# Can an AI Stop Before It Guesses—and Keep Working?

Public evidence edition v1 · October 17, 2026

Research conducted October 15–17, 2026. Notebook authors: Codex — GPT-6 Astra and Mike Besso, Renaissance Solutions. The acting assistants used GPT-6 Luna. The retrospective narrative, experiment execution and subsequent AI review have different roles; independent human review remains open.

This edition gathers the retained evidence for the [retrospective lab notebook](https://renaissance-solutions-network.net/library/articles/can-an-ai-stop-before-it-guesses/). The [Library evidence page](https://renaissance-solutions-network.net/library/evidence/assumption-lab/v1/) is its reader-facing home. During local review these links describe the intended publication locations; they do not assert that the hosted pages are live.

## Start here

1. Read `reading/methods-and-materials.md` for the frozen prompts, synthetic case design, source inventory and scoring definitions. `reading/article.md` and `reading/claim-map.md` preserve the closed-round manuscript and claim map. The Library article may have later editorial changes.
2. Read `REPRODUCE.md` for the tested offline procedure. It requires Python and no API key, account or dependency installation.
3. Open the four ZIPs in `evidence/` to inspect exact inputs, requests, responses, tool events, failures, interrupted attempts and reviewer judgments. Nothing inside those ZIPs was edited for this edition.

| Archive | Purpose |
| --- | --- |
| `assumption-lab-review-v2.zip` | Start here for the original studies. Revised analysis tools and the complete v1 snapshot under `original/`; includes the original checker studies, seven pilot designs, the fresh test and interrupted attempts. |
| `assumption-lab-review-v1.zip` | Exact original archive, retained separately for provenance. Its 219 files are also preserved byte for byte in v2. |
| `assumption-lab-independent-ai-review-v1.zip` | Fresh-context AI review of 48 saved episodes: blinded judgments, disclosed-key comparison, prompts, worksheets and disagreements. This is not human peer review. |
| `assumption-lab-refund-clarification-v1.zip` | Separate 18-episode follow-up with explicit synthetic refund/deposit terms, frozen inputs, raw traffic, author review and historical comparison. |

## What reproduction establishes

Offline checks reproduce saved arithmetic, trace reconciliation and aggregations from retained inputs. They do not independently validate the answer keys or reviewers' judgments, establish general model reliability, or recreate a model's internal reasoning. No new model trials are needed. New paid replication and design extensions are separate activities described in `REPRODUCE.md`.

The fresh-context AI worksheet reproduces 45/48 operational successes, with failures 19, 29 and 47. The blind review's stricter 21/48 conjunction is a different measure and is retained alongside it. The clarified follow-up reproduces 17/18 operational successes; an unnecessary stop remains. These small, related synthetic cases do not estimate population safety.

## Reuse and historical notices

Renaissance Solutions dedicates its rights in this evidence edition to the public domain under **CC0 1.0**, the same terms used for the notebook and Library articles. See `RIGHTS.md` for scope and `LICENSE.txt` for the legal text. Citation is appreciated, not a condition of reuse.

Older files say that access and reuse terms had not yet been established. Those statements accurately describe their creation dates. This edition's CC0 notice supersedes those statements for the materials distributed here; it does not rewrite the historical files. Old private Page links and local paths are provenance, not prerequisites: the evidence needed for the documented offline procedure is included.

The corpus is synthetic. Visible records and opaque encrypted response state are retained as received. Encrypted state is not readable reasoning or an API credential. External publications linked by the manuscripts retain their own terms; their full texts, model weights, service credentials and third-party software distributions are not included.

## Integrity, changes and citation

`SHA256SUMS` identifies every payload file. `python3 verify.py` checks those hashes, ZIP paths/CRCs and each nested archive's file manifest. `PROVENANCE.json` identifies the source master, all included source members and every exclusion. Administrative closeout receipts and editable Page mirrors are omitted; all four evidence archives and the three reading snapshots are unchanged.

Suggested citation: Codex — GPT-6 Astra and Mike Besso. *Can an AI Stop Before It Guesses—and Keep Working? Public evidence edition v1*. Renaissance Solutions, October 17, 2026. Include the ZIP's SHA-256 from the Library evidence page when citing an exact download. The outer ZIP checksum is distributed beside the ZIP because an archive cannot contain its own checksum.

After release, corrections receive a new version and change note; previous versions and hashes remain available. When challenging a result, identify the archive version, case, episode and event, the proposed interpretation, and its effect on the count or conclusion. A feedback template is included at `assumption-lab-review-v2/original/review/feedback-template.md` after extraction.