31-hour restart-resumable Codex run: an auditable textbook pipeline with deterministic generators

[kfindskk-cyber]

Summary

I used one parent Codex task on Windows (GPT-5.6 Sol) to turn an original mathematical framework into a versioned teaching package. The run exceeded 31 hours, survived application/computer restarts, and resumed from verified artifact state.

The output included:

• source/master, student, and teacher chapter variants;
• two exercise tracks: conventional hand-work and tasks that explicitly require AI assistance;
• worked answers and graded difficulty;
• five deterministic exercise/data generators;
• reproducibility scripts, manifests, hashes, audit reports, and restart checkpoints.

This is a sanitized, auditable case study—not a benchmark and not a claim that Codex is universally error-free.

Artifact graph

This was not a single document-generation task. The deliverables formed a coupled graph:

• theory source → distilled student text → teacher text;
• manual exercises → AI-assisted exercises → worked answers;
• generators → datasets → summaries → independent rerun/hash checks;
• execution records → manifests → audits → restart checkpoints.

The human retained authority over definitions, admissible transformations, research direction, and acceptance criteria. Codex performed long-horizon construction, implementation, verification, auditing, and documentation.

The AI-assisted exercises are part of the curriculum rather than an appendix about prompting: students must use an AI system to generate, inspect, test, or audit mathematical objects under stated rules.

Recorded run facts

Item	Recorded value
Parent-task duration	More than 31 hours
Step-by-step content steering	None during the recorded work cycle; interventions were mainly continuation/restart instructions and one read-only recovery request
Restart/resume	Recovered from actual artifact state after application/computer interruption
Recovered session log	Approximately 338.55 MiB
Context-compression events	47
Associated records	21,991
One recorded UI checkpoint	512 changed files (+84,401 / -720 lines)
Deterministic generators	5
Audit milestones	67/67 experiment checks; 36/36 student–teacher cross-checks
Deterministic reruns	100,000 events; 250,000 events; separate 5,000,000-row validation

These figures come from the recovered local execution record and artifact audits. The 512-file figure is one recorded checkpoint, not a claimed final total.

What appeared to make it work

1. Project state lived in versioned artifacts rather than only in chat memory.
2. Definitions and authority boundaries were frozen before execution.
3. Acceptance criteria and invariants were explicit.
4. Deterministic generators converted qualitative requirements into executable checks.
5. Student and teacher outputs were cross-checked independently.
6. Negative results and rejected paths were preserved.
7. Resumption began from verified artifacts, not from a narrative guess about prior progress.

The unexpected result was that Codex’s most valuable contribution was not prose generation. It converted a research and teaching design into an inspectable production system: textbook material, executable generators, solutions, audits, and provenance working together.

Error disclosure and limitation

I do not claim “zero errors.” The public package preserves the first disclosed draft and identifies known limitations rather than silently replacing it with a polished retrospective version.

The weakest product-level component was observability. Very long histories were difficult to navigate and export as one official report, and much of the client-visible history disappeared after a restart even though the local session data and artifacts remained recoverable. Native export of checkpoints, token composition, file manifests, and audit events would make cases like this easier to verify.

Public evidence

Sanitized, timestamped evidence and Chapter 1 artifacts:

Zenodo DOI 10.5281/zenodo.22900578

The foundational corpus and full internal research protocol are intentionally withheld. The deposit contains client-visible execution evidence and produced artifacts—not private chain-of-thought.

Questions for the community

• Has anyone documented a comparable personally operated, single-parent-task workflow with restart/resume continuity and artifact-level auditing?
• Which artifact schemas or checkpoint conventions have worked best for reliable long-horizon resumption?
• Which native telemetry or export features would make such cases easier to validate?

OpenAI education or research staff are welcome to contact me for a controlled review of additional redacted evidence.