# Camera-ready data handoff — October 18, 2026

The original Claude Opus 4.8 outputs have been recovered. New DeepSeek paraphrase and baseline runs are complete and privately available to collaborators. The original July DeepSeek paraphrase outputs remain unrecovered.

The October 1 accepted PDF, submitted source, scored JSONs and published findings remain unchanged. Vittoria Lanzo is handling the sentence-segmentation correction and rescoring. This data handoff does not establish revised results or a new camera-ready manuscript. Distribution version 1.1.1 records the added provenance.

## Private data and access

**All Hugging Face links below require collaborator access.** These raw files are excluded from the public code/data package. Each uploaded file was downloaded again and checked against its SHA-256. The [machine-readable inventory](camera-ready-data-2026-10-18.json) records exact repository revisions, paths, sizes and checksums.

| Dataset | Contents | Verified private revision |
| --- | --- | --- |
| Original Opus 4.8 | 3,000 July generations; 600 self-refine drafts and 600 critiques | [5f9e1bc — collaborator access required](https://huggingface.co/datasets/burstiness-engine/opus48-news-gen/tree/5f9e1bcdb993171dbd586f3c4b8ad1f053d3c0c9/data) |
| New DeepSeek paraphrases | Three variants × 200 scenarios × one iteration: 600 outputs | [69dc8a3 — collaborator access required](https://huggingface.co/datasets/burstiness-engine/deepseek-v4-news-gen/tree/69dc8a31412b45bf06848b3b1f0ec53cfc37af34/paraphrase/new-run-2026-10-18) |
| New DeepSeek baseline | Five modes × 200 scenarios × three iterations: 3,000 final outputs, plus 600 drafts and 600 critiques | [e35f0b8 — collaborator access required](https://huggingface.co/datasets/burstiness-engine/deepseek-v4-news-gen/tree/e35f0b872b489151fc2e430b879499b20092884c/benchmark/new-run-2026-10-18) |

The Opus files are `data/opus-4.8.jsonl` and `data/opus-4.8.batchstate.json`: the original `effort=high`, `max_tokens=32000` run. They are separate from the earlier `effort=max` near-miss run.

The DeepSeek folders contain the selected raw outputs, manifests, attempt histories, repair records and checksums. Existing historical baseline files were not overwritten. Baseline self-refine records retain the draft, critique and rewrite stages.

## DeepSeek rerun disclosure

Both new runs completed on October 18 using the API alias `deepseek-v4-pro`, documented at run time as **DeepSeek-V4-Pro-0813**, the August 13 snapshot. Responses returned the alias; the API did not pin the original July snapshot. These are new generations from a later model snapshot, so differences from the July results cannot be attributed solely to prompt wording or sentence segmentation.

Both runs use the same frozen 200-scenario input revision, `3f27ce645ce4f21aa91bd9b8b52a9ca33214fdee`, with SHA-256 `fa73bffe8b2a312ae0902a90a3b3e72b07de74ddbc8a3dd49134c39377859ed6`. The generator source snapshot is `959853c6aace188975fb907ee6946bcc8032afe2`. The baseline's 4,200 selected stage prompts were checked against the historical benchmark prompt functions with no mismatches. Iterations denote repeated requests; no API seed parameter was supplied.

Thinking was enabled with `reasoning_effort=max` and `max_tokens=32000`. The 32,000-token cap is a disclosed correction from the checked-in 16,000-token configuration. The original July paraphrase run's effective cap could not be established. Frozen settings and provenance accompany both new datasets.

All 600 paraphrase outputs and 3,000 baseline outputs are nonempty and finish with `stop`. The paraphrase package preserves 602 attempts, including two repaired failures. The baseline preserves 4,271 stage attempts: 4,200 selected stages and 71 failed attempts. Repairs retained successful predecessors and the original failure records.

One baseline response is flagged as **off-topic**: `self_refine`, scenario `93`, iteration `1`, rewrite stage (51 words). Its recorded request matches the frozen prompt; the cause is undetermined. The response is retained unchanged, without retry or exclusion. Structural completeness does not imply that every response satisfies its content task.

## Public archive and pending analysis

The submitted eight-family archive, `data/released/hf-datasets.zip`, remains byte-for-byte unchanged. It still excludes Opus and the October DeepSeek runs; its Qwen mirror remains incomplete. The public package includes this note and sanitized metadata, while the newly available raw files remain private.

Sentence boundaries after curly closing quotes need correction and the affected data need rescoring. Until that work is complete, the website's tables and claims describe the October 1 accepted snapshot. No scores or significance claims have been recomputed from these October runs.