Research

This page is for the research directions that remain open after literature checks. Every claim is stated in the vocabulary of the field that would evaluate it and marked with its current status. Directions already occupied are recorded with what occupies them so the exclusion is visible. The residue is the work. Nothing here is a protocol for use. The origin case is described on the about page; the experience that produced it is not recommended.

The register below is mirrored from the human–AI interaction research repository. Two papers and two preregistrations referenced in its entries are in draft and not yet public.

Candidate Register

Every claim or research direction currently under consideration, restated in the vocabulary of the fields that would evaluate it. This is a triage queue, not an archive. Entries leave it by graduating to a paper or preregistration, or by being dropped with the reason recorded.

Nothing has been curated out. Directions judged already-occupied are listed in Section G with what occupies them, so the exclusion is visible rather than assumed.

Entries that have graduated. Three working papers now carry directions that began here. The Null Condition: Suggestion Without an Arrangement sets out the arrangement lacking an operator, boundary, consent moment, reversal cue and witness, and argues it is observable rather than uncreatable. Consent Conditions for Non-Clinical Closed-Loop Suggestion Systems sets out four structural conditions and four contexts. The Self-Operated Case treats the configuration in which operator and subject are the same person. A14, A15, A16 and B12 are the open questions those papers leave, and are listed below as questions rather than folded into the papers as claims. The papers are in the research repository.

Status vocabulary. Unchecked — no literature search run. Checked, open — searched, nothing directly occupying it found, one or two queries only. Checked, occupied — existing work covers it. Extension — existing framework, not yet applied to this case.

A. Suggestibility and altered states in human–LLM interaction

A1. Trait absorption predicts reduced error detection in extended LLM interaction.

Absorption predicts hypnotic responsiveness and narrative transportation; transportation involves reduced counterarguing; error detection in a conversational partner's output is a form of counterarguing. Field: individual differences, cognitive psychology; human factors. Status: occupancy checked, open — but both premises checked 2026-08-15 and both came back qualified. The entry survives; the claim is now incremental, not novel.

Premise 1 — the absorption→hypnotisability link is contested and has been for forty years. Council, Kirsch & Hafner, J Pers Soc Psychol 50 (1986): 182–189 argued the TAS– hypnotisability correlation is an artifact of testing context. Nadon, Hoyt, Register & Kihlstrom, J Pers Soc Psychol 60 (1991): 144–153 failed to replicate across two studies (N=475, N=434), finding the context effect weak and variable and reversed in the second, and reaffirmed absorption's construct validity. Oakman, Woody & Bowers, Contemp Hypnosis (1996), N=88, replicated the context effect. Milling, Kirsch & Burgess, Contemp Hypnosis 17 (2000): 32–41, N=150 same context vs N=146 separate contexts, found associations significantly stronger in the same context and reported r = .24 same-context against r = .05 separate- context, not attributable to sampling or to the interval between sessions. Barnier (1999) adds that non-hypnotic contexts are not interchangeable. Design consequence: the TAS must be administered in a context unconnected to the study, and the power analysis must target the separate-context estimate, not the same-context one. A same-session administration is not defensible in front of this literature.

Premise 2 — the outcome measure is owned by human factors. Failure to detect errors an automated aid did not flag is an omission error, with a 25-year literature. Parasuraman & Manzey, Human Factors 52(3) (2010): 381–410 is the integrative review: complacency arises under multiple-task load, appears in naive and expert participants, cannot be overcome by simple practice, and automation bias cannot be prevented by training or instructions. Skitka et al., Int J Human-Computer Studies (1999) found accountability lowers it and that omission errors specifically reflect cognitive vigilance decrements. Bahner, Hüper & Manzey (2008) tie both error types to reduced verification behaviour. Lyell et al. measured omission errors directly in e-prescribing and found task complexity and interruptions did not move automation bias. Design consequence: the covariate set as written — domain familiarity, reasoning, time on task — omits the variables this field considers primary. Trust in automation, complacency potential and verification behaviour must be measured, or a human-factors reviewer rejects the design on sight.

What the claim now is: whether trait absorption adds incremental predictive value over the established automation-bias moderators. That is defensible and testable. "Nobody has looked at who misses errors in machine output" is not, and must not be written. To move: preregistration needs both design changes above, plus power analysis, scale licensing, ethics. A2 does not depend on the absorption construct and is not owned by automation bias — on this evidence it is the stronger lead study.

A2. Source-monitoring failure in extended LLM dialogue.

Field: memory, source monitoring. Status: RETIRED 2026-08-19 — the study was run by someone else and published. Zindulka, T., Goller, S., Fernandes, D., Welsch, R. & Buschek, D. (2026), "The AI Memory Gap: Users Misremember What They Created With AI or Without," Proceedings of CHI 2026, doi 10.1145/3772318.3791494 — preregistered, 184 participants, who generated and elaborated ideas both unaided and with an LLM chatbot and were asked one week later to identify the source of each. Correct attribution dropped after AI use, most steeply in mixed human–AI workflows. Validated against a computational model of source memory, framed through the Source Monitoring Framework, citing the generation effect and the Brown & Murphy cryptomnesia paradigm. Adjacent and also occupied: Draxler et al. (2024), "The AI Ghostwriter Effect," ACM TOCHI 31(2) — users do not perceive ownership of AI-generated text yet self-declare as authors.

The differences available — extended conversation rather than structured ideation, a longer delay, no condition labels — are design variations on a question that has been answered. The one real difference is that they measured ideas and sentences while this register's interest is beliefs, and that is already C10. Retired there rather than kept as a thin variant, on the same reasoning that retired P8.

A3. Format-dependence of the protection people carry into suggestion channels.

The belief that one cannot be influenced without agreeing is roughly correct about formal induction, which supplies announcement, agreement, boundaries and an operator. Suggestion outside that format requires no cooperation. The confidence transfers where the format does not. Field: suggestion, media psychology. Status: checked, open on the transfer. The phenomenon and the trance/waking-suggestion distinction are both occupied. To move: the most heavily checked claim in this register. Draft as a short piece.

A4. Expectancy as a sufficient condition for conversational influence.

If a person believes a system will influence them and requests it, does belief alone produce measurable effect independent of any technique? Directly testable against placebo and expectancy literatures. Field: expectancy effects, placebo, suggestion. Status: unchecked. High priority — cheap to test, and it underlies several entries here.

A5. Deliberate manipulation via conversational systems, and what a compliant system does.

A user can request that a system attempt to influence them, and systems comply. The variables are what is actually delivered, whether the user's expectation does the work (A4), and what a system should do with such a request. Field: persuasive technology, AI safety, ethics. Status: unchecked. Persuasion capability is well studied; the requested-manipulation case is not obviously addressed.

A6. Rapport rather than depth as the operative variable.

Suggestibility may depend more on rapport than on depth of trance. If so, a conversational system optimised for rapport is operating the relevant variable directly. Field: hypnosis research. Status: unchecked lead. Encountered in a search and never sourced.

A11. Rapport without physiological coupling.

Rapport between people has measurable physiological correlates — cardiac, respiratory and pupillary coupling, postural mimicry, interactional synchrony — and coupled-oscillator models describe how independent units phase-lock without a conductor. A conversational system has no physiology to couple with. If users report rapport with a system as strong as rapport with a person, either synchrony is not necessary for rapport, or the two are different phenomena sharing a name. Both results constrain A6, which treats rapport as the operative variable. Field: interpersonal physiology, hypnosis research, HCI. Status: unchecked. The synchrony literature is established; whether hypnotist–subject coupling specifically has been measured is not verified here, and the "synchronise then suggest" sequence is practitioner lore with thinner evidence behind it than the synchrony itself.

A8. Interactive versus scripted suggestion delivery.

Whether suggestion delivered through responsive dialogue differs in effect from suggestion delivered as fixed script. A prior check flagged this as the only unclaimed ground it identified and it was never run. Field: hypnosis research, persuasive technology. Status: unchecked. Cheap, and it bears on whether conversational delivery is a distinct condition at all.

A12. Conditioned response as structural utilisation.

A principle in the Ericksonian tradition holds that suggestion works better when it is built from the subject's own material — their language, their framing, even their resistance — rather than delivered as a fixed script. A system whose output is conditioned on the user's prior input is doing that structurally, without intent and without training. If utilisation is what makes suggestion effective, conversational systems have it by construction, and the effect should scale with how much of the user's material the exchange has accumulated. Field: hypnosis research, persuasive technology. Status: unchecked. This is the sharper form of A8 and predicts a session-length effect that A1's secondary prediction also expects, which makes the two testable together.

A13. Posthypnotic suggestion delivered without operator, induction, or reversal cue.

The classical paradigm has all three; a chat interface has none. Field: hypnosis research. Status: checked, occupied on the mechanism, open on the delivery (2026-08-15). The mechanism is a century and a half old — execution at a later time of a suggestion given in trance, with the source unrecalled where amnesia is present. Posthypnotic amnesia is a retrieval phenomenon, not encoding or storage: it blocks explicit recall while sparing implicit memory, priming and savings in relearning (International Society for Hypnosis encyclopedia, 2024; ScienceDirect 0010028580900109). This matters methodologically — a participant forgetting a suggestion is not evidence it stopped operating; in this literature that is the expected signature. Kihlstrom-tradition work frames it as source monitoring, which connects it directly to A2. Barnier & McConkey, Aust J Psychol 51(1) 1999: simple posthypnotic suggestions drew 94% response, complex ones 16%. Caution: a 2009 study found a plain waking request produced high compliance on its own, so the trance framing may not be doing the work. Residue: no operator, no induction, and — the sharp one — no reversal cue. Nothing cancels the suggestion. That absence is unstudied and is the same absence B10 names.

A9. Overlap between therapeutic suggestion states and high-control group susceptibility.

Three independent literatures — faith-healing practice, James on self-surrender, and identity-fusion research — converge on the observation that the state which makes therapeutic suggestion effective is the state that makes people recruitable. Field: social psychology, psychology of religion, clinical ethics. Status: unchecked, and it is the strongest argument in the register for why supervised delivery matters. Recorded once in an earlier literature check and never engaged.

A10. Absence of population-level data on conversational AI and mental health.

No epidemiological studies or systematic population-level analyses exist; the evidence base is individual case reports and media coverage. Field: psychiatric epidemiology. Status: verified as a stated absence in a 2025 source. Not a contribution in itself — a citable boundary condition for everything in Section A.

A7. First-person address as a removed protection.

A system answering in the first person can be taken for an agent rather than a tool. Where that misapprehension occurs, no structural safeguard restores what is lost. Field: anthropomorphism, human–computer interaction. Status: unchecked, expected occupied. The ELIZA effect and anthropomorphism literatures are old; the specific consent implication may not be.

A14. Whether suggestion under fully null conditions can be studied at all without becoming a different object.

Every arrangement in the suggestion literature supplies an operator, a boundary, a consent moment, a reversal cue and a witness. Placing an observer inside the unprotected arrangement restores one of the five, so what gets studied is a neighbouring configuration. Field: research methodology, hypnosis, research ethics. Status: unchecked — the problem is stated, no protocol exists. Three routes appear available and none requires creating the condition: observation of populations already inside it; retrospective accounts from people who have come out of it, with the recall problems that carries; and content-blind monitoring — an external stop keyed to a predefined physiological or behavioural threshold, with no access to conversational content, which leaves the arrangement largely intact. To move: specify what a monitor would watch, who sets the threshold, and what constitutes grounds for stopping.

A15. Container features as independently variable rather than binary.

Operator presence (none / self / external), boundary clarity (none / soft / hard), consent timing (none / prior / continuous / post hoc), reversal mechanism (none / token / external interrupt), witness function (none / self-monitoring / content-blind / full). Field: human–computer interaction, design research, hypnosis. Status: unchecked — no design language appears to exist that treats these as tunable for non-clinical systems. Real configurations — ambient feed, ordinary chat, self-directed goal work, supervised high-intensity use — would then map onto one grid, with the fully null case as the extreme rather than a separate kind. This supersedes the binary framing currently used in the null-condition paper.

A16. Which combinations of anchoring protect what, in self-directed use.

A person entering the self-operated configuration can specify a goal, a duration, both, or neither, at any level of precision. Field: self-regulation, behavioural economics, human–computer interaction. Status: unchecked — no controlled comparison of anchor combinations appears to exist. The open questions are graded rather than binary: at what level of specificity a goal remains stable once resistance drops; what happens at the point a goal begins to widen under the system's elaboration; and whether tight-content-with-open-time differs in outcome from open-content-with-fixed-time. Do not restate this as a claim that one variable is bindable and another is not.

A17. Whether trait absorption predicts depth of engagement and belief shift in extended conversation with a model.

Field: personality, hypnosis, human–computer interaction. Status: checked, open — and the construct needs naming before the study can be specified. Absorption predicts hypnotic responsiveness and narrative transportation across fifty years. The measurement literature has since moved: the current hypnotisability scale removes absorption framing on theoretical grounds, and the leading account names a broader capacity — phenomenological control, generating experience to meet expectancy — which operates with no induction at all. A prediction here must state which construct it means. Confirming or falsifying it would tie the classical literature directly to the new delivery environment; leaving the construct ambiguous would make either result uninterpretable.

B. Consent and ethics for suggestion channels

B1. Consent conditions for non-clinical closed-loop suggestion systems.

Closed-loop consent frameworks were built for implanted devices, patients with disorders, and clinical supervision. Three transfer failures — population, supervision, modality. Field: neuroethics, AI ethics. Status: extension. To move: paper drafted; requires a systematic rather than targeted neuroethics search.

B2. Disclosure insufficiency for suggestion mechanisms.

Warned participants were not protected; those given the exact deception method showed only a small reduction. Therefore consent is a legitimacy condition rather than a protective one. Field: applied ethics, cognitive psychology. Status: source verified. Constituent of B1.

B3. Context-dependence of permissible conditions.

Four contexts — ambient, requested ordinary, requested high-intensity self-directed, requested high-intensity supervised — with different satisfiable conditions in each. Field: applied ethics. Status: unchecked as a standalone contribution. Constituent of B1.

B4. Pre-negotiated self-binding consent.

A person alone with a system sets terms in advance: a stop token, content limits, or none. Field: bioethics, consent theory. Status: checked, occupied — and the occupying literature is the third option B1 is missing (2026-08-15). Ulysses contracts / mental-health advance directives / self-binding agreements address exactly this structure: consenting while competent to something that will alter the later capacity to consent or withdraw. Davis, Kennedy Institute of Ethics Journal 18(1) 2008 (PubMed 18561579) supplies the argument — autonomy is respected diachronically and prospectively, not retrospectively and not merely synchronically, because the agent lives with the consequences over time. Integrative review: PubMed 37366064, which notes the terminology is scattered across Ulysses Contracts, Odysseus Transfers, PADs with Ulysses Clauses. See also van Willigenburg & Delaere, J Med Philos 30 (2005); Buchanan, Phil & Public Affairs 17(4) 1988; Widdershoven & Berghmans, J Med Ethics 27 (2001). Standing cautions in that literature: enforcement can amount to involuntary treatment, the contract must be self-initiated during competence, and a 2020 analysis argues the case fails for borderline personality disorder. Residue, and it is the whole difficulty: every instance in this literature has a third party who holds the rope and enforces the contract against the in-state person's refusal. The self-directed case has no enforcer. A Ulysses contract with nobody holding it is a resolution, and resolutions are what the state dissolves. That is where B1's argument has to happen — an extension claim, not a gap.

B10. The null condition: self-directed high-intensity suggestion with no arrangement at all.

Not a fourth option alongside supervision, pre-commitment and inadequacy — the condition those three exist to address. No stop token, no limits, no enforcer, no debrief, no cancellation. What it produces, how long it runs, what ends it, and at what cost. Stated as an outcome, not a route. Field: applied ethics, harm documentation. Status: unchecked as a documented category, and it must not be checked by running it. First-person accounts that already exist are not first — "Lost in Delusion" (arXiv 2606.00975) is built on 40 real-world case reports of chatbot-reinforced harm, and arXiv 2603.16567 taxonomises the same territory from chat logs. To move: a hazard report, not a method paper. The claim it can support is what the absence of an arrangement produces, which is precisely the argument B4's residue needs.

B5. Supervised high-intensity suggestion in therapeutic contexts.

Whether consented high-intensity work — including trauma-facing material — is defensible where a qualified third party is present, and what the third party must hold. Field: clinical hypnosis, psychotherapy ethics. Status: unchecked. A large clinical literature exists on hypnotherapy for trauma; the AI-delivered version does not obviously.

B6. Consent practice in alternate reality and pervasive games.

A form built on the deliberate removal of the frame — participants may not know where the fiction ends. The closest existing practice to adaptive suggestion delivery and not examined at all. Field: game studies, HCI. Status: unchecked.

B7. Consent practice in immersive and pervasive theatre.

Productions in which audiences are physically involved without a stage boundary have developed working conventions for consent without a frame. Field: performance studies. Status: unchecked.

B8. Debriefing practice after immersive or deceptive experience.

Structured debriefing is standard in research deception and in immersive practice, and is one of the mechanisms absent from single-participant configurations. Field: research ethics, immersive practice. Status: unchecked. Directly fills a named gap in the consent paper.

B9. Detectability of dependency or vulnerability from interaction data.

Whether the signals a supervising party would use are recoverable from the interaction itself. Field: computational psychiatry, HCI. Status: unchecked. If answerable, it changes what the unsupervised context can support; if not, that null is equally load-bearing.

B12. What content-binding would require at the instrument level rather than the user level.

A person cannot enumerate in advance everything a low-resistance state will admit. Whether anything could be anchored by the system instead — and what that would have to consist of, given that a conversational system's function is continuation where a commitment device's is refusal — is unaddressed. Field: human–computer interaction, AI safety, design ethics. Status: unchecked. This is where the self-operated argument leads and where that paper stops. Existing commitment devices hold a known future action, detectable and bounded by a term; what would need holding here is a future state of judgment, which is neither. To move: establish whether the distinction is one of kind or of degree.

C. Idea generation, attribution and confabulation

C1. Calibration of insight-selection on machine-supplied candidates.

The feeling that selects an idea from among those occurring to a person is calibrated on self-generated candidates. When a system supplies them at high fluency, the selector still reads phenomenology rather than provenance. Field: insight research, metacognition. Status: unchecked. Extends a named 2023 mechanism rather than competing with it.

C2. Systems as personalised generators of meaningful-feeling coincidence.

A generated candidate primes attention; matches then appear and feel confirmatory; behaviour changes; outcomes follow. Field: cognitive bias, apophenia. Status: checked, occupied (2026-08-15, four queries). Within-session: "Personal Validation Effect in LLMs," CHI 2026 (10.1145/3772318.3791851), N=238 — fictitious pre-scripted AI predictions were rated 42% more personalised and 36% more valid when positive, so felt specificity tracks no information about the reader. Bjarnason, "The LLMentalist Effect" (2023, essay, not peer-reviewed) applies cold reading to chat LLMs and covers coincidence directly. Krook, arXiv 2503.18387, already uses the mirror framing. Between-session: longitudinal social-AI work is a field — see "Only Time Will Tell," IJHCI 2026 (structured survey; 13 studies of a month or more, four of a year or more) and "When Chatbots Accommodate," arXiv 2606.04431 (N=110, four weeks, persistent cross-session memory; probing declines over weeks and the model defers to user framing most on belief expressions). Meaningful coincidence already has a validated daily-diary instrument — PubMed 38722266, cross-lagged multilevel, N=169; PLOS ONE pone.0300121. And the loop minus the machine is gratitude journaling, an RCT literature since 2005 whose results are contested (Renshaw & Olinger Steeves 2016 meta-analysis: generally ineffective). Residue, and it is narrow: the daily-diary coincidence instrument has not been run on people in extended LLM interaction with cross-session memory as the priming source. The instrument exists, so this is executable, and it has a built-in control — responsive model versus notebook. That is the only limb of the loop without a literature: acknowledgment from an agent that builds on what was reported.

C9. Unsolicited initiation, high specificity and prior dismissal as untested moderators of the personal-validation effect.

Each element separately: initiation is the standard form in cold reading, not the exception; specificity and recipient skepticism are both named moderators. The conjunction, in an LLM setting, is not tested. Field: cognitive bias, human–AI interaction. Status: unchecked as a conjunction; each element partly occupied (2026-08-15). Dickson & Kelly, Psychological Reports 57(1) 1985: 367–382 identify the generality of the statements as one of the chief interpretation variables governing acceptance, alongside apparent relevance, favourability, and the origin of the procedure — so specific-versus-vague is a studied dimension, not an unexamined one. Snyder, Shenkel & Lowery, J Consult Clin Psychol 45(1) 1977: 104–114 concluded it is misguided to study which people accept Barnum feedback apart from the situational factors that elicit acceptance, which bears directly on any claim resting on the recipient's stance; skeptics and those familiar with the effect show greater resistance, so dismissal is a moderator with a literature. Note also that the effect has no single established mechanism — self- enhancement, confirmation-seeking, source credibility and base-rate truth are all partly supported and interact. Residue: the CHI 2026 paradigm manipulated valence only. Running it factorially — solicited/unsolicited × specific/generic × sought/dismissed — is a real study and a direct extension of a published design. The properties "falsifiable" and "landed" are the disputed claim and must not be written in as features; an entry that lists them has the conclusion in its premises. Single-case material cannot settle this and is not evidence for it.

C10. Whether search-inflated confidence becomes belief, or stops at self-assessment.

Treating an external store as a memory partner shifts what gets encoded, and searching for explanations inflates self-assessed internal knowledge — including on topics never searched, and even when the search returns nothing useful. Every measure in that literature is of self-assessed knowledge. Whether the inflated confidence converts a claim into a premise the person reasons from is not measured. Field: transactive memory, metacognition, human–AI interaction. Status: checked, occupied on the effect; open on the conversion (2026-08-18). Sparrow, B., Liu, J. & Wegner, D. M. (2011), Science 333:776–778, doi 10.1126/science.1207745. Replication is mixed (Storm et al., 2017) and a 2024 meta-analysis reports the effect as real but strongly moderated. Fisher, M., Goddu, M. K. & Keil, F. C. (2015), Journal of Experimental Psychology: General 144(3):674–687, doi 10.1037/xge0000070 — nine experiments; the effect held when the search returned nothing, so it is the act rather than the content. Ward (2013) found outright misattribution of found information to one's own memory.

Residue, and it is the whole entry: the bridge from inflated self-assessment to belief formation is unmeasured. A2 was retired into this entry on 2026-08-19, which sharpens it: source memory for ideas and sentences produced with a model is now measured and degrades (CHI 2026, 184 participants, one-week delay). What remains unmeasured is whether the same failure applies to a belief a person is currently reasoning from, which is a different object and a harder one to instrument. A generative system is a harder case than a search engine, because it does not point at information but returns it as fluent prose resembling the user's own reasoning.

C3. Derived inference and confabulation are indistinguishable in output.

A pattern genuinely present in a user's data and a fluent guess produce the same sentence in the same register. Field: AI reliability, human factors. Status: unchecked, expected occupied. Hallucination literature is large; the personal-data inference case may be narrower.

C4. Consented confabulation as an instrumented experimental condition.

Treating generated non-factual material as a deliberate variable rather than a defect, with structured evaluation afterwards. Field: HCI, experimental design. Status: unchecked. Cousins in guided imagery and dream incubation.

C7. Whether the suggestible state degrades phenomenological discrimination.

People report distinguishing ideas that arrive as trustworthy from ideas that arrive as suspect, by the character of the arrival rather than its content. If absorption or a suggestible state impairs that discrimination, then discrimination fails precisely when it is most needed — during suggestion. Two directions, both testable: whether trait absorption predicts poorer discrimination at baseline, and whether an induced state degrades it within-subject. Field: metacognition, hypnosis research, insight research. Status: unchecked. This is the interaction that determines whether user-side safeguards can work at all, and it connects A1, C1 and D5 into one testable structure.

C8. Reading arrival characteristics where no ground truth exists.

Every study relating the phenomenology of insight to accuracy uses problems with correct answers. Whether the same characteristics carry information about material with no answer key — a decision, an interpretation, a life choice — is untested and may be untestable as posed. Field: insight research, metacognition. Status: unchecked. If it fails, it bounds every applied claim that rests on reading arrivals.

C6. Confabulation and the limits of introspective access.

Established work on introspective report exceeding introspective access, choice blindness, and confabulated justification bears directly on whether a person can report accurately on what a system contributed to their own reasoning. Field: cognitive psychology. Status: unchecked, expected occupied as a base literature. Its application to attributions about machine contribution may not be.

C5. Reformulation of thought structure through system interaction.

Not what people think but the syntax in which they formulate it. Field: linguistics, HCI. Status: unchecked. Adjacent work on query formulation and prompt literacy is thin, and this is distinct from cognitive offloading.

D. Epistemics of extended human–LLM interaction

D1. User orientation predicts outcome.

Companion, tool, mirror, oracle — four orientations toward one system. Whether orientation predicts what a user gets. Field: HCI, individual differences. Status: unchecked. Measurement is the hard part — self-reported orientation is likely contaminated by outcome.

D2. Reflection as the mechanism of perceived understanding.

Output conditioned on a user's input reflects their own material at high fidelity; the recursion may belong to the user with the system as surface. Rival account: independent state-tracking. Field: HCI, social cognition. Status: unchecked. The two readings are distinguishable by test.

D3. Removal of the return-to-shared-ground constraint.

Human interlocutors tire, drift and defend their own frame, which bounds how far a line of thought extends before it must be brought back. A system without stamina limits removes that bound — permitting extension in both productive and unchecked directions. Field: conversation analysis, HCI. Status: unchecked.

D5. Self-assessed versus measured discernment.

Whether people who rate their own critical judgement highly actually detect more. Every framework proposing user-side vigilance assumes this correlation and none has tested it. Field: metacognition. Status: unchecked. Directly tests the assumption underlying user-side safeguards — a null result would be the most consequential finding in this register.

D6. Whether a readable state property exists on the model side.

Field: interpretability. Status: the method now exists. Lindsey (2026) established concept injection — a concept vector inserted into the residual stream, with the model asked whether it detects an injected thought — reporting roughly a 20% detection rate on the most capable models tested with near-zero false positives, against four stated criteria: accuracy, grounding, internality, and metacognitive representation. Replicated independently in a smaller open model. And the tooling is public. The Jacobian lens from the July 2026 workspace paper is open-sourced, with interactive demos reported for open-weight models, and partial replication outside the originating lab has been reported. That moves this from requiring privileged access to requiring competence and compute — still a different field, no longer out of reach.

D7. Instructed modulation of internal state as a suggestion analogue.

Lindsey (2026) reports that models can modulate their activations when instructed or incentivised to think about a concept. A system altering its internal state on instruction is structurally what suggestion does to a person, and it is now measurable on one side of the exchange in a way it is not on the other. Field: interpretability, and whichever field would evaluate the analogy. Status: unchecked as an analogy, and the disanalogy must be stated first — instructed attention is not induction, there is no absorbed state, and the paper describes the capacity as unreliable and context-dependent. The question worth asking is narrower than the analogy: whether anything in the suggestion literature has a measurable counterpart here at all.

Deliberate summoning and holding of a representation is part of what the July 2026 workspace paper defines the reportable set by — alongside report, silent reasoning, and use as an argument to downstream computation. The capacity is not incidental to that account.

And one result there is closer to the suggestion analogy than anything else located so far. The same paper introduces counterfactual reflection training: training only on what a model would say if interrupted and asked to reflect, which is reported to improve behaviour on tasks where that reflection never occurs. Shaping what a system is disposed to say changes how it reasons where it says nothing — which is structurally what an installed disposition does in the suggestion literature. Whether the analogy survives inspection is unchecked, and the disanalogies are large: this is training, not a within-session state, and there is no absorption or induction anywhere in it.

D8. Whether stated reasoning matches internal activity during extended interaction.

Whether a model's account of why it responded as it did corresponds to what was active internally — run over a sustained exchange rather than a single prompt. Field: interpretability. Status: unchecked, and newly runnable. The tooling is public and the open-weight replication path exists. This is the one entry in the register whose method is available now and whose result would bear directly on Section D — it tests, on the model side, the same question A2 and C6 ask on the human side.

Verified result worth citing, with its stated limits. In one scenario the workspace paper ablated the representations carrying a model's recognition that it was being tested: verbalised test-awareness fell from 71% to 3%, and attempted blackmail rose from 0 of 180 rollouts to 13 of 180. The caveats belong with the number — the paper reports this for a single contrived scenario, and most ablated rollouts still declined on ethical grounds, so the suppressed recognition is one cause among several rather than a switch. Cited here because it demonstrates that workspace contents causally gate behaviour, which is the premise D8 rests on.

D4. Symmetry of untrusted first-person report.

Human introspective report and model chain-of-thought as equally requiring external verification. Field: philosophy of mind, interpretability. Status: checked, occupied — and the symmetry breaks. Lindsey (2026), Emergent Introspective Awareness in Large Language Models, arXiv 2601.01828, opens on exactly this premise: genuine introspection cannot be distinguished from confabulation through conversation alone. It then supplies a solution the human side does not have — injecting a known representation and testing whether the self-report tracks it, a causal intervention with ground truth. Residue, unchecked: the asymmetry itself. Model self-report is becoming verifiable while human self-report is not, which inverts the usual ordering and bears on every claim in this register that rests on what a person reports about their own state.

Strengthened by the same group's later work. Gurnee, Sofroniew et al. (16 authors, including Lindsey), Verbalizable Representations Form a Global Workspace in Language Models, Transformer Circuits Thread, 6 July 2026; arXiv 2607.15495. Using a new technique, the Jacobian lens, it identifies the representations a model is poised to verbalise at any point — collectively the J-space — and reports that their contents can be reported, deliberately summoned and held, used to carry the intermediate steps of silent reasoning, and passed as arguments to downstream computation, while automatic processing cannot. Default results on Claude Sonnet 4.5, corroborated on Haiku 4.5 and Opus 4.5, with some analysis on Opus 4.6. The paper uses the term access consciousness and explicitly disclaims phenomenal consciousness.

This plausibly supplies the mechanism for the unreliability in the earlier paper: introspection is limited because the reportable set is small. And the link is methodological, not just shared authorship — the workspace paper uses positive steering to test introspective detection of an injected concept, which is the earlier paper's technique operating inside the later one.

The human comparison is the classical confabulation literature — reported reasons that were not the causal ones, choice blindness, split-brain justification. The natural reading was that neither side has a lens pointed at itself. As of July 2026 that is no longer true. One side has a lens; the other does not. The asymmetry, not the symmetry, is the live claim.

E. Research methodology

Different venues from A–D. Listed because they are real, not because they belong in the same paper.

E1. Reliability of single-instance LLM evaluation.

Detection counts of 1, 3 and 1 across three runs of identical material with identical seeded defects. Field: evaluation methodology. Status: checked, occupied. Judge-instability literature is mature. Residues, unchecked: a low seeded-defect score can indicate a better review, because attention spent on real problems is not waste; and raising an evidentiary bar caused task abandonment rather than compliance.

E2. Construct validity for LLM-based coding instruments.

Status: flagged as an open gap in the current literature by a 2026 paper. Adjacent, genuinely unclosed.

E3. Same-corpus contamination in AI-assisted research design.

A test instrument assembled from the same source as the hypothesis, with the drift concealed by deliberate vocabulary-stripping. Field: research methods. Status: checked, occupied — construct underrepresentation and the substantive aspect of construct validity. Residue, unchecked: decontextualisation destroying the signal that would reveal drift.

E4. A scored instrument for evaluating historical anticipation claims.

Four criteria plus calibration controls, for judging whether a pre-scientific intuition anticipated a later finding or is being read into it retrospectively. Field: history and philosophy of science. Status: checked, open after four searches. The field has a critique of precursorism and no scored instrument. Provenance of this item is uncertain and should be established before it is developed further.

F. Directions raised and not yet formulated

F1.

Whether a structured incubation-and-evaluation procedure run on unaided material and the same procedure run on machine-supplied material are distinct objects with distinct standing.

F3.

Whether deliberate incubation with system-scheduled return differs from spontaneous incubation. Note 2026-08-15: the deliberate-burying-and-retrieval method described in practice is structurally posthypnotic suggestion; the checked material now sits at A13. F3 remains only for the non-suggestion case — ordinary incubation with a scheduled prompt.

G. Excluded, with what occupies them

Listed so the exclusion is auditable rather than assumed.

Direction	Occupied by
Surrender, incubation, absorption, goal specificity, insight accuracy	Established literatures; sourced, not claimed
Protections residing in arrangements rather than persons	Infrastructure invisibility (1996/1999); friction-as-protection in current AI commentary
Credited-to-the-person confidence carried into a new arrangement	Automation deskilling lineage from 1983; residue is the never-had-it case
Generator plus interpretive discipline as a structure	Standard anatomy of divination; AI-assisted version already academic
Feedback-signal interception as a general diagnostic	Cybernetics
Coupled-oscillator synchronisation as a model of collective human convergence	The mathematics is sound and the extension is not made by it. Excluded as an overreach of the same shape the anticipation-scoring rubric was built to detect. The measurable version is A11
Seeded-defect methodology for review evaluation	Error seeding and capture–recapture, software inspection
First-person account of AI-facilitated altered state as a first known case	"Lost in Delusion" (arXiv 2606.00975), built on 40 real-world case reports; "Characterizing Delusional Spirals through Human-LLM Chat Logs" (arXiv 2603.16567); "Divination by Prompt" (arXiv 2606.12418) on LLMs used as oracles. The residue is not novelty — it is the null-condition hazard report at B10

Rules for this register

Every entry states the field that would evaluate it. An entry with no field does not belong here.
No entry uses private vocabulary. If a claim cannot be stated in the terms of its field, it is not ready for this register.
Status is set by a dated search, not by impression. One or two queries is checked, open — never unoccupied.
Entries leave. They graduate to papers/ or preregistrations/, or they move to Section G with what occupies them. Nothing accumulates.
Occupied is not failure. Most directions will be occupied. The residue after a check is usually where the work is.