| language: | |
| - en | |
| tags: | |
| - evaluation | |
| - writing-feedback | |
| # E1 frozen historical writing references | |
| Public evaluation-only exports: verbal FCP Bootstrap step15 and RFT+mean-reward | |
| GRPO step15, seeds 0/7/42/60, 1000 unique test prompts each. Temperature 0, | |
| top_p 1, top_k -1, max_tokens 4096. All IDs and original prompts were matched | |
| to the current original test manifest. Text and row identities are preserved; | |
| environment paths and server logs are omitted. See provenance.json for source hashes. | |
| These are historical outputs, not newly trained results, not training data, and | |
| not checkpoint weights. They permit rescoring and paired comparisons without | |
| access to the original development machine. Use the same independent evaluator | |
| for candidate and reference outputs. Exact Table 3 result/statistic correspondence | |
| must still be checked before quoting paper numbers. | |