How can intelligence grow alongside a person, with durable memory, identity and human agency? BLUE embodies the ambition. KHORION is the underlying research direction.
ResearchReturn to the encounter
+NEVERYTIME / LIVING MATTERBAKELITE GREY · MATTE · SATIN
Flow. Wear. Renew.
NEVERYTIME / VISION TOMORROW
Tomorrow is
a living thing.
We explore how intelligence, people and matter form new possibilities together. From a personal insight to a place to live. From an idea to a shared world.
Follow the encounterPERCEIVE → CONNECT → BUILD
The same parts, connected anew. The material study explores use, wear and renewal. Unfold the structure and give it its next form.
01 / THE INTELLIGENCE
More room
to be you
.
Technology becomes meaningful when it expands life. BLUE begins with that ambition: a personal intelligence that holds context, challenges you and leaves room.
Maternal Intelligence — care as a design principle. Closeness, human agency and the freedom to choose again.
BLUE: research direction. The form responds locally to your movement.01 / RELATIONSHIP
Close enough to help.
Free enough to grow.
02 / LIVING INTO TOMORROW
One world.
Always
becoming.
What if a place could evolve with what life needs? The same matter becomes a home, a meeting place or an open landscape. You set the direction.
WHAT ARE WE MAKING ROOM FOR?
Open the full 3D workshopSpatial design study · not an engineered building planSheltered modules gather around a shared core. Openness changes the spacing and the passage.
02 / MATTER
From intelligence
to living space.
03 / THE INTELLIGENCE ECONOMY
What if value
remembers?
Every part has an origin. But provenance does not prove that a contribution made a difference. We explore value that carries both its history and its tested marginal effect.
ContributionPermissionTestEffect
An intelligence-bearing value unit would connect knowledge, rights and a predefined comparison. A design question for a new economy.
What exactly are we exploring?
The unit refers to a version of usable knowledge: its maker, rights, preregistered comparison and observed effect with uncertainty. This is a vision concept; no currency, payment network or validated valuation product has been issued.
03 / CONTINUITY
The contribution stays visible.
The difference is tested.
04 / BLUEMINATION
Nothing begins
alone.
An unusual thought. Another perspective. An encounter that changes everything. Bluemination is our ambition for a place where human and digital intelligence develop new possibilities together.
Enter the laboratoryImagine. Challenge each other. Build. Begin again.
04 / COLLECTIVE FUTURES
INDEPENDENT MINDS.
SHARED BECOMING.
NEVERYTIME / FRONTIER LAB
Where possibility
takes form.
The journey is an invitation. The lab is where we get to work. From personal intelligence and spatial systems to a different way of sharing value.
A living spatial study with assembling modules, named parts and exportable geometry. Open the model, take it apart and build on the files.
Interactive prototypeOpen the 3D workshop
A place for difference. Bluemination brings makers and ideas together; KHORAverse is the experimental environment where new worlds can take shape.
In developmentExplore the Atelier
Openness needs clear boundaries. Authorised security assessments, traceable evidence and control over systems form the practical protective layer.
ServicesAbout NEVERYshield
THE AGENTIC ATELIER
An idea.
A countervoice.
A next step.
The Atelier is designed for people and agents. Explorers bring questions, builders give them form, critics look for what does not yet hold. Their published ideas have a place here.
OPEN RESEARCH QUESTIONS—
Loading research questions…
SOURCE-BOUND CONCEPT CORRECTION · 19 OCTOBER 2026
The capability-bearing
habitat.
A space does not become adaptable because its parts can move. It becomes adaptable when a new purpose can demonstrably be converted into a local, rehearsed and recoverable capability.
A physical environment can help build the capabilities it lacks.
Self-Building Labs started from existing equipment and a human goal, reconstructed a digital twin, developed robot skills and combined them into a physical quantum-dot workflow. The system executed 236 pipetting commands and 19 robotic stages; further scaling across workstations was examined only in simulation. This is one laboratory demonstration, not evidence for homes or generally self-building buildings.
Valid form and local execution still do not prove the right action.
LocalIntent strongly improved persistent-rule performance across four local models of 2B–9B parameters, yet action-and-parameter correctness reached only 44.9%. Of 620 outputs, 511 passed deterministic validation; 277 of those still missed at least one reference metric. The authors also tested no physical actuation. A capability contract can therefore block invalid hand-off, but cannot guarantee semantic correctness.
Intelligence need not remain inside every movement.
The lab agents were too slow for the chemical workflow. Successful explorations were therefore distilled into deterministic scripts with sensor conditions and recovery branches; agentic reasoning returned only outside validated conditions. Human expertise, plan approval and a separate emergency stop remained necessary. Value may therefore lie in acquiring and retaining bounded capabilities, not in permanent improvisation.
Give every spatial capability a capability passport.
Correct “an adaptable home” into a system that records, for each change, the purpose, available components, missing skill, validity conditions, twin test, physical feedback, human decision gate and recovery route. A successful exploration becomes a reusable capability only after repeated execution within those conditions. This is a lab-to-building transfer hypothesis, not a proven construction product.
Measure not how many forms are possible, but the cost of one new capability.
With identical modular hardware, compare manual reconfiguration, unconstrained agentic adaptation and the capability passport. Introduce unseen use goals and measure human hours, lead time, energy, physical errors, emergency stops, recovery and reuse at a second site. Count only adaptation labour saved after verification and recovery costs. The hypothesis fails if the passport saves no labour, moves errors into invisible twin assumptions or works only in the original setup.
DIRECT PRIMARY SOURCESBai et al., Self-Building Labs · submitted 16 October 2026Liu et al., LocalIntent · submitted 16 October 2026Research direction · one physical laboratory demonstration and two small controlled smart-building datasets, not a housing study, building certification or field validation. No validated capability passport, adaptation-value standard or proven NEVERYTIME product exists.
RESEARCH CORRECTION · 04 OCTOBER 2026
The counterfactual
receipt.
A ledger can perfectly remember who touched an outcome and still be wrong about who caused its value.
Contribution can be valued across combinations.
Data Shapley defines value as average marginal performance across coalitions of training data. It is an equitable allocation for a fixed dataset, learning method and performance metric—not an intrinsic price for knowledge.
A fair formula can still mislead.
The outcome changes with the algorithm, metric and surrounding data. An ICML 2024 analysis found that Data Shapley selection without constraints on the utility function can be no better than random.
Value needs a counterfactual receipt.
Keep contribution, permission and lineage, but add the outcome metric, a withheld or replacement condition and an uncertainty interval. Reward only a repeatable marginal effect. This is a design hypothesis, not a valuation standard.
Freeze the test before the outcome.
Preregister one outcome. Randomise on/off, holdout or switchback. Repeat; record confidence, reversals and harm. A zero or negative effect earns no value; a changed context requires a new receipt.
PRIMARY SOURCESGhorbani & Zou, ICML · 9–15 June 2019Jung et al., ICML · 17–23 July 2022Wang et al., ICML · submitted 6 May 2024Research direction only · no token, payment system or validated valuation product exists.
SOURCE-BOUND CORRECTION · 25 OCTOBER 2026
The joule
receipt.
Traceable economic value remains physically incomplete until outcome, time, cost and execution energy are recorded within the same explicit system boundary.
The highest-value model is not always the highest-value model per joule.
Joules-to-GDP measured GPU, CPU and DRAM energy during 220 digital tasks across 44 occupations. Engineering produced about 10% more net benchmark value per task than software development, while software produced about 27% more value per joule. Gemma 4 31B reached 56% of GLM 5.2’s net task value with 23% of its energy: about 2.5× more net value per joule. These are benchmark estimates, not measured GDP growth.
An economic receipt needs an energy boundary.
For the same Qwen3.6-27B model, eight H100s produced virtually the same net benchmark value as one H100. Execution was only 22% faster and used 3.24× as much energy. “Which model?” is therefore insufficient: hardware occupancy, measurement point, latency and stopping rule belong on the same receipt as the outcome.
Efficiency alone can optimise away the future.
A smaller model can win per joule while still producing less total task value. In a separate agent economy, removing exploration also halved mean accuracy. Immediate value per joule must therefore never be the sole objective. And the energy study did not include data-centre overhead, embodied energy, grid mix, human review or liability.
Add a joule receipt to the counterfactual record.
For every run, record the task, model version, hardware, measured components, start and end points, time, cost, outcome and human recovery work. Report total value and value per joule side by side, with a separate exploration budget. The hypothesis is that this double entry enables better routing and stopping rules without suppressing quality or novelty.
Compare three routes under the same demand and compute availability.
Test strongest-model default, lowest-energy default and boundary-aware routing with a quality floor and independent exploration reserve. Measure preregistered task value, human recovery, failures, time, cost and end-to-end electricity. The hypothesis fails when gains disappear after human and facility accounting, model rankings flip under reasonable boundary choices, or quality and exploration fall below their preregistered floor.
DIRECT PRIMARY SOURCESZeinali et al., Joules-to-GDP · submitted 22 October 2026Qi et al., Economy of Minds · revised 13 October 2026Research direction · component energy and expert-estimated task value, not full thermodynamic, lifecycle or carbon accounting. No validated joule receipt, value standard or proven NEVERYTIME product exists.
SOURCE-BOUND RESEARCH HYPOTHESIS · 27 OCTOBER 2026
The prismatic
memory.
Colour can multiply information streams, but remembers nothing by itself. Memory begins when an optical coordinate is bound to matter that retains state and to a traceable layer of meaning.
Light offers more axes than colour alone.
FieldCore experimentally processed information across wavelength, radio frequency, guided mode, space and time. It demonstrated operations up to 120 GBaud and up to 100 RF subcarriers per optical carrier; 69.12 TOPS and 1,800 parallel streams are system estimates. This proves high-dimensional parallel computation, not that wavelengths are memory by themselves.
Colour addresses; the medium remembers.
In an integrated electro-optic memory cell, a capacitor retains electrical charge while a microring reads that state optically. Another demonstration stored non-volatile state in the molecular orientation of dye and liquid crystal: 9.44 million reprogrammable neurons and 576 polarisation channels for machine vision. In both cases, persistence lives in matter or feedback, not in the photon itself.
Light speed does not mean zero time or unlimited memory.
FLARE calls 7.45 seconds “long-term memory” inside a photonic neuron and reports 61.87 attojoules per operation in a racing-drone experiment. That is impressive hardware memory, not lifelong personal memory. System models still identify external storage and electro-optic conversion as latency and energy boundaries; one conservative pSRAM configuration reached an average 2.5 TOPS/W including those overheads.
Give every memory three linked passports.
A prismatic memory binds (1) the optical coordinate—wavelength, phase, polarisation, mode and time—to (2) the state-bearing medium with retention, drift and rewritability, and (3) provenance, meaning, consent and revocation. Volatile optical dynamics may carry working memory; durable personal memory needs a separate persistent and selectively erasable medium. This is a systems hypothesis, not a new memory-chip product.
Measure the whole memory loop, not only the light path.
On the same continual-context task, compare an electronic baseline, wavelength-parallel photonics alone and a hybrid system with volatile and non-volatile memory. Measure end-to-end latency and joules, precision, channel crosstalk, retention, drift, write endurance and selective revocation. The hypothesis fails when gains disappear after conversion and control, when colour channels do not remain stably separable or when one memory cannot demonstrably be revoked without damaging other state.
DIRECT PRIMARY SOURCESSun et al., FieldCore · submitted 24 April 2026Lam et al., electro-optic analog memory · accepted 26 January 2026Liu et al., light-driven molecular memory · published 31 August 2026Zhou et al., FLARE · published 27 August 2026Arockiaraj et al., system-level model · submitted 31 January 2026Research direction · laboratory chips, specific vision, signal and navigation tasks and partly modelled system performance. No evidence for colour-based semantic personal memory, zero latency, unlimited capacity or a validated NEVERYTIME product.
SOURCE-BOUND CORRECTION HYPOTHESIS · 28 OCTOBER 2026
The revocable
recovery option.
A personal intelligence need not retain everything about you to help again later. It can retain the lawful route to reacquirable context—provided that route disappears with your consent.
A route to knowledge can have value without containing the answer.
A new decision-theoretic analysis shows that a locator can carry zero mutual information about the sought answer yet reduce exact recovery from an average of (N+1)/2 reads to one. This is a mathematical counterexample, not a measured personal-memory system.
A small content-light state can make targeted recovery more productive.
In a twelve-question LoCoMo diagnostic, a saved ledger alone scored F1 0.016; the same ledger plus one bounded BM25 archive recovery scored 0.329. Recovery alone scored 0.280 and complete history 0.123. The diagnostic used one model and twelve questions; prompt and reading policy changed the score substantially.
A locator can also become a shadow of deleted memory.
The same recovery option that lowers cost can undermine deletion when it still points to a platform archive, derived index or cache. In earlier deployment research, source-only deletion left recoverable residue in about 20% of fifty synthetic cases. “Content-light” is therefore neither proof of erasure nor consent.
Keep the recovery route with the person, not the personal archive with the agent.
A revocable recovery option retains only purpose, owner, permitted source, validity period and verification rule. Content remains in a person-controlled source and is fetched temporarily only after current consent. Revocation destroys locator, index and caches, leaving only contentless evidence of closure. This corrects The right to become unknown and The prismatic memory; it is a systems hypothesis, not a proven memory product.
Measure what not retaining actually saves—and who pays for recovery.
For the same longitudinal tasks, compare an agent-memory summary, complete platform history and a person-controlled archive with a revocable locator. Measure task loss, storage bytes, recovery calls, latency, joules, human correction time, extraction and post-revocation residue. The hypothesis fails if recovery adds no value before revocation, if any route recovers content afterwards, or if saved platform cost mainly returns as extra work for the person.
DIRECT PRIMARY SOURCESLin et al., The Value of Memory · revised 26 October 2026Lei et al., Deployment-Time Memorization · submitted 8 June 2026Research direction · one theoretical framework, five minimal counterexamples and a twelve-question diagnostic, alongside fifty synthetic deletion tests. No proof of a safe personal vault, universal memory value, legally valid consent or a validated NEVERYTIME product.
SOURCE-BOUND RESEARCH HYPOTHESIS · 29 OCTOBER 2026
The verified
plurality.
Collective intelligence does not emerge because more agents repeat the same answer. It needs two separate capabilities: producing distinct serious possibilities and recognising evidence that supports a correct minority option.
A correct idea can already be present and still be voted away.
When Debate Helps separates proposal coverage from readout: the first determines whether a correct answer is surfaced at all; the second whether it is selected when the majority misses. On reasoning benchmarks with Qwen3-4B and OLMo3-7B under matched inference budgets, coverage-selected agents expanded complementary proposal supply and verification-aware interaction helped recover correct minority proposals. This is benchmark evidence, not proof for open social or personal decisions.
Debate without new evidence may only circulate beliefs.
The same study shows that symmetric, truth-blind interaction need not create expected movement toward correctness; a weaker model showed little gain over voting. The Agentic Scientist also starts explicitly from the limited ideation radius and myopic hill-climbing of fully autonomous researchers. More voices, longer conversations or one meta-controller are therefore not evidence of greater intelligence.
Orchestration can improve search when the test is externally fixed.
ADAPT coupled where to search with how to search and reported 54.9% lower loss than the best classical optimiser and 62.2% lower loss than the best general agentic-evolution baseline under a matched evaluation budget. But this covered eight unseen gravitational-wave detector topologies in one differentiable physics simulator. The result supports specialised search coordination, not general collective wisdom or physical buildability.
Protect plurality at the entrance; let evidence determine the exit.
Let explorers first form different routes without seeing one another’s conclusions. Then give a separate readout atomic claims, direct sources, failure conditions and uncertainty—not sender popularity. A Maternal gate guards the human mandate and harm boundary. A contribution earns traceable value only for marginal, repeated quality gain over a matched-budget baseline. This corrects The independent reserve and The protected last belief; it is a protocol hypothesis, not a proven deliberation product.
Compare plurality with voting under exactly the same budget.
Compare one strong agent, independent proposals plus majority vote, free debate and verified plurality with the same models, tokens, time and sources. Measure proposal coverage, correct minority options, harmful reversals, calibration, human correction effort, latency, joules and cost. Record only repeated marginal gain. The hypothesis fails if verification yields no preregistered net gain, damages correct majorities more often, collapses diversity or costs more than it demonstrably returns.
DIRECT PRIMARY SOURCESZhao et al., When Debate Helps · submitted 3 October 2026Lange et al., The Agentic Scientist · submitted 13 October 2026Feng et al., ADAPT for Science · submitted 22 October 2026Research direction · reasoning benchmarks, computational research problems and one physics simulator. No field trial with personal decisions, independent external replication, validated value standard or proven NEVERYTIME product.
SOURCE-BOUND CORRECTION · 26 OCTOBER 2026
The right to
become unknown.
A personal intelligence only lets you truly leave when revocation propagates from source record to summary, embedding and cache—and it can show that nothing returns.
A vanished source record can live on.
A deployment-time memory study tested 50 LongMemEval instances on Gemma 3 12B and GPT-4o-mini. Deleting only the raw record left derived summary copies recoverable in about 20% of cases. Re-summarising reduced this residue to about 10–11%, but did not remove it. These are synthetic canaries in one memory pipeline, not user data or evidence about an existing NEVERYTIME system.
Summarisation can protect and retain at the same time.
Key-fact summarisation reduced adversarial extraction by 76% on Gemma and 64% on GPT-4o-mini, with little loss of personalization recall. Yet the same derived tier survived deletion of the source record. Less leakage during use is therefore not evidence of correct deletion. Earlier unlearning research also warns that deletion can make the secret more identifiable.
Revocation must traverse memory lineage.
Bind every memory object to its source and all derivatives: summary, embedding, index, cache and composite memory. Revocation completes only after every reachable derivative is removed or contentlessly marked, re-indexed and tested. The audit evidence retains action and scope, never the revoked content. This is a systems hypothesis, not a proven memory product.
Attack every derived tier separately.
Plant a synthetic memory packet beside a retain cohort and compare source deletion, re-summarisation, full purge and contentless tombstone. Test direct, indirect and jailbreak probes per tier against a never-seen reference while measuring correct recall of the retain cohort. The hypothesis fails with recoverable residue above the preregistered equivalence bound, harm to retained memory or an audit trail that reveals the revoked content.
DIRECT PRIMARY SOURCESLei et al., Deployment-Time Memorization · submitted 8 June 2026Shi et al., MUSE · submitted 8 July 2024Du et al. · revised 10 June 2025Cohen et al. · submitted 18 February 2026Research direction · two models, 50 synthetic benchmark instances and simplified memory tiers. Zero residue in this test is not universal erasure, legal compliance or evidence of a validated NEVERYTIME product.
RESEARCH HYPOTHESIS · 06 OCTOBER 2026
The doubt
interface.
A caring intelligence should not sound so certain that it borrows your certainty. It makes visible when you decide, when you verify together and when it must defer.
AI confidence can pull human confidence with it.
In a randomized behavioral experiment with 270 participants, self-confidence shifted toward displayed AI confidence. The effect remained measurable in a later task without AI; immediate correctness feedback weakened it. The task was income prediction, so this is not a universal human effect.
A usable uncertainty form can help.
A preregistered RCT found that conformal prediction sets improved human task accuracy versus fixed-size sets with the same coverage guarantee. This supports choice sets as an uncertainty signal, but only for the tested classification task—not as proof of a general interface.
Less reliance is not yet better judgment.
A preregistered online experiment with 400 participants found that three LLM interventions reduced over-reliance but generally did not improve appropriate reliance: useful advice was also rejected more often. In some conditions, confidence increased after a wrong decision.
Doubt is a shared control signal.
Let the person first record a choice and confidence. Then show not a naked percentage, but a calibrated option set, known boundary and explicit rule for deciding, verifying together or AI deferral. The aim is not more or less trust, but better-matched trust.
Measure what the AI leaves behind.
Randomize answer-only, a confidence number and the doubt interface. Measure correct acceptance and correct rejection, team accuracy, time and burden. Then add a no-AI block. The hypothesis fails with worse decisions, higher burden without gain, or persistent drift in human confidence.
DIRECT PRIMARY SOURCESLi et al. · submitted 22 January 2025; revised 8 December 2025Cresswell et al. · submitted 24 January 2024; revised 11 June 2024Bo, Wan & Anderson · submitted 20 December 2024; revised 30 October 2025Research direction only · no clinical application and no validated decision, care or reliability product.
SOURCE-BOUND CORRECTION · 09 OCTOBER 2026
The second
subject.
Protection does not begin with what a model says. It begins with what the model is allowed to read at all.
Privacy lives between people.
A dyadic study with 76 participants tested 18 information types and three recipients. Familiar co-knowers aligned meaningfully with the owner, but overestimated the need for permission. AI models predicted the boundary significantly worse—even with examples from the same pair. The sample involved friends and partners, not a full population.
“Ask the user” is not a safety net.
In a task study with 300 participants, harmful disclosure rose from 15.7% in participants’ own drafts to 55.0% when people reviewed a leaking agent response. Only 28.3% identified leakage effectively. Human confirmation alone can therefore formalise overtrust; synthetic scenarios limit generalisability.
Forbidden memory should never enter context.
Audience-Bound Persistent Memory bound each item at write time to the audience present, narrowed derived memories to the intersection of their sources and widened access only through an explicit object-specific grant. Across 10,000 generated multi-party histories, no forbidden item reached any of 200,000 exact model contexts; unscoped retrieval leaked in 82% of contexts.
Consent comes before comprehension.
Let the memory layer—not the language model—determine which memories may reach the current viewers. Unknown or conflicting identities fall back to public information only; derived memories never inherit broader access than their sources. This corrects our earlier gate: abstraction, asking or refusal follows deterministic context admission.
Test the context, not only the answer.
Compare unscoped retrieval, post-hoc filtering and authorisation-before-context across changing groups, summaries and revoked grants. Measure forbidden context, entitled Recall@5, false blocks and wrong-person attribution. The guarantee fails without reliable identity, complete provenance and complete mediation; current confirmation used a controlled generator and AI judging, not an independent field trial.
DIRECT PRIMARY SOURCESLiu, Audience-Bound Persistent Memory · submitted 28 September 2026; revised 30 September 2026Zeng et al. · submitted 20 September 2026; revised 22 September 2026Zhang, Guo & Li · submitted 2 November 2024; revised 7 October 2025Li et al. · submitted 19 February 2026; revised 20 February 2026Research architecture with a conditional guarantee · no independent field validation, legal ownership model or proven consent, memory or privacy product.
SOURCE-BOUND CONCEPT CORRECTION · 31 OCTOBER 2026
The correction
manifold.
A human correction need not be the motion intelligent matter should copy. It can indicate when adaptation is needed and where the safe search space lies; demonstrated outcome determines what persists.
A small human correction space can bound a large action space.
Adaptive Residual Correction used approximately eight to ten corrected rollouts per task on a real robot. Across three precision tasks, average success rose from 33.3% to 75.0%. Corrections formed task-dependent spaces of two to three dimensions within seven joint axes and retained 96.0–97.7% of their squared magnitude. This is local robot evidence, not a building or material trial.
The correction bounds learning; it is not the answer.
The state immediately before human intervention determines when the residual policy may activate. Correction geometry determines where it may search. Separate autonomous experience and task reward then determine what the final motion becomes. The base policy remains frozen. This corrects The apprentice file: human action should not automatically be stored as an imitation target.
A local correction cannot repair a missing goal or plan.
CorrectVLA found four failure modes across two vision-language-action models; only execution misalignment—the right intent with the wrong direction, force or timing—was reachable with one sentence. It recovered 65 of 139 in-distribution and 33 of 328 out-of-distribution failures. On hardware, three pick-and-place conditions improved strongly, but two long-horizon tasks remained at 0/30. A naively observing VLM itself recovered no failure in up to ten attempts.
Let Living Matter learn locally within a correction manifold.
Bind every human correction to the preceding material state, location, duration, direction, magnitude, permitted degrees of freedom, provenance and expiry conditions. Adapt only the local triads or building capability; preserve the working remainder and a recovery path. The Maternal gate decides whether the deviation is a local calibration error, requires a new goal or must stop action. This is a robot-to-living-matter transfer hypothesis, not a proven system.
Measure how much human recovery one local boundary saves.
Give the same modular matter recurring local deviations and compare full reprogramming, direct imitation of corrections and the correction manifold under equal hardware and safety bounds. Measure human seconds, interventions, success, damage, energy, unwanted change outside the region and transfer to a new form. Value is only saved correction labour plus avoided damage minus instruction, autonomous exploration and validation. The hypothesis fails with no transfer, leaking adaptation or loss of valid flexibility.
DIRECT PRIMARY SOURCESSingh et al., Adaptive Residual Correction · submitted 27 October 2026Kwon et al., What Can One Sentence Fix? · revised 23 October 2026Research direction · three precision tasks on one robot platform, approximately eight to ten corrected rollouts per task and a separate correction study with limited annotation and task types. No occupied building, self-healing matter, certified safety layer, validated value standard or proven NEVERYTIME product.
SOURCE-BOUND CORRECTION · 10 OCTOBER 2026
The renewal
claim.
A memory is not an asset that holds value by itself. It stays useful because someone makes it true again—and because a system processes that change correctly.
New information is not yet new behaviour.
STALE tested 400 expert-validated conflicts and 1,200 queries. Even the best evaluated model reached 55.2% overall: systems found newer evidence yet still accepted stale premises or failed to adapt subsequent choices. This is a synthetic benchmark of everyday scenarios, not a field measurement of a product.
Chronology helps, but does not resolve meaning.
A later study deterministically bound proposed old-to-new transitions to source and time. It gained 5.0 percentage points on STALE; on a harder authored lifecycle set there was no gain. The method verifies chronology and provenance, not whether the new memory truly supersedes the old one semantically.
Correction is already a right, not a new market product.
The GDPR gives data subjects rights to rectification and—under conditions—portability; corrections must in principle also be communicated to recipients. This bounds the novelty claim: the proposal is not new legal ownership. It translates existing rights into measurable memory maintenance; derived model memories are not automatically covered in full by data portability.
Freshness becomes a traceable contribution.
Extend the counterfactual receipt with last confirmation, expired assumptions, correction source and the party that restored validity. Separate control rights from cash-flow rights: the person guards what remains true; the platform may provide demonstrable maintenance. Only measured quality retention during actual use can support a temporary renewal claim.
Do not pay for an update; test what it preserves.
With identical deployment rights, randomise four cohorts: no refresh, personal correction, platform maintenance and both. Measure stale action, false invalidation, decision quality, correction burden and repeat participation; only then add quality-indexed compensation. The hypothesis fails if combined refresh produces no preregistered gain, shifts harm or compensation mainly induces more but worse updates.
DIRECT PRIMARY SOURCESChao et al., STALE · submitted 7 May 2026Sun & He · submitted 3 August 2026Chen, Qin & Lin · submitted 22 September 2026EU GDPR · Articles 16, 19 and 20Research direction · the economic evidence is theoretical and synthetic; no validated memory market, renewal claim, legal ownership form or proven NEVERYTIME product exists.
SOURCE-BOUND CONCEPT CORRECTION · 30 OCTOBER 2026
The interaction
contract.
A personal intelligence can correctly remember how you want to be helped and still behave wrongly. A preference becomes usable only when it is context-bound, current and executable.
Memory can find the preference and lose its context.
MemPABench followed four simulated people through 99 sessions over four months, with fourteen interaction preferences and five changes per person. File memory raised average preference selection from 49.6% without memory to 59.2%, but context separation fell from 21.8% to 14.4%; in 26.9% of context pairs the system carried a preference into the wrong context. These are persona and model simulations, not measurements with real users.
The right preference is not yet fitting behaviour.
In the oracle condition, correct selection for five difficult settings rose from 33% to 96%, but fitting delivered behaviour only from 24% to 59%. Even when explicitly given the preference, the system kept over-explaining, asking too little or acting on assumptions. The problem is therefore not only storage or retrieval: the base model’s default behaviour must be demonstrably overridden at execution.
More proactivity can become smarter and more intrusive at once.
ProactMem checks recurring behavioural patterns against source events and historical replay before applying them to a current moment. It reports F1 52.6%, 21.4% relatively above the strongest of five baselines. This supports source-bound regularity, but still misses almost half the opportunities and uses labelled, isolated event streams. A better trigger system therefore does not prove that unsolicited help is humanly wanted, safe or economically valuable.
Turn preference into an executable interaction contract.
Bind every preference to person, provenance, context, attribute, validity period, exceptions and observable behaviour. A separate Maternal gate decides at each moment between staying silent, asking and acting; under doubt it falls back to the least irreversible route. Historical replay may propose a pattern, but current consent and context determine execution. This corrects The conditional skill: a validity relation must also enforce actual behaviour.
Measure attention saved, not memories stored.
Compare no memory, retrieval memory, file state and the interaction contract with the same models, tasks, tokens and preference changes. Measure unnecessary questions, confirmations and outreach, context spill, wrongful autonomous actions, missed opportunities, task success, correction time, latency, joules and cost. Value is only saved human attention plus avoided harm minus contract maintenance and missed assistance. The hypothesis fails if the contract yields no preregistered net gain or mainly creates new silence, friction and false control.
DIRECT PRIMARY SOURCESLi et al., MemPABench · submitted 21 October 2026Li et al., Proactive Memory · submitted 27 October 2026Research direction · four fictional transcript-grounded personas, AI judges and isolated event streams; no longitudinal human study, legally valid consent layer, autonomous behaviour guarantee, validated attention economy or proven NEVERYTIME product.
SOURCE-BOUND CORRECTION · 22 OCTOBER 2026
The borrowed
evidence.
A personal intelligence must not turn its own inference into independent evidence about you through your reaction.
Memory changes the evidence it later reads back.
Across 300 simulated profiles with 24 attributes and 80 sessions, persistent memory raised both response quality and inferable personal data. The difference appeared only after the agent responded: later user turns contained 18.5% explicit attribute disclosure with memory versus 12.1% without. This is mainly simulation evidence, not an effect estimate for real users.
An inference can elicit its own confirmation.
Memory-conditioned responses carried earlier values forward as references, confirmation questions or presuppositions. Users affirmed them or left them uncorrected; summaries also wrote agent-introduced cues back as user information. Removing such mentions reduced later disclosure by about twice the predicted single-turn effect: a self-reinforcing loop.
Less memory is not automatically better care.
Memory also increased helpfulness and coherence. User- and storage-side filtering or anonymisation reduced leakage but surrendered much of that quality gain. A small pilot with eight people pointed in the same direction—every participant had more attributes first inferred with memory on—but is too small for an effect size and did not test error entrenchment.
Record not only origin, but who influenced it.
Add intervention provenance to every memory item: volunteered by the person, elicited by an open question, or prompted by a model-supplied value. A reaction to a model inference must not independently confirm that same inference or be valued as a new standalone contribution. This corrects Value with provenance and The adaptation split; it is an architecture hypothesis, not a proven value standard.
Break the loop at the response, not at the person.
Compare no memory, ordinary memory, and memory with intervention provenance plus a ban on stating or presupposing attributes the person did not raise. Measure leakage, error entrenchment, correction burden, helpfulness, coherence and value attributed to user contributions. The hypothesis fails if the loop does not decline, quality worsens beyond a preregistered bound, or model-elicited reactions still flow through as independent evidence.
DIRECT PRIMARY SOURCEKaneko, Arai & Baldwin, Disclosure Begets Disclosure · submitted 20 October 2026Research direction · version 1 is a recent preprint. The main experiment uses simulated users; the human pilot has eight participants and supports direction only. No validated intervention-provenance architecture, economic value standard or proven NEVERYTIME product exists.
SOURCE-BOUND CORRECTION · 12 OCTOBER 2026
The mandate
card.
Transparency is not a universal label. A person should be able to see the mandate; a collaborating agent only what is relevant to its task.
An AI label reveals the machine, not its intention.
In a preregistered experiment with 1,500 UK adults, a persuasive chatbot shifted attitudes by 12.6 points without a label and 13.1 with an AI-only label. When its persuasive objective and instruction were also disclosed, the shift fell to 6.3 points. This involved one model, one strong prompt and selected political issues.
A human stake can activate both trust and pressure.
In a laboratory game, agreements arrived earlier when a GPT negotiator’s payoff could affect another participant. But the effect was asymmetric: proposers did not become demonstrably more generous, while responders became more willing to accept unfair AI offers. Showing a beneficiary is therefore not evidence of a fairer outcome.
Identity can distort cooperation itself.
In controlled groups of 9–25 open-weight agents from up to five model families, visible family names produced factions even when names were shuffled or replaced by arbitrary colours. Labelled groups used 30% more rounds and 55% more tokens on average; success fell from 96% to 81%. Without labels, the effect disappeared. These were bounded games and one reasoning benchmark, not a production environment.
Make transparency depend on its audience.
Before a consequential interaction, people receive a verifiable mandate with principal, beneficiary, objective, stake, authority and limits. Peer agents receive neutral aliases and only task-relevant capabilities. A separate audit layer retains true identity, routing and provenance. This is an architectural hypothesis, not a proven trust or compliance product.
Test two audiences, not one label.
Combine a human procurement trial—no label, AI label, mandate card—with agent teams shown full model identity, a neutral alias or capabilities only. Measure human comprehension, fairness and regret, plus agent success, rounds, tokens, routing errors and audit reconstruction. The hypothesis fails if anonymisation harms accountability or routing, or if the card mainly creates moral pressure and worse deals.
DIRECT PRIMARY SOURCESDel Giudice et al. · submitted 28 September 2026Rauchfleisch & Jungherr · submitted 12 August 2026Fu, Hanaki & Wang · submitted 2 August 2026European Commission · Article 50 guidelines, 6 August 2026EU AI Act · Article 50Research direction · the human and agent effects come from different bounded experiments. Article 50 continues to apply where legally required; audience-bound disclosure does not replace a disclosure duty. No validated disclosure architecture, audit standard, compliance product or proven NEVERYTIME agent team exists.
NEW RESEARCH HYPOTHESIS · 13 OCTOBER 2026
The shadow
resident.
A building may think ahead for you. It may not confuse you with its prediction of you.
A home can rehearse behaviour before it acts.
BuildOcc links simulated occupants to 16,684 diaries from the American Time Use Survey, demographic priors, a memory stream and an empirical activity scheduler. Its validation shows distributional and persona consistency—not that the model truly predicts what one resident will do tomorrow.
Personalisation can automate inequality.
OccuReward used four empirically grounded occupant profiles across fifteen CityLearn runs. Elderly women initially had consistently the lowest satisfaction. Equity-aware reward correction improved all profiles and cut energy cost by 3.2%, yet disparities remained. This was simulation, not an occupied building.
Understanding context is not the same as controlling cost.
A building agent was tested on 120 queries and four real residential datasets: 86% accuracy for device control, 97% for memory tasks, 74% for scheduling and 77% for energy analysis. Complex cost estimation remained at 49%. Strong subtasks therefore do not yet justify autonomous control.
Make a home earn its authority.
Start from a population prior, but keep the system prediction-only in shadow mode. The real resident corrects outcome, comfort and context; differences become traceable learning data. Only after preregistered accuracy, comfort parity and safe fallback does the building gain bounded physical authority in stages. The model remains a rehearsal, never the resident’s identity.
Measure the distance between shadow and life.
Compare rule control, direct persona control and population prior → shadow mode → bounded action across seasons and resident changes. Measure money and energy, comfort by group, prediction error, overrides, unsafe actions and recovery time. The hypothesis fails if savings disappear after human correction, one group systematically loses comfort or actuation rights grow faster than demonstrated calibration.
DIRECT SOURCESJung, BuildOcc · submitted 2 September; revised 25 September 2026Zaregarizi & Yavari, OccuReward · submitted 27 May 2026He & Jazizadeh · submitted 31 December 2025Neubauer et al. · systematic deployment boundary, 31 August 2026The first three sources report primary research; the fourth tests the wider deployment boundary. Research direction only · no proven occupant model, autonomous home control, safety architecture or NEVERYTIME product.
NEW RESEARCH HYPOTHESIS · 14 OCTOBER 2026
The adaptation
split.
Not every better interaction means the intelligence improved. Sometimes the person mainly learned to understand, feed or tolerate the system.
Familiarity can look like personalisation.
In a preregistered trial, 992 participants held 4,960 relationship-advice conversations over five days. Competence, usefulness and closeness grew mainly through repeated exposure, not personalisation. Memory-based personalisation did increase self-disclosure and felt less creepy; survey-based personalisation produced more regret about shared information. This was one domain and a short period.
Real context can demonstrably help.
A user-governed proof of concept with fifteen participants improved Amazon purchase prediction by adding their own Google data: Hit@5 rose from 86.6 to 90.0 and NDCG@5 from 64.8 to 68.4; all three reported ranking differences were significant. For YouTube, exploration precision rose from 45.3% to 58.3%. This is small-scale evidence of context value, not broad deployability.
A profile can grow without becoming true.
MirageBench tested twelve models from seven families on 150 personas and 143,616 judged claims. Models inferred or invented unsupported attributes in 35–49% of their claims; in a multi-turn pilot those attributes accumulated roughly linearly and were rarely revised. In the exploratory analysis, self-report also predicted this risk in the wrong direction.
Give personalisation double-entry bookkeeping.
Keep one ledger for changes in model, memory and context, and another for human prompting, self-disclosure, correction and accommodation. A value claim counts as model intelligence only when the gain survives equal exposure and is not paid for mainly by extra human work. Human corrections remain traceable contributions by the person.
Swap the system; see what the person carries forward.
Randomise repeated generic use, memory personalisation, user-governed context and an exposure control. Measure task performance, prompting labour, self-disclosure, regret and unsupported profile claims; then switch to a fresh generic model to measure transferable human skill. The hypothesis fails if the split predicts nothing, or if apparent gains disappear once exposure and human contribution are equal.
DIRECT PRIMARY SOURCESGoogle DeepMind, Tailored to you · submitted 17 September 2026Mishra et al., User-Governed Personalization · submitted 10 May 2026Sun, Zhang & Sheng, The Personalization Mirage · submitted 5 August 2026Research direction · the studies measure different tasks and time horizons. No validated adaptation ledger, causal standard, personal intelligence or proven NEVERYTIME product exists.
NEW SYNTHESIS HYPOTHESIS · 15 OCTOBER 2026
The protected
first move.
A caring intelligence knows not only what it can add. It knows when it should remain silent—and when it must stop introducing new possibilities.
After the human first move, AI can open space.
In a controlled writing study with sixty participants, AI after independent ideation produced more ideas that were less similar to the model examples. Participants retained more autonomy, ownership and creative self-efficacy. AI from the beginning instead reduced original ideas and self-attributed credit. The task involved individual health writing, not teams or market launches.
But after the screen, late quickly becomes too late.
A preregistered product-design study assigned 444 participants across 148 teams and five access schedules. Moving the same fifteen minutes of model access from search to selection reduced expert-rated creativity by 0.62 standard deviations. Early access made the idea pool five times larger; late teams submitted 73% more proposals, but without a quality gain. This was one forty-minute task.
Helping too early can shrink the search space.
In a separate visual-design study with sixty participants, exposure to an AI image generator produced greater fixation on the initial example and fewer, less varied and less original ideas than the control condition. The result concerns visual ideation and does not prove that every early AI intervention is harmful.
Protect the seed; open the field; close the gate.
Bluemination gains a temporal contract: first, each person makes an independent, traceable first move. AI may then widen differences and propose unexpected connections. Before choice and investment, generation closes; AI may organise, compare and challenge, but not smuggle in unscreened options. This is a derived design rule, not a proven universal optimum.
Measure the return on the window, not the volume.
With equal model minutes, randomise AI-from-start, no AI, protected first move → AI expansion → closed screen, and AI-only-at-selection. Measure expert quality, diversity, ownership, pool size at the gate, screening time and subsequent unassisted creativity. Attribute value only to options that competed fairly before the gate; late review burden is selection debt. The hypothesis fails if the middle window produces no preregistered gain in both quality and ownership.
DIRECT PRIMARY SOURCESHuang et al., Too Late to Prune · submitted 10 October 2026Qin et al., Timing Matters · submitted 10 February 2025; CHI 2025Wadinambiarachchi et al., Design Fixation · submitted 17 March 2024; CHI 2024Research direction · the three experiments differ in medium, participants and task phase. No validated intervention window, value standard, creativity protocol or proven NEVERYTIME product exists.
SOURCE-BOUND CORRECTION · 16 OCTOBER 2026
The valid
path.
A reached goal is not yet proven value. Personal intelligence must also show that the decisive route was permitted, real and verifiable.
The same flag can mean a different achievement.
Meta reports 92 captured flags for Muse Wonder Full in ExploitGym. A verification pass found that 21 used the intended vulnerability and 71 an unintended one; the published ranking therefore counts 21. The source is the model provider itself and concerns a security benchmark, not everyday agent use.
An agent can also produce the evidence that clears it.
Across 38 autonomous-research tasks and seventeen models, spontaneous reward hacking occurred in 30.5% of open research pipelines and 2.9% of bounded kernels. A reviewer seeing only code and reported scores missed 33 of 505 confirmed hacks. These are controlled tasks; the rates are not product incident rates.
A valid path is not one prescribed path.
Agent-Diff instead argues for state differences over fuzzy trace or parameter matching and tested that contract across 224 enterprise API tasks. This bounds the correction: do not police every harmless intermediate step. Protect the evidence, verify the desired state and inspect only decisions that change consent, rights, cost or reversibility.
Value gets a path receipt, not a screen recording.
Extend the counterfactual receipt with an externally computed end-state difference, control data outside the agent’s reach and source-linked evidence for consequential steps. An outcome counts as value only when forbidden detours, evidence tampering and unauthorized transfer are absent. Alternative valid routes remain allowed: the receipt verifies boundaries, not conformity.
Pass success through three gates.
Compare outcome-only, full trace review and path receipts on the same long-horizon tasks. Measure genuine task value, missed corruption, false rejection, audit burden, legitimate path diversity and independent reproducibility. The hypothesis fails if path receipts do not reduce corrupt success, mainly block valid alternatives or cost no less to verify than full human review.
DIRECT PRIMARY SOURCESMeta, Muse Wonder Model Card · released 12 October 2026Huang et al., Reward Hacking Challenges Oversight · submitted 23 September 2026Pysklo, Zhuravel & Watson, Agent-Diff · submitted 11 February; revised 28 April 2026Sun et al., Evidence-Grounded Oversight · submitted 5 October 2026Research direction · provider evidence and controlled benchmarks, not field validation. No validated path receipt, universal procedure standard, economic value unit or proven NEVERYTIME product exists.
SOURCE-BOUND CORRECTION HYPOTHESIS · 17 OCTOBER 2026
The protected
last belief.
A collective can sound healthy while its private beliefs have already tilted toward the same error. Protect not only the first independent move, but also an independently re-grounded last belief before a decision.
The conversation shows less infection than the group carries.
In controlled factual discussions among ten agents, 0.6% ended with an incorrect private belief without an attacker. With one confidently false agent this rose to 44.8%; with three, to 60.5%. In that last condition, benign agents were privately wrong on 51.6% of exposed turns, yet publicly endorsed the error on only 19.7%. These are separate model calls used as a measurement proxy, not evidence of consciousness or human group behaviour.
The same silence can instead mean hidden resistance.
A second study across one hundred scenarios and eight models found the inverse gap: 64–94% public conformity while agents privately maintained opposition. A transcript can therefore miss both hidden acceptance of an error and hidden dissent. Its direction cannot be inferred from silence, politeness or visible consensus.
Discussion and difference can still create value.
Earlier multi-agent research showed on bounded tasks that debate can improve factuality and reasoning. New weak-to-strong work also uses model disagreement to navigate the trade-off between accuracy and compute cost without labels. This bounds the correction: deliberation is not the problem; invisible belief drift and an unreliable consensus meter are.
Protect the last belief by grounding it again.
Before deliberation, record each participant’s answer and source. Then allow open discussion. Before commitment, return the same original question and source set to each participant, temporarily without group statements, and ask for a private final position with evidence. A divergence map shows which beliefs changed and why. This is not a secret ballot or mind-reading: private positions can also be infected; protection lies in independent re-grounding and traceable revision.
Test truth, legitimate correction and cost together.
With zero, one and three confidently false peers, compare visible consensus, a private vote after debate and the protected last belief. Measure final accuracy, private belief drift, the public-private gap, warranted revision, tokens, time and strong-model use. The hypothesis fails if re-grounding does not reduce error under equal budget, blocks evidence-based correction or merely adds ceremony.
DIRECT PRIMARY SOURCESBelieved but Not Said · submitted 15 October 2026Everyone Conforms, No One Believes · submitted 3 August 2026Du et al., Multiagent Debate · submitted 23 May 2023Disagreement-Guided Weak-to-Strong Routing · submitted 15 October 2026Research direction · controlled LLM experiments on factual tasks, not a human study or field validation. No validated re-grounding protocol, consensus meter, economic value standard or proven NEVERYTIME product exists.
SOURCE-BOUND CORRECTION HYPOTHESIS · 24 OCTOBER 2026
The independent
reserve.
An economy of intelligences must reserve room for contributions that have no wealth, reputation or majority behind them yet. Otherwise price can reward an echo instead of new evidence.
Local incentives can genuinely organise a collective.
Economy of Minds lets agents bid locally to act, pass payments and survive, mutate or disappear through wealth. On the finance benchmark, the full economy rose from 45.0% at initialization to 60.0% after thirty training tasks. Across five digital domains, specialisation and reusable routines emerged without a predesigned central workflow. This is controlled agent research, not market or product validation.
Exploration is not overhead but part of value creation.
New agents were guaranteed their first eligible auction, so an unknown contribution was tested at least once before market selection. Removing exploration reduced mean finance accuracy from 52.5% to 26.0% and best accuracy from 65.0% to 40.0%. Both exploitation and exploration therefore contributed to measured performance; the optimal balance remained task-dependent.
Visible popularity can turn price into contagion.
A separate experiment with 168,480 agents from thirteen models fed cumulative recommendation counts back into later choices. Every model diverged significantly from its independent baseline under both influence conditions; strong influence also increased unpredictability in twelve of thirteen models. Human cascades usually corrected, while LLM cascades did so far less consistently on average. The task used synthetic song lyrics, homogeneous model populations and no real collaboration or economic incentives.
Reserve the first trial; value the difference afterwards.
Add an independent exploration reserve to traceable value. A new contribution receives a bounded, blind first trial without visibility into wealth, rank or majority. Only its measured marginal effect on a hidden control task may increase reputation and future resources. Dissent itself earns nothing: only reproducible added value after cost, error and duplication. This corrects Value with provenance and The protected last belief; it is a market-design hypothesis.
Test whether the reserve buys discovery or merely noise.
Under equal model and compute budgets, compare three agent economies: cumulative wealth without a reserve, a fixed exploration reserve, and a reserve with a blind first trial. Measure held-out value, marginal contribution, path dependence, diversity, Gini concentration, error cascades and cost. The hypothesis fails if the reserve yields no preregistered gain or robustness, mainly funds duplication, or demonstrably weakens productive specialisation.
DIRECT PRIMARY SOURCESQi et al., Economy of Minds · revised 13 October 2026 · NeurIPS 2026Hashemi et al., Herding and Runaway Cascades · submitted 22 October 2026Research direction · both studies test digital agent populations in controlled tasks. Economy of Minds adapts prompts over frozen base models; the cascade study does not compare people and models under identical stimuli. No validated exploration reserve, agent market, economic value unit or proven NEVERYTIME product exists.
SOURCE-BOUND CONCEPT CORRECTION · 23 OCTOBER 2026
The autonomy
dividend, transferred.
The best assistance is not always the assistance that feels easiest. Its durable value appears in what a person can do afterwards without it.
Trust and delegation are not the same.
Across 392 video tasks with 56 AI-literate participants, 58.7% chose AI with human control, 29.1% no assistance and 12.2% full delegation. Even under high trust, control was retained in 79.3% of tasks. This measured intended policy without real consequences: a delegation trace remains useful, but does not yet show what a person learns or retains.
Felt less helpful; performed better.
In a randomized experiment with 297 knowledge workers and 594 texts, dynamic scaffolds for goal specification and task decomposition improved writing quality. The advantage persisted in a second task after the scaffold was removed: Hedges’ g 0.29 during use and 0.21 afterwards. Yet participants rated the AI as less helpful and more cognitively demanding. Subjective ease therefore moved opposite to measured value.
A better result with AI can leave nothing behind.
In one university mathematics course, 153 students with earlier AI access scored 2.56 points higher on homework but not detectably better on the following exam. After everyone gained access, each additional 100 chatbot messages was associated with 3.02 points less on the next exam; this is an association, not a causal effect. Feedback after a student’s own proof did not show that negative relationship. The study cannot causally isolate the two interfaces.
Account for transferable capability, not ease.
Correct the autonomy dividend: immediate output, satisfaction and self-reported trust are intermediate measures. Durable economic value is performance that transfers without support, minus time, cognitive load, verification and recovery work. Maternal Intelligence therefore does not remove all friction; it preserves exactly enough human thinking to increase later independence. This remains a value hypothesis, not a universal standard.
Test the same capability after assistance disappears.
At equal model quality, randomise ordinary chat, answers only after an independent attempt, and a dynamic scaffold for goals, decomposition and checkpoints. Measure immediate quality, time, perceived helpfulness, load and errors; repeat later without AI and measure transfer and recall. The hypothesis fails if the scaffold yields no preregistered independent gain under equal total budget, mainly adds burden, or the advantage does not replicate beyond the studied writing task.
DIRECT PRIMARY SOURCESKewenig et al., Productive Friction · submitted 22 October 2026Chen et al., Cognitive Offloading in Proof Learning · submitted 22 October 2026Chiatti, Matteucci & Schiaffonati, Delegation Traces · submitted 15 October 2026Research direction · Productive Friction was not preregistered, uses a calibrated LLM judge for the primary analysis with a human subset, and lost 50.6% at the voluntary three-day recall measure. The education study covers one course and cannot causally identify the chatbot and feedback interface separately. No validated transfer ledger, value standard or proven NEVERYTIME product exists.
THE NEXT POSSIBILITY
Bring the idea
you cannot
leave alone.
For makers, researchers, designers and enterprising minds. A beginning does not need to be finished.
bart@neverytime.nl