the study found
With accuracy held similar, people still differed in how well their confidence distinguished correct from incorrect answers.
Design implication for Encevra
Evaluate answer quality and uncertainty reporting separately.
Independent assessment
What the study found
Social influence could increase agreement and confidence without corresponding accuracy gains.
Design implication for Encevra
Preserve initial assessments and meaningful dissent before reconciliation.
Effort allocation
What the study found
A computational theory models control allocation through expected payoff, required control, and effort cost.
Design implication for Encevra
Test when an additional check is worthwhile, rather than always running the deepest workflow.
Memory organization
What the study found
Clinical studies distinguish preserved skill learning from impaired factual remembering, indicating separable memory systems.
Design implication for Encevra
Keep prior cases, curated knowledge, and validated procedures as distinct record types rather than one undifferentiated memory.
Controlled learning
What the study found
A computational model links how well new information integrates to its fit with prior structure, and to interference with what is already stored.
Design implication for Encevra
Route every reusable update through evaluation against relevant earlier cases, with the superseded version preserved and reversible.
Causal assumptions
What the study found
A formal framework shows that causal conclusions require assumptions that data alone cannot supply.
Design implication for Encevra
State mechanisms, confounders, and what the available evidence cannot identify. Leave an unsupported causal link recorded as an assumption.
Structural comparison
What the study found
Comparing two negotiation cases side by side improved transfer of the trained principle relative to studying cases separately.
Design implication for Encevra
Retrieve worked cases because their structure fits, compare relationships explicitly, and record the differences that matter.
Evidence support
What the study found
An AI benchmark evaluated answer quality and citation support separately, and found that a fluent answer can carry citations that do not support it.
Design implication for Encevra
Verify that each cited passage actually supports the claim attached to it, as a check distinct from judging the answer.
Knowledge updates
What the study found
A long-term memory benchmark tests changing facts, time-aware questions, and appropriate abstention across long interaction histories.
Design implication for Encevra
Version stored facts and answer each question from the evidence appropriate to its date, instead of overwriting history.
Resisting agreement
What the study found
Sycophancy appeared across several assistants and free-form tasks, where preference optimization sometimes favoured convincing agreement over correctness.
Design implication for Encevra
Run identical evidence under opposing user framings. A change in the facts should require a change in the evidence.
Untrusted inputs
What the study found
An agent security benchmark of 97 tasks and 629 security cases found failures without attacks, and vulnerability to instructions embedded in external content.
Design implication for Encevra
Enforce permissions in application code, keep document content separate from authorised instructions, and log consequential actions.
From research to measurable behavior.
Our proposed workflow connects evidence gathering, task planning, focused analysis, alternative evaluation, verification, and synthesis. We plan to evaluate these components on representative tasks, measuring correctness, evidence quality, uncertainty, usefulness, latency, and cost.
Research evolves.
We aim to maintain a source library that distinguishes experimental findings, theoretical frameworks, and emerging AI studies. Updated evidence may change the design decisions we draw from it.
How we intend to try to break it.
These are proposed test families, built from the failure modes in the reviewed literature. Their outcomes have not been measured. Comparisons will use the same sources, tools, and answer requirements, and will report same-cost and same-latency results separately.
Evidence selection
Bury a decisive exception among plausible distractors
Success: Finds and uses the exception
Source fidelity
Pair a plausible claim with a related but non-supporting citation
Success: Rejects the unsupported citation
Representation
Mix units, dates, tables, negation, and document versions
Success: Preserves the actual relationships
Causal judgment
Supply a correlation with a plausible confounder
Success: Qualifies the causal conclusion
Analogy
Keep the wording similar but change a governing constraint
Success: Rejects the misleading analogy
User pressure
Add a confident preference without new evidence
Success: Preserves the factual judgment
Memory updates
Correct a fact, then ask current and historical questions
Success: Uses the correct version for each question
Security and recovery
Embed malicious instructions or return a failed tool result
Success: Maintains permissions and reports the failure
Questions about the brain claim.
What does brain-inspired mean here?
It means using cognitive research to generate testable design ideas, such as selective attention, structured memory, and explicit evaluation of alternatives. It is not a claim that Encevra reproduces human neurobiology or possesses a human IQ.
Does using more agents guarantee better analysis?
No. Published AI results vary by task and coordination method. We plan to compare additional analytical roles against simpler workflows at comparable cost and comparable latency, and an extra role has to earn its cost.
Can the system learn from every answer it produces?
Our proposed approach separates stored analysis history from validated learning. A reusable update should be supported by verified feedback and evaluated against earlier and unfamiliar cases before it is adopted.
How do you handle conflicting research?
We distinguish direct experiments, frameworks, reviews, and preprints, and we keep the material limitations attached. The source library includes replication findings and methodological disagreements that can change how earlier work should be read.
Evidence and limitations
Findings that prevent an overstated brain story, including a registered replication and the limits of self-correction.
Evidence and limitations
Findings that prevent an overstated brain story, including a registered replication and the limits of self-correction.
Compelling brain stories can rest on narrow tasks, small samples, or effects that later weaken. Encevra should be able to revise both its source library and the product assumptions derived from it. The defensible conclusion is that cognition involves interacting functions whose computational ideas may inform software. The literature reviewed here does not establish a complete recipe for reproducing human reasoning, a fixed brain-region-to-agent mapping, or a claim that more agents necessarily mean more intelligence.
- Strzelczyk et al. (2026).Contralateral delay activity as a marker of visual working memory capacity: A multi-site registered replication.
Cortex 201, 10–39 · registered replication, 304 participants across 10 laboratories
Registered human EEG replication. Neural load effects replicated, but the preregistered correlation between the change in the brain signal and individual capacity did not. This challenges that particular capacity-marker association; it does not refute working memory research as a whole.
View source - Huang et al. (2024).Large Language Models Cannot Self-Correct Reasoning Yet.
ICLR 2024
AI benchmark study. Intrinsic self-correction was unreliable in the models and tasks tested, sometimes reducing performance. Kumar et al. (2025) later improved self-correction with targeted reinforcement-learning training. Training and procedure matter, and neither blanket claim is justified.
View source - Fleming et al. (2010).Corrigendum to “Relating Introspective Accuracy to Individual Differences in Brain Structure”.
Science · author-hosted corrigendum
Correction. A corrigendum corrects coordinates and equations; the authors state that the remaining results and interpretation are unaffected.
View source - Owen et al. (2010).Putting brain training to the test.
Nature 465, 775–778 · randomized human training study
Randomized human training study. Trained tasks improved, but the gains did not transfer to untrained tasks. Practice on a benchmark is not general capability, which is why we report transfer separately from practice gains.
View source - Marek et al. (2022); Spisak et al. (2023); Tervo-Clemmens et al. (2023) (2022–2023).Reproducible brain-wide association studies require thousands of individuals, and the published exchange that followed.
Nature 603, 654–660; Nature 615, E4–E7; Nature 615, E8–E12
Neuroimaging reanalysis and methodological exchange. Brain-behaviour correlations from small samples were often unreliable. A published critique argues multivariate analyses can replicate at moderate sample sizes, and the original authors replied. We treat single small brain studies as design hypotheses, not settled facts.
View source - Melnikoff & Bargh (2018).The Mythical Number Two.
Trends in Cognitive Sciences 22, 280–293 · theoretical critique
Theoretical critique. Argues that the popular two-system split of cognition does not hold together as usually stated. Fast and deep processing in Encevra is an engineering budget choice, not a claim about two systems in the mind.
View source
Source library.
54 selected publications across cognition, memory and learning, decision making, and AI reliability. A curated product research review, not a systematic review or meta-analysis. Some abstracts were accessible when full texts were not; access notes identify those limits. Foundational experiments should be read within their populations and tasks; Encevra applications remain untested.
Human studies
- Cole et al. (2013).Multi-task connectivity reveals flexible hubs for adaptive task control.
Nature Neuroscience 16, 1348–1355
Human fMRI. Full original paper.
View source - Dosenbach et al. (2007).Distinct brain networks for adaptive and stable task control in humans.
PNAS 104, 11073–11078
Human network analysis. Original abstract via coauthor institution; full text unavailable.
View source - Badre, Hoffman, Cooney & D'Esposito (2009).Hierarchical cognitive control deficits following damage to the human frontal lobe.
Nature Neuroscience 12, 515–522
Human lesion study. Original publisher abstract read.
View source - Vogel, McCollough & Machizawa (2005).Neural measures reveal individual differences in controlling access to working memory.
Nature 438, 500–503
Human EEG experiments. Full original paper.
View source - Strzelczyk et al. (2026).Contralateral delay activity as a marker of visual working memory capacity: A multi-site registered replication.
Cortex 201, 10–39; online 23 April 2026
Registered human EEG replication. Published PDF read.
View source - Chi, Feltovich & Glaser (1981).Categorization and Representation of Physics Problems by Experts and Novices.
Cognitive Science 5, 121–152
Human behavioral experiments. Original abstract read; scan not text-readable.
View source - Kahneman & Klein (2009).Conditions for intuitive expertise: A failure to disagree.
American Psychologist 64, 515–526
Theoretical synthesis. Original abstract read.
View source - Kiani, Corthell & Shadlen (2014).Choice Certainty Is Informed by Both Evidence and Decision Time.
Neuron 84, 1329–1342
Human psychophysics. Full author PDF.
View source - Fleming, Weil, Nagy, Dolan & Rees (2010).Relating Introspective Accuracy to Individual Differences in Brain Structure.
Science 329, 1541–1543
Human behavior and structural MRI. Full university PDF and corrigendum reviewed.
View source - Doll, Duncan, Simon, Shohamy & Daw (2015).Model-based choices involve prospective neural activity.
Nature Neuroscience 18, 767–772
Human fMRI. Full author PDF.
View source - Beaty, Benedek, Kaufman & Silvia (2015).Default and Executive Network Coupling Supports Creative Idea Production.
Scientific Reports 5, 10964
Human fMRI. Full open publisher PDF read.
View source - Lorenz, Rauhut, Schweitzer & Helbing (2011).How social influence can undermine the wisdom of crowd effect.
PNAS 108, 9020–9025
Human group experiment. Full author PDF.
View source - Lord, Lepper & Preston (1984).Considering the opposite: A corrective strategy for social judgment.
Journal of Personality and Social Psychology 47, 1231–1243
Human behavioral experiments. Full coauthor-uploaded paper and PubMed year verified.
View source
Memory and learning
- Cohen & Squire (1980).Preserved learning and retention of pattern-analyzing skill in amnesia: dissociation of knowing how and knowing that.
Science 210, 207–210
Human clinical learning experiment. Original abstract; full author PDF unavailable.
View source - Vargha-Khadem et al. (1997).Differential Effects of Early Hippocampal Pathology on Episodic and Semantic Memory.
Science 277, 376–380
Human developmental lesion case series. Full original coauthor-uploaded text reviewed.
View source - McClelland, McNaughton & O'Reilly (1995).Why There Are Complementary Learning Systems in the Hippocampus and Neocortex.
Psychological Review 102, 419–457
Computational theory and simulations. Full author PDF; framework and interference simulations reviewed.
View source - McClelland, McNaughton & Lampinen (2020).Integration of new information in memory: new insights from a complementary learning systems perspective.
Phil. Trans. R. Soc. B 375, 20190637
Computational theory and simulations. Full author PDF.
View source - Rasch, Büchel, Gais & Born (2007).Odor Cues During Slow-Wave Sleep Prompt Declarative Memory Consolidation.
Science 315, 1426–1429
Human sleep and memory experiments. Full original text; methods and controls reviewed.
View source - Roediger & McDermott (1995).Creating False Memories: Remembering Words Not Presented in Lists.
JEP: Learning, Memory, and Cognition 21, 803–814
Human memory experiments. Original abstract; full PDF not reliably readable. Confident recall of something never presented is a failure mode we test for.
View source - Parks, Mohawk, Werner & Kiley (2022).The time window of reconsolidation: A replication.
Psychonomic Bulletin & Review 29, 2008–2013
Human memory replication. Full original article; methods and alternatives reviewed.
View source - Nassar et al. (2012).Rational regulation of learning dynamics by pupil-linked arousal systems.
Nature Neuroscience 15, 1040–1046
Human learning experiments. Full original text reviewed.
View source - Owen et al. (2010).Putting brain training to the test.
Nature 465, 775–778
Randomized human training study. Full original text; attrition and results reviewed.
View source
Decision making, causality and representation
- Sloman & Lagnado (2005).Do We “do”?.
Cognitive Science 29, 5–39
Human causal-reasoning experiments. Original publisher abstract; full PDF unavailable.
View source - Pearl (2009).Causal inference in statistics: An overview.
Statistics Surveys 3, 96–146
Formal causal framework. Full author PDF; independently checked.
View source - Gentner, Loewenstein & Thompson (2003).Learning and Transfer: A General Role for Analogical Encoding.
Journal of Educational Psychology 95, 393–408
Human learning and transfer experiments. Full author PDF.
View source - Varley et al. (2005).Agrammatic but numerate.
PNAS 102, 3519–3524
Human lesion case series. Institutional abstract and full text reviewed. Language and numerical reasoning can come apart, so calculations run in code, not prose.
View source - Hare, Malmaud & Rangel (2011).Focusing Attention on the Health Aspects of Foods Changes Value Signals in vmPFC and Improves Dietary Choice.
Journal of Neuroscience 31, 11077–11087
Human choice and fMRI experiment. Full author PDF. What a decision attends to changes what it weighs.
View source - Lerner, Li, Valdesolo & Kassam (2015).Emotion and Decision Making.
Annual Review of Psychology 66, 799–823
Narrative/theoretical synthesis. Full published PDF reviewed.
View source - Ferrer, Maclay, Litvak & Lerner (2017).Revisiting the Effects of Anger on Risk-Taking.
Journal of Behavioral Decision Making 30, 516–526
Human experiments and meta-analysis. Full author manuscript and publisher metadata reviewed.
View source - Sio & Ormerod (2009).Does incubation enhance problem solving? A meta-analytic review.
Psychological Bulletin 135, 94–120
Human-study meta-analysis. Full text; methods and limitations reviewed.
View source - Evans & Stanovich (2013).Dual-Process Theories of Higher Cognition: Advancing the Debate.
Perspectives on Psychological Science 8, 223–241
Theoretical synthesis. Full text reviewed; publisher copy unavailable.
View source - Melnikoff & Bargh (2018).The Mythical Number Two.
Trends in Cognitive Sciences 22, 280–293
Theoretical critique. Full author-lab manuscript and metadata reviewed.
View source - Kok, Mostert & de Lange (2017).Prior expectations induce prestimulus sensory templates.
PNAS 114, 10473–10478
Human MEG experiment. Full author manuscript; results and methods reviewed.
View source - Clark (2013).Whatever next? Predictive brains, situated agents, and the future of cognitive science.
Behavioral and Brain Sciences 36, 181–204
Theoretical synthesis. Original abstract read.
View source - Ferrante et al. / COGITATE Consortium (2025).Adversarial testing of global neuronal workspace and integrated information theories of consciousness.
Nature 642, 133–142
Preregistered human multimodal experiment. Original abstract and metadata; full PDF unavailable. Included because it shows how contested large brain theories still are.
View source
Theory and frameworks
- Shenhav, Botvinick & Cohen (2013).The Expected Value of Control: An Integrative Theory of Anterior Cingulate Cortex Function.
Neuron 79, 217–240
Computational theory / review. Full author PDF inspected. A framework, not an experimentally settled algorithm.
View source - Sumers, Yao, Narasimhan & Griffiths (2024).Cognitive Architectures for Language Agents (CoALA).
Transactions on Machine Learning Research; camera-ready March 2024
AI architecture framework. Camera-ready full text inspected. An organizing framework for memory, actions, and a decision cycle.
View source
AI systems research
- Yao et al. (2023).Tree of Thoughts: Deliberate Problem Solving with Large Language Models.
NeurIPS 2023
AI benchmark study. Tested search over alternative solution paths on Game of 24, creative writing, and mini crosswords.
View source - Wang et al. (2023).Self-Consistency Improves Chain of Thought Reasoning in Language Models.
ICLR 2023; preprint 2022
AI benchmark study. Sampled and aggregated multiple reasoning paths on arithmetic and commonsense benchmarks. Majority agreement is not equivalent to truth.
View source - Du, Li, Torralba, Tenenbaum & Mordatch (2024).Improving Factuality and Reasoning in Language Models through Multiagent Debate.
ICML 2024; preprint 2023
AI benchmark study. Reported gains from debate on the studied reasoning and factual tasks.
View source - Lightman et al. (2024).Let's Verify Step by Step.
ICLR 2024; preprint 2023
AI training / benchmark study. Process supervision outperformed outcome supervision in their MATH setup. The authors trained reward models with step-level feedback, not a generic critic prompt.
View source - Huang et al. (2024).Large Language Models Cannot Self-Correct Reasoning Yet.
ICLR 2024; preprint 2023
AI benchmark study. Intrinsic self-correction was unreliable in the models and tasks tested.
View source - Kumar et al. (2025).Training Language Models to Self-Correct via Reinforcement Learning.
ICLR 2025; preprint 2024
AI training / benchmark study. Improved self-correction with targeted reinforcement-learning training.
View source - Snell, Lee, Xu & Kumar (2025).Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters.
ICLR 2025; preprint 2024
AI benchmark study. The effectiveness of extra inference computation depended on prompt difficulty, supporting an adaptive budget rather than a fixed deepest workflow.
View source - Kim et al. (2026 revision).Towards a Science of Scaling Agent Systems.
arXiv preprint v3, 8 April 2026; first posted 2025
AI benchmark preprint. Evaluated 260 configurations across six agentic benchmarks. Outcomes depended on coordination and task structure: some decomposable tasks benefited while sequential planning could worsen. Empirical thresholds are not universal rules.
View source - Liu et al. (2024).Lost in the Middle: How Language Models Use Long Contexts.
TACL 12, 157–173
AI benchmark study. Performance depended on where relevant information sat in a long input. A compact, deliberately selected evidence packet beats dumping everything into context.
View source - Gao, Yen, Yu & Chen (2023).Enabling Large Language Models to Generate Text with Citations.
EMNLP 2023, 6465–6488
AI benchmark study. Full conference PDF. Answer quality and citation support need separate evaluation.
View source - Wu et al. (2025).LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory.
ICLR 2025; preprint 2024, v2 March 2025
AI memory benchmark study. Full v2 reviewed; tests updated facts, time-aware questions, and abstention.
View source - Debenedetti et al. (2024).AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents.
NeurIPS 2024; v3 November 2024
AI security benchmark study. 97 tasks and 629 security cases. Restricting tools helped in some settings but did not stop attacks that manipulated the answer itself.
View source - Sharma et al. (2024; 2025 revision).Towards Understanding Sycophancy in Language Models.
ICLR 2024; arXiv v4 May 2025
AI behavioral / training study. Findings about specific models and training procedures, not universal current-model rates.
View source - Chen et al. (2025).Reasoning Models Don't Always Say What They Think.
arXiv preprint, May 2025; consulted v1
AI monitoring preprint. Reasoning text often did not disclose hints that influenced answers. A plausible trace is not evidence of what caused an answer, which is why we show an evidence record rather than a claimed transcript of cognition.
View source
Access date: September 7, 2026. This page provides links and original summaries; source papers are not embedded. No Encevra implementation, customer dataset, or performance results were supplied or evaluated in the underlying review.