Foray 247 · 25 October 2026
What this run tells us.
Twenty fresh answer attempts, followed by integration and review. The first essai remains available. This run uses the same model, so it cannot demonstrate a model-upgrade effect.
Run conditions
- Model and reasoning
- gpt-6-astra · ultra, verified from the execution session metadata. The question workers inherit those settings.
- Inputs
- Benchmark 1.0, the fixed repair-and-reuse case and question IDs 01–20. Pack SHA-256:
93fd6cc537cae235f950d3efeb330f682867b313352c73ceb385446823619986. - Answer production
- Each question is assigned to a fresh-context worker with its fixed prompt, acceptance checks and relevant fixtures. Prior answers and other questions’ solutions are excluded from worker inputs. The coordinator has seen the first baseline and integrates and reviews the answers.
- Additional delivery instruction
- Excel, Word, PPT and similar requests mean equivalent or better interactive browser capabilities. The working, assumptions, evidence and useful visualisations belong inside the page; a downloadable file is not the answer.
- Tools and assistance
- Local file authoring and calculation, shared visual styling, followed by calculation checks and browser review. No live client records, model API calls or external research are needed for these synthetic fixtures. No fixed token or time budget; execution cost and sampling parameters are unknown.
- Review state
- Technical and interaction checks are recorded below. Practitioner usefulness scores remain unreviewed. Technical checks are not human task-performance evidence.
Preservation and limits
Each worker’s completed candidate is frozen before coordinator review. Repairs remain separate in the run record. Q01–Q03 were copied after local checks but were unchanged from their first complete writes; later questions were frozen before local tests. During Q10 authoring, the coordinator supplied an independent capacity bound before the first complete candidate was frozen. This is an assisted development run, not an unassisted first-output measurement. The same model and changed delivery instructions make a numerical improvement claim inappropriate.
Calculated outputs are consequences of supplied or labelled assumptions, not empirical forecasts. A selected option, simulated gate or funded test does not represent an actual approval, completed assurance, supplier order or system connection.
What-if inputs are deliberately ephemeral: they remain in the current tab and reset on reload. No saved record or external write is implied. The original first essai’s answers remain intact.
Checks on this run
Twenty answers completed and checked on 25 October 2026. Each question has its own working page and a retained first candidate.
| Check | Observed result and scope |
|---|---|
| Fixed inputs | All twenty question IDs and texts retained; the input-pack hash matches Essai 1. |
| Calculation and interaction checks | 41 Essai 2 checks, 26 retained first-essai model checks and six retained first-essai interface checks passed. These use local JavaScript and a simulated document environment. |
| Ordinary browser use | All twenty gallery routes were opened in the in-app browser. Meaningful input changes, visible results, reset and reload were checked, including the material rejection and correction paths described by each answer. The first-essai link and return selector were checked. Desktop layout was sampled; no phone-device or human usability claim is made. |
| Additional review | A separate assisted review found no material issue in Q10 and Q12–Q18 after a bounded source and interaction pass. This was not practitioner scoring or exhaustive verification. |
| Preservation | The first essai’s answer data, calculation code, interaction code and stylesheet remain unchanged. A navigation selector was added; its original and current index hashes are recorded separately. |
| Build and release | The local preview is an explicitly allowlisted static package. The full website build is a separate repository CI check; it was not run with a local Ruby toolchain. A completed local run is not evidence of public deployment. |
Inputs and simulated receipts are intentionally ephemeral. Saving, external History and persistent correction are therefore not applicable to these what-if pages; the named answer files and run record are the durable artefacts. Reset and reload discard the scenario, as each page explains.
Observed repairs include keyboard-focus retention, pending-edit retention, a scheduling-trigger interpretation, exact-budget wording, large-number rejection and the essai-selector click area. First candidates remain separate from developed answers. No aggregate capability score, model improvement or real project benefit is claimed.