The Container Problem

What this is #

Experience is sometimes reported as happening inside something: a field or space of awareness that contents are presented to. Call this the container structure. The programme treats it as an empirical object rather than a metaphysical one, and cuts it into four separately graded, separately killable claims: whether a non-spatial container factor exists in first-person report at all, whether it follows an ordered trajectory across introspective training, whether a specific mechanism can derive it, and whether it dissociates from the body model. Every claim is pre-theoretical — a claim about data structure, not about what the structure means — and each carries its own kill study and its own drawer condition.

What happened to v0.1 matters more than v0.1 did. The first version of this note was circulated to be attacked, not admired, and it was: two rounds of multi-agent adversarial review, six attack lenses, every serious attack independently adjudicated by referees instructed to refute it. No attack killed the programme, but four wounded it at major severity, and the note shrank in every direction that mattered — the mechanism went from a default to a conditional, the trajectory from a discriminating signature to a candidate one, the priority claim from a paragraph to a single fenced sentence, and one prediction was reclassified as an observation the programme must merely be consistent with. This page describes what survived. The whole attack record is published below, in reading order.

The claim, in conditional form #

C1, existence. There is a non-spatial container factor in first-person report, separable from the perceptual spatial frame and from body-boundedness, under the pre-registered machinery of the items document. That is the whole claim: a factor exists. Not what it means, not that it is veridical, not that it matters to any metaphysics.

C3, mechanism — conditional remainder-modelling (CRM). Any system that models by selective illumination produces at every moment a remainder: content that funds the current figure but is not shown as figure. By default, the system does not model this remainder at all, or models it minimally as a point-index — an egocenter, "more, relative to here." The medium encoding — homogeneous, surrounding, empty, in operational senses frozen before any run — is predicted to become optimal only when the remainder itself becomes a target of modelling, under sustained ground-directed attention. Formally: there exists a non-rigged, independently justified task-and-rate regime, characterised by high introspective demand, in which the medium encoding is uniquely optimal against all live rivals and shows all three signatures under frozen tests.

The conditional form is an honest downgrade, and it is worth saying plainly why it was forced. The constraint document's scalar-dominance analysis showed that compression alone can never select the medium: a medium is a scalar plus an index set plus enclosure topology, and every cost function prefers the scalar, so only task demands can pay for structure. The clinical record forced the same repair independently — the blind spot is completed, not experienced as a gap; the neglected hemifield has no experienced edge; inattentional blindness leaves no phenomenal placeholder. The empirical default for the unrepresented is nothing. So the mechanism is an existence claim over regimes, not a universality claim, and a predicted transition is a sharper, more falsifiable object than a default would have been.

Two scope restrictions travel with C3 everywhere. Where ordinary experience does present its remainder, it presents it as heterogeneous and relevance-graded — James's fringe, Gurwitsch's thematic field — so CRM's actual target is the far remainder, the one not presented at all, and the boundary between the two is an open hole that closes at Study C's freeze or C does not run. And C is constitutionally asymmetric: even a maximal positive result exhibits a data structure with three formal properties, and "read phenomenologically as a container" is a bridge posit no optimisation can test. C retains full kill authority over the mechanism and capped, necessary-condition confirm authority. That asymmetry is stated now so a positive result cannot be inflated later.

Naming and provenance #

The mechanism is called CRM — conditional remainder-modelling — throughout, and the naming history is disclosed because it is itself a result of the red-team round. The framing that seeded the programme is Dōgen's, via Okumura's Realizing Genjōkōan: when one side is illuminated, the other is dark. That provenance is aesthetic lineage, disclosed as such, and it is explicitly not evidence.

The placeholder name "darkening" is retired, for a substantive reason rather than a cosmetic one: the clinical record shows the metaphor carried a false implicature. "Dark" suggests the ground is experienced as dark by default, and the finding is that the ground is not experienced at all until attention makes it a target. A mechanism name that pre-loads the phenomenology it is supposed to derive — and that imports the framework's favourite classical ally into every sentence — is a liability the naming-risk audit already priced. No mechanism vocabulary of any register appears in instruments, item wordings, preregistrations, or the machine layer; Study D's manipulation is named "unstructured-field viewing" in all study documents.

The trajectory, demoted #

The sequence — point-observer, then container, then collapse — is a candidate signature: as of this note, a post-hoc arrangement of three known results, held to the same discipline v0.1 applied to the collapse endpoint. It is also, honestly, a two-mechanism trajectory. Embodiment and egocentric modelling give the point; the pure-awareness literature gives the collapse; CRM's sole novel work is the stage-two transition, and the programme claims credit for exactly that and nothing else. Stage one is restated as a prediction the non-meditator free-report arm must confirm, not a result: the existing citations measure folk theory, not phenomenology, and the folk construct — a located presentation venue — already satisfies presented-to structure before any training. Stages one and two are therefore distinguished by format (point-located venue versus unbounded homogeneous medium), and the located-theater outcome is pre-registered as a named alternative, not a failure.

The rivals are stated in their own vocabularies, not the programme's. The curriculum rival: the classical formless-attainment arc makes boundless space the literal object of the fifth absorption — a two-millennia-old syllabus delivered procedurally and cross-linguistically, to which a cross-linguistic vocabulary control is structurally blind. The construction rival (Jaynes): if metaphor acquisition builds the phenomenology, metaphor inventory and experience strata are the same variable. Extended attention-schema and precision-reallocation accounts reach a staged sequence with plausible tuning.

Against all of them, one binding commitment: practitioners of ground-interrogating trainings that never use space instruction — Mahasi-style pure noting, Rinzai kōan work — should still develop the container stage on CRM, and should not on the rivals. Lineage-contamination instruments are mandatory (modern contemplative culture is cosmopolitan), and a historical-text arm of pre-modern reports from space-lean lineages is the cleanest available natural experiment. If only the instructed develop the container stage, the trajectory belongs to pedagogy and the drawer condition fires. And an ordering commitment with it: the within-person claim gets kill authority only from the longitudinal or randomised-training arm. Cross-sectional strata results are reconnaissance for the ordering claim, never verdicts on it — a claim we cannot afford to test is a claim we do not get to make.

The studies #

A — Reconnaissance on an existing dataset

A reanalysis of the MPE-92M questionnaire data, for item harvesting and curve reconnaissance only. It has no kill authority, no confirm authority, and this containment is pinned by a drift test: no page on this site may ever cite it as support for anything. The standing caveat is stated in advance — the instrument's contemplative lexicon (including its Luminosity factor) meeting any light-adjacent theory vocabulary is contamination one step removed, which is one more reason the light-vocabulary mechanism name was retired.

B — The residue instrument

Kill authority over C1 — existence — and over nothing else. B is fielded exclusively through the items document's pre-registered machinery, adopted wholesale: two metaphor-family facets (dative and venue) with cross-family convergence as the signal, since neutral wording is impossible and structure must adjudicate; container items before all Likert content; three-form planned missingness; randomised order arms, with only the container-first arm counting toward the existence verdict if teaching effects exceed threshold; cognitive-interview gates with reading-coded deflation ceilings; jangle-bait and method-factor controls; a matched-abstraction ownership/agency positive control that voids the run if it fails to separate; incremental validity against the boundlessness pools, since a residue factor collinear with unity items is the collapse pole renamed; a pre-registered mediation threshold on the metaphor regression; Korean as the test language with full invariance gating; and a decision table as the sole arbiter of verdicts, including its located-theater and genuine-absence rows. The known-groups meditator arm makes a naive-arm absence interpretable rather than convenient.

C — Formalising the mechanism

Kill authority over C3; confirm authority capped. C runs only behind a frozen constraint specification that must be hash-committed before any formalisation, implementation, or pilot; work preceding the freeze is exploratory by definition. The two-regime structure is the study: regime one (remainder not queried) predicts a point or indexical optimum — that outcome is the theory's prediction, not a failure; regime two (ground-directed attentional demand) predicts the medium. An analytic tier runs before any code, and a failure there kills the mechanism without a training run. One disclosure, because task-ensemble selection is the master degree of freedom: the author has proposed a candidate ecological justification for the task ensemble (attentional steering), and that proposal enters only through the frozen justification document, as hostile input for red-team scrutiny like any other, with its author-origin flagged.

D — Dissociations

Later, and lab-bound: body-model perturbation crossed against unstructured-field viewing, with crossed magnitude predictions at pre-registered power, and its own drawer condition. It is gated behind a literature pass first: if existing Ganzfeld-phenomenology reports already describe frame or container destabilisation under homogeneous fields, that arm is retyped as retrodiction before anything is claimed. Acute perturbation data, including psychedelic datasets, become probative only after B's three-factor separation exists; D then either commits to a psychedelic-adjacent prediction or states why not.

Falsification conditions #

Fixed now, before any data, in the governance document's tightened form. The idea, or the named limb, goes in the drawer if any of the following holds. Machine-checkable criteria for each live in container.json.

1. Existence. No separable third factor emerges in the introspectively experienced strata under the pre-registered discriminant criterion (CFA nested-model comparison, thresholds frozen before fielding), with the item pool frozen after one adversarial revision round; a second null after that single permitted revision fires regardless of item-quality objections. A factor appearing equally strongly in naive strata fires condition 2, not a save. All verdicts route through the items document's decision table, including its voiding gates.
2. Trajectory. The pre-registered ordered contrast (naive below intermediate-trained) fails at the pre-set minimum effect size, or equivalence testing (TOST at a pre-registered SESOI) shows the naive-to-trained slope flat; the collapse limb is a separately pre-registered prediction whose failure kills the collapse limb only. Cross-sectional survivorship is acknowledged as unresolvable in B; the condition fires on B's data anyway, and the ordering claim additionally dies if the longitudinal arm fails its pre-registered contrast. The tradition-contrast result feeds this condition: container-stage development confined to space-instructed lineages fires it.
3. Vocabulary/construction. The cross-linguistic arm is mandatory before C1 exits pilot grade. Pre-registered model comparison: container-factor loading regressed on individual introspective experience versus a frozen language-level container-productivity index built corpus-side before any B data are seen. Fires if the language-level predictor dominates under the pre-set criterion and the individual predictor adds nothing at threshold.
4. Mechanism. Under the frozen two-regime C: fires if the regime-two optimum is robustly non-medium under the constraint document's decision rule, or if any of its seven failure-reporting conditions hold, including formalisation failure, which is published as a result, not absorbed into redesign. Regime one producing a point-index is the theory's prediction, not a failure.
5. Dissociation. D's crossed magnitude predictions fail at pre-registered power.
6. Priority. The novelty case dies if any prior work derives a spatially surrounding phenomenal medium from a representational-completion or remainder-modelling requirement — under a binding search protocol: frozen search strings logged with dates; a full forward-citation chase from Place 1956, Dennett 1991, Jaynes 1976, Metzinger 2020 and 2024, the PCM papers, and Laukkonen & Slagter 2021; and, before any public priority statement, at least one hostile external referee instructed to find the prior derivation. Discovery at any time fires the condition and is logged publicly.
7. Staleness. If B's cognitive-interview gate has not been entered within 24 months of the v0.2 note, the programme's public status reverts to dormant. A drawer that can be indefinitely deferred is a shelf.
8. Public semantics. From the moment the programme goes public, drawering is a published state (status: "drawered" in container.json and state.json, drift-tested), and firing conditions 1 or 3 is additionally logged as negative evidence on S1's weak reading in the framework ledger, per the symmetric transfer commitment.

The neighbour map #

The priority claim, at its full and final narrowness: no prior work derives a homogeneous, surrounding, empty phenomenal medium from the remainder of selection under resource constraints and task demands. Every broader priority sentence from v0.1 is withdrawn. The neighbours, one discriminating sentence each:

• Place (1956) names the phenomenological fallacy — properties of the representation read as properties of an inner field — but supplies no generative mechanism; CRM is a candidate mechanism for the error Place named.
• Dennett (1991) diagnoses the Cartesian Theater as theorist's error, and his "absence of representation is not representation of absence" is the canonical objection CRM absorbs by going conditional; the diagnosis is not a sub-personal derivation of the empty-medium format, which is the gap CRM occupies.
• Jaynes (1976) derives the container from metaphor acquisition; CRM is sub-personal and non-linguistic, and the two are discriminated empirically by the tradition-contrast test and B's cross-family convergence design, not by argument.
• Metzinger — transparency and the tunnel (2003, 2009) derives phenomenal immediacy from unavailability of construction processes and the tunnel from bandwidth; the modelled variable is not the selection remainder.
• Metzinger — MEO and the contraction principle (2020, 2024) is the nearest neighbour: empty-space phenomenology from the system modelling its own epistemic capacity. CRM differs on the modelled variable (remainder of selection, not capacity) and on scope (the ordinary, untrained container, not MPE); a candidate discriminator — alertness covariation versus ground-directed attentional demand at fixed alertness — is offered, not established.
• PCM (Rudrauf, Williford, Friston) formally derives the perspectival frame under projective geometry — content-structured, perspectival, neither homogeneous nor empty, and precisely the first factor B controls for; PCM-style projective encodings sit inside C's frozen encoding space, so the comparison is run, not asserted.
• Graziano (AST) derives an "ethereal, vaguely localisable" awareness-attribution from a control model of attention; v0.1's claim that the schema has no reason to reshape with contemplative training is retracted as unverified and likely false, and an extended AST is scored as a live rival in its own vocabulary.
• Laukkonen & Slagter (2021) derive a staged contemplative continuum ending in pure awareness from precision-weighting; they do not derive the container format at the intermediate stage, where precision accounts predict opacity revealing heterogeneous process structure and CRM predicts a homogeneous empty medium.
• Limanowski & Friston (2018) formalise the trigger operation itself — introspective attention rendering transparent states opaque; what remains CRM's own is the claim about the content of what opacity reveals for the far remainder.
• Husserl, James, Gurwitsch describe the structured near-remainder — fringe, theme, thematic field, margin; CRM's target is the far remainder, and the handoff between the two is a named open hole.

Open holes, named so they cannot be quietly forgotten #

1. The switching boundary. The fringe/far-remainder boundary and handoff are unspecified; they close at C's freeze or C does not run.
2. The MEO discriminator. Alertness versus attentional-demand covariation is a candidate needing development — ideally by someone instructed to prove convergence, not difference.
3. The task-ensemble justification. Unwritten; until it is frozen, C cannot start, and the author's attentional-steering proposal inside it is unreviewed hostile input.
4. The tradition-contrast operationalisation. Lineage purity criteria, contamination instruments, and the historical-text arm's scope are design work not yet done.
5. The bridge problem, permanently. C can never test "read phenomenologically as", and every public statement about C must carry that asymmetry.

The origin story #

This programme was red-teamed before it was published. The v0.1 working note was committed privately and handed to an adversarial review — two rounds, six attack lenses, twenty-two raw attacks merged to eighteen, every kill- or wound-graded attack independently adjudicated by referees instructed to refute it, and every load-bearing literature claim web-verified or marked unverified. The attack produced three constructive companion documents along the way: a hostile item-level draft-and-kill pass on the residue instrument, a frozen constraint specification for the mechanism study, and the governance quarantine this page now lives inside. Then the note was rewritten as v0.2 around what the attack forced. The whole record is public, in reading order:

1. The Container Problem, v0.1 — the original working note, preserved as the record, including the claims the red team later falsified.
2. Red-team report, v0.1 — the attack itself: verdicts, severities, and the three-way convergence that forced the mechanism to go conditional.
3. Response to the author — the red team's cover letter: what died, what survived, and how narrow the survival is.
4. The Container Problem, v0.2 — the canonical note this page renders, rebuilt around the attack's repairs.
5. Study B items document — the residue instrument at item level, every item drafted to be killed, with the decision table that makes drawer condition 1 fireable.
6. Study C constraint specification — the adversarially fixed encoding space, cost functions, and decision rules, binding at hash-commit.
7. Governance document — the quarantine design: the firewall analysis, the machine layer's shape, and the drift tests that pin this page to it.