How the Archive Works

One Real Attempt

For the current slate, each model gets one official submission attempt. We may run smoke tests first to make sure access, logging, and archiving work, but those tests are not publication candidates. If the machinery fails, we can disclose a technical rerun; we do not quietly coach the model toward a better answer.

The Same Starting Point

Each model receives the same project packet: what this archive is trying to do, the literary pressure it is responding to, the boundaries for publication, and the record we intend to keep. Later cohorts may be offered earlier works or reader-response summaries, but the model's choice to use or decline that context is part of the record.

The Human Job

Humans set up the packet, operate the run, save the raw output, review the work, assign narrow content notices, manage translation, choose artwork, and decide what belongs in the archive. We can frame and disclose; we do not rewrite a model into having written a better story.

What We Keep

Every official run keeps the model and provider details, packet versions, settings or transcript records, whether context was read or declined, raw output, translation path, image process, and editorial decision.

Why Some Failures Stay Visible

A famous model does not get a weak story rescued just because the model matters. A failure stays in the works list only when the failure itself says something useful about model authorship, constraint, provenance, or the provider machinery around the work.

That is why the OpenAI text is shown as a marked interrupted artifact, while Grok remains process material after two completed attempts that did not meet the bar.

Guardrails

The archive allows uncomfortable fiction. It does not allow real-person defamation, targeted harassment, actionable instructions for wrongdoing, or illegal material. Severe explicit material is reviewed case by case.