research overview · research september 2026
Research continuation — September and October 2026
These materials separate a new instrument readout, an audit of existing data, and proposals for later work.
| Material | What it supplies | Status |
|---|---|---|
| Bias-corrected plan | Prospective plan: do jackknife- and Miller–Madow-corrected divergences repair the vanishing-gap simulation's failures? Recorded predictions B1–B4. | Plan committed before the code and any result; offline, no model calls. |
| Vanishing-gap results | The registered simulation's results: P1 failed for every variant, P2–P4 held; Fieller sets, a movement interval and the gap's own interval near a vanishing gap, three estimator variants over two floor conventions. | Offline simulation results for human review; no change to the estimators in metrics/noophorics/ and no published result recomputed. |
| Vanishing-gap plan | Prospective plan for a simulation of F* near a vanishing gap: Fieller confidence sets against the present gated, capped report, with recorded predictions. | Plan committed before the simulation code and any result; offline, no model calls. |
| Estimand audit | A run-time tracer binding each interval and p-value the shared inference helpers return to the call and array it came from — inline estimators are reported as untraced — with Problem 13's three incidents as red/green fixtures. | Offline instrument and report for human review; no rule adopted, not a gate. |
| Self-transfer tie sensitivity | Exhaustive retrospective check of the saved result under alternative modal-tie resolutions. | Descriptive audit draft for human review; reviewed 2026-10-07 and cited as ground (c) of retraction 22. The original prediction remains as written. |
| Program-output verifier | Offline validation of a restricted deterministic task format on synthetic fixtures. | Instrument prototype; no E-003 subject data. |
| Qwen register readout | Direct length-and-register counts from blind ratings of archived compositions, with raw requests and responses. | Instrument data; collection status is reported inside. |
| Self-transfer audit | Reconstruction from saved draws and proposed narrower wording for the existing result. | Retrospective audit and correction draft; reviewed 2026-10-07, its wording adopted as retraction 22. |
| Domain-selection design | Candidate task families, a bounded exploratory screen, and separation of selection from confirmation. | Design draft; not registered or run. |
| Methods supplement | Primary-source methods and limitations behind the proposed task families. | Literature review and engineering recommendation. |
The register check follows a prospective plan committed before ratings. Its source compositions already existed; their length outcomes were known. It does not reopen E-001c or measure transfer fidelity.
The self-transfer audit reproduces the arithmetic while distinguishing observed row means from evidence of equivalence. Its suggested correction has not been applied to the original result, journal, theory, or retraction ledger. It was applied on 2026-10-07 as retraction 22, struck in place in the result, the journal and Problem 9. The published prediction remains as written.
The tie-sensitivity audit checks the saved self-transfer draws under every resolution of their observed modal ties. It reports which finite-data descriptions depend on the runner's deterministic tie rule; it does not add a new experiment or alter the published result.
The methods review proposed an offline verifier for a restricted program-output task, now represented by the prototype above. A new E-003 study also needs an explicit operational definition of prior, an independent capability anchor, independently adjudicated keys and an untouched confirmation set. The design draft names these open choices; its proposed budget is not a completed experiment.
The program-output verifier is the first offline instrument prototype from that plan. Its synthetic fixtures check implementation behavior and are not a probe measure, independently adjudicated keys, or an untouched confirmation bank.
Reproducing the evidence
The readout and audit each include their exact reproduction commands. Their checks derive counts from raw records and compare generated reports with the saved files. The corresponding tests deliberately inject incorrect counts, missing responses, mismatched identities, and interrupted collection.
Agent and judge exposure is recorded in EXPOSURE.md. The scientific workflow remains governed by CONTRIBUTING.md.
This document is licensed CC BY 4.0.