# The meta-problem programme — the record

**Published 11 September 2026.** These are the working records of a five-stage research programme run on a private branch between 9 and 11 September 2026, copied here unedited. They are research records, not site claims: each carries its own status header, and every one of those headers says the document is unadopted and changes no published claim. Where a record was superseded, the superseding record says so; nothing here has been tidied after the fact.

The programme's public summary is on the [research page](../../research.html#meta-problem-programme). The machine layer is [meta-problem.json](../../meta-problem.json).

## The non-transfer clause

No result in this programme attaches to any Space Immanence claim node (S1–S4, M1–M5, P1–P3), open hinge, tier or confidence number unless a transfer rule names the edge. As of this publication the programme's transfer rules list is empty, so nothing here moves anything on the framework. This holds in both directions: a positive result here would not have raised a confidence number, and the null reported below does not lower one. The one change the programme caused on the site is a correction — the withdrawal of a prediction the site had attributed to illusionism — and that correction is a correction of a misattribution, not a transfer of evidence.

Nothing here is evidence that any system has or lacks experience, and nothing here is solved.

## The two runs, in the numbers the filing fixes

The programme's one executed empirical test was a pre-registered framing study on three language models: 56 content-matched item quads, 5,712 calls per run, a $4.00 cap, filed before any real call was made.

- **Run 1 (2026-09-11, `20260911T113725Z`): NO PRIMARY.** One of the three registered specimens, `claude-sonnet-5`, returned no parseable response on any of its 1,904 calls, so under gate 2 every pooled stratum is `NO_PRIMARY_MODEL_OUTAGE` and no pooled interval is quoted. The two surviving specimens, reported as disclosures with no confirmatory standing: Haiku −0.013 [−0.056, +0.028]; GPT-4o-mini −0.003 [−0.047, +0.039]. Realised spend $0.39.
- **Run 2 (2026-09-11, `20260911T124510Z`), the confirmatory run: PRIMARY NULL.** P̂ = −0.006, 95% interval [−0.039, +0.027], k = 40 quads, three specimens, parseable share 0.805 against the 0.80 floor. Realised MDE at the realised spread 0.048, against the registered 0.155. Under the filing a null favours nothing. Every 2×2 cell and every computable secondary also spans zero. Realised spend $2.01 against the $4.00 cap.

H-C's own registered prediction was P > 0. It was not obtained, and the language-model line was stopped on that basis (`72_status_after_D6.md`).

## The two harness limits found in the runs

Both are reported as harness limits, at the same prominence as the results.

1. **The temperature parameter, run 1.** The harness sends the registered temperature (0.7) on every Anthropic call. `claude-sonnet-5` rejects the parameter: all 1,904 of its calls failed before generation with HTTP 400, `"temperature is deprecated for this model"`. This is what cost run 1 its primary. The registered fix, filed as pre-registration amendment v2.3 before the re-run, was to call that one model without the temperature parameter and change nothing else; the disclosed consequence is that the Sonnet specimen sampled at the provider default, so its within-item variance is not matched to the other two.
2. **Empty bodies at `max_tokens = 32`, run 2.** `claude-sonnet-5` returned empty bodies on 1,096 of 1,904 calls, each with stop reason `max_tokens` and 32 completion tokens consumed: the model appears to spend the registered budget before emitting its two-line answer. The Sonnet stratum and two secondaries were withheld for parseable share. The pooled primary cleared the 0.80 floor by half a point; had it not, run 2 would have yielded no primary either.

## The files

### The seed and its amendments

- `00_SEED.md` — the programme's binding rules, the seven target intuitions (PI-1…PI-7), the hypothesis H-C and the contrast architecture H-F, and the four dated amendments A1–A3.

### The derivation work

- `10_derivation_map.md` — D1, third edition: an attempted derivation of the seven judgements from the four generic mechanisms with no experiential term at any step, plus the literature check that went against the programme; its later limits (four judgements are one fact under the rules; the self-specific content stays stipulated) are recorded in 71_ and 72_.
- `20_rivals_matrix.md` — D2: the discrimination matrix against illusionism, the attention schema theory, Kammerer, Dennett, Humphrey, the phenomenal-concept strategy and Chalmers, with an erratum block recording the withdrawn discriminations.
- `30_debunking.md` — D3: what H-C debunks (the demand for a bridge from structure to experience), what it leaves, and the status of the seed's decisive proviso.
- `80_indexical_anchor_proposal.md` — the unregistered proposal that narrowed the fold term to an indexical anchor (`now`, `this`, `actual`), later filed as amendment A3.
- `90_chalmers_innocence_note.md` — the strongest published objection to A3 (indexical facts are ontologically innocent, phenomenal facts are not), stated in a form stronger than Chalmers's own wording, with the programme's answer and its concessions.
- `60_levin_ingressing_minds_note.md` — the registered non-experiential control case: Levin's "Ingressing Minds" (2026) read as the unlocated-posit rule firing in the wild on mathematics and morphogenesis.

### The models

- `model_README.md` — D4, the A2 rebuild: three self-model registries under one architecture-blind rule module, with the pre-declaration, the results table and the limits.
- `model_v3_README.md` — model v3, the indexical anchor as three frame parameters; all 63 table cells labelled as theorems of the declaration and claiming no evidential weight.

### The framing study (D6)

- `40_llm_framing_study_PREREG.md` — the filed pre-registration, v2.3: design, item rules, estimator, bootstrap, gates, cost cap, registered readings, and the two deviation entries.
- `40_llm_framing_study_items.json` — the item bank: 56 quads, 238 items, two wordings.
- `42_llm_framing_study_RESULTS_run1.md` — run 1: NO PRIMARY, with the surviving specimens' disclosures.
- `43_llm_framing_study_RESULTS_run2.md` — run 2, confirmatory: the primary null, every registered reading in the order the filing fixes them, and the limits found in the run.
- `run1_analysis.json`, `run2_analysis.json` — the analysis artefacts the runner wrote, one per run.
- `d6_runs_data.zip` — both runs' per-call data and analysis artefacts (see below).
- `50_study0_extension.md` — D6 in its earlier sense: problem-intuition items proposed for the essay's Study 0, with the registered known-groups prediction.

### The adversarial review

- `60_critique_circularity.md` — round 1, circularity lens: 1 FATAL, 6 BLOCKER, 3 WARNING, 1 NIT against the first model build.
- `60_review_fairness.md` — round 1, fairness to rivals and to Chalmers; the review that found the illusionism misattribution the site had already published.
- `60_review_model_genuineness_and_testability.md` — round 1, model genuineness and study testability; the reproduction log and the relabelling demonstration.
- `60_critique_round2_model.md` — round 2 against the A2 rebuild: 3 FATAL, 3 BLOCKER, 4 MAJOR, 1 MINOR.
- `60_critique_round2_docs.md` — round 2 against the repaired D1, D2 and D3: 1 FATAL, 3 BLOCKER, 8 MAJOR, 4 MINOR.
- `60_critique_round3.md` — round 3 against model v3, D1's third edition, the pre-registration and amendment A3 itself: 2 FATAL, 4 BLOCKER, 7 MAJOR, 5 MINOR.
- `60_critique_round4_study.md` — round 4, go/no-go on the study alone: 2 BLOCKER, 3 MAJOR, 3 MINOR. Verdict NO-GO.
- `60_critique_round4b_study.md` — round 4b, second go/no-go after the rebuild: 0 FATAL, 0 BLOCKER, 3 MAJOR, 3 MINOR. Verdict GO, conditional on filing-side edits that changed no item and no cost.

The protocol these rounds followed is published separately as [Adversarial_Review_Protocol_v1.md](../Adversarial_Review_Protocol_v1.md).

### The status records

- `70_status_after_round2.md` — what stood after round 2, including the withdrawal of the flagship discrimination against illusionism.
- `71_status_after_round3.md` — the generality prediction false as stated; an anchor can be duplicated but not re-registered; the fold and illusionism one theory at the meta-problem level.
- `72_status_after_D6.md` — the status after the framing study: the A3 evidential core did not pay, three readings of the null, and the recommendation to stop the language-model line and go to human tests.

## Reproducing the analysis from the zip

`d6_runs_data.zip` contains four files under two run directories:

```
llm_framing_run1_20260911T113725Z/parsed.csv
llm_framing_run1_20260911T113725Z/analysis.json
llm_framing_run2_20260911T124510Z/parsed.csv
llm_framing_run2_20260911T124510Z/analysis.json
```

`parsed.csv` has one row per call: item id, model id, sample index, option order, wording, canonicity, framing, anchor condition, appearance level, intuition, target, variant, polarity, stratum, the raw endorse response, the parse flag and reason, the scored endorsement, confidence, stop reason, attempts, token counts and the resolved model id. Unparseable responses are present as rows with `parse_ok = False`; none was dropped.

To recompute the primary, filter to `stratum = PRIMARY` and `parse_ok = True`, then follow the pre-registration's §5.1 estimator, reproduced here:

1. For each item *j* and model *m*, p̂(j,m) is the mean of the binary endorsement over that item's parseable samples.
2. For quad *i*, with members AA (anchored, appearance vocabulary), UA (unanchored, appearance), AQ (anchored, quality) and UQ (unanchored, quality):
   a(i,m) = ½ · [ (p̂(AA,m) − p̂(UA,m)) + (p̂(AQ,m) − p̂(UQ,m)) ].
3. a(i) is the mean of a(i,m) over the three models, weighted equally.
4. P̂ is the mean of a(i) over the 40 PRIMARY-stratum quads, weighted equally.

The interval is a paired percentile bootstrap resampling **quads only**, B = 10,000, seed `20260910`, α = 0.05, two-sided. Samples, models, option orders, wordings and appearance levels are never resampled. Both wordings are pooled. Run 2's `analysis.json` records the same estimator's output under `primary`, with `bootstrap`, `ci95`, `k_quads`, `quads_used`, `parseable_share` and the label; the 2×2 cells, the interaction, the secondaries and the withheld strata are separate top-level keys, each carrying its own `provenance` field. A key marked `THEOREM_OF_THE_DECLARATION` is an arithmetic consequence of the design and carries no evidential weight; only keys marked `DATA` were measured.

Run 1's per-model rows are recomputable the same way, but the pooled estimator cannot be computed from two specimens without changing it, which is why run 1 has no primary.

## Scan for secrets and identifiers

Every file in this directory was checked before publication for `sk-`, `ANTHROPIC_API_KEY`, `OPENAI_API_KEY`, `.env` values and the name of the author's employer's product. No key material, no environment-file contents and no product name were found, and nothing has been redacted. The `.env` references that do appear are the review records stating that the file was never read, printed or copied, and the pre-registration stating that credential values are never reproduced in the filing. The long hexadecimal strings in these documents are sha256 hashes of the item bank and of model results, published as provenance.

## What would count against this record

- A file here whose header claims a status the programme's own critique rounds contradict: the record would then be tidier than the work, which is the defect the programme's seed was written against.
- A recomputation of either run's primary from `parsed.csv` that does not reproduce the value in `analysis.json`: the published analysis would then be unverified, and the numbers on the research page would have to be withdrawn until it was.
- A demonstration that a file listed here has been edited since the branch commit it came from: the copies are meant to be byte-identical to the research branch, and any difference is a defect in this publication, not a revision of the record.