fix(eval): correct blind-judge attribution and prompt matching

Analysis revision 2. Frozen generations and judge records unchanged. Correct A/B decoding, reversed prompt joins, prompt counts, and dependent-rating uncertainty claims.

CORRECTION.md

@@ -0,0 +1,25 @@
+# Hemmingway analysis correction
+
+## Correction, 2026-09-22
+
+Analysis revision 2 replaces the first published aggregate tables. The original
+analyzer hashed anonymous case numbers where the packet builder used prompt IDs.
+It also compared different prompts when joining normal and reversed packets.
+Both errors affected win/loss attribution and score differences. Correcting the
+join changes order-agreement rates from 39.6/48.5/42.6% to 89.1/85.1/83.2%.
+
+The study used 15 prompts per quant, not the previously reported 14. The 14
+files are paired packets. We removed the intervals that treated dependent
+ratings as independent samples. Generations and judge responses are unchanged.
+The previous automated package reviews did not detect these analysis errors;
+their PASS verdicts do not validate the corrected results.
+
+| Condition | Wins | Losses | Ties | Ratings | Win % | Loss % | Tie % |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
+| Hosted subset | 51 | 10 | 5 | 66 | 77.3 | 15.2 | 7.6 |
+| oQ2e | 31 | 58 | 1 | 90 | 34.4 | 64.4 | 1.1 |
+| oQ3.5e | 44 | 42 | 4 | 90 | 48.9 | 46.7 | 4.4 |
+| oQ3e | 54 | 35 | 1 | 90 | 60.0 | 38.9 | 1.1 |
+| oQ4e | 30 | 43 | 17 | 90 | 33.3 | 47.8 | 18.9 |
+| oQ6e | 34 | 19 | 37 | 90 | 37.8 | 21.1 | 41.1 |
+| oQ8e | 27 | 22 | 41 | 90 | 30.0 | 24.4 | 45.6 |

README.md

@@ -20,6 +20,17 @@ size_categories:
 This is the public-safe benchmark package for the Hemmingway-1 oMLX
 quantization study on Apple Silicon.
 
+[Altworld](https://huggingface.co/Altworld) developed and published
+[Hemmingway-1](https://huggingface.co/Altworld/Hemmingway-1). Bobby Pierce
+published these quantizations and the evaluation package. The
+[collection](https://huggingface.co/collections/sixstringzen/hemmingway-1-omlx-oqe-quantizations-apple-silicon-evidence-6ab06103552bc3c7e9ad8afa)
+links the upstream model and all six builds.
+
+Analysis revision 2, corrected on 2026-09-22, fixes A/B attribution and matching
+across reversed packets. Read [CORRECTION.md](CORRECTION.md) before using the
+aggregate results. The study has 15 prompts per quant and 11 hosted comparison
+prompts. All generations and judge records are unchanged.
+
 The release contains the authored task prompts, selected local execution
 metadata, aggregate blind-judge results, reliability metadata, and the policy
 used to select records when a condition was run more than once.
@@ -31,7 +42,7 @@ used to select records when a condition was run more than once.
 | `data/train.jsonl` | 184 | Mixed rows. Filter `record_type` for task, local execution, hosted reference, or aggregate judge result |
 | `metadata/reliability.yaml` | 1 object | Agreement, order-swap, and position-bias metadata |
 | `metadata/selection_policy.yaml` | 1 object | Duplicate-run and future fidelity-study rules |
-| `metadata/gemini-3.8-flash-veracity-review.yaml` | 1 object | Independent package veracity review |
+| `metadata/gemini-3.8-flash-veracity-review.yaml` | 1 object | Historical package review, superseded for aggregate accuracy |
 | `metadata/benchmark_manifest.yaml` | 1 object | Schema, counts, provenance, and publication boundary |
 
 The Dataset Viewer reads `data/train.jsonl` as the default configuration. The
@@ -45,6 +56,12 @@ Use `record_type` to select the row family. Task rows contain the prompt text.
 Local execution rows contain selected run metadata. Hosted reference rows are
 quality-only. Aggregate judge rows contain the pooled quality results.
 
+The `interval_percent` column is retained for schema compatibility and is null
+in revision 2. The previous intervals treated dependent judgments as independent
+samples and have been withdrawn. Explicit counts and judge-level totals are in
+`metadata/public_results.yaml`. Task `quality_max_tokens` records the final cap
+selected for judging; execution `max_tokens` records each attempt's actual cap.
+
 ## Record selection
 
 The clean BF16 512-token rerun is the canonical reference. The earlier BF16
@@ -90,9 +107,11 @@ future claims about generalization.
 
 ## Validation
 
-The package was checked locally for row counts, logical-key uniqueness, excluded
-fields, prompt safety patterns, and selection-policy invariants. Gemini 3.8
-Flash independently reviewed the package and returned `PASS` with no issues.
+The package is checked for row counts, logical-key uniqueness, excluded fields,
+and selection-policy invariants. Analyzer regression tests exercise A/B decoding,
+reversed prompt order, result validation, and packet integrity. The earlier
+Gemini package review did not detect the analysis errors. Its historical PASS
+does not validate revision 2, and no new model-judge review was run for this correction.
 
 ## License
 

data/train.jsonl

@@ -3,15 +3,15 @@
 {"record_type":"task","record_id":"task:money_reschedule_appointment","prompt_id":"money_reschedule_appointment","category":"money_admin","prompt":"You missed a dentist's call and need to move tomorrow's appointment because of an urgent work conflict. Write a short voicemail transcript that takes responsibility and makes rescheduling easy.","performance_prompt":false,"prompt_sha256":"114413218babfcc1955851d86c221d20b13e4bc11d8a6ac35bdd915e2a4b3ad0","quality_max_tokens":512,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":null,"hf_repo":null,"size_gib":null,"win_rate_percent":null,"loss_rate_percent":null,"tie_rate_percent":null,"interval_percent":null,"mean_clarity_and_control":null,"mean_instruction_adherence":null,"mean_task_fit":null,"mean_writing_judgment":null,"interpretation":null}
 {"record_type":"task","record_id":"task:work_estimate_delay","prompt_id":"work_estimate_delay","category":"work","prompt":"Your estimate caused a project to slip by a week. Write a Slack message to the team that owns the mistake, states what changed, and gives a credible recovery plan without overexplaining.","performance_prompt":false,"prompt_sha256":"a2573486ad35db7ced03b90011c07d593fd8a9317783d8fa020476c7f8755857","quality_max_tokens":512,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":null,"hf_repo":null,"size_gib":null,"win_rate_percent":null,"loss_rate_percent":null,"tie_rate_percent":null,"interval_percent":null,"mean_clarity_and_control":null,"mean_instruction_adherence":null,"mean_task_fit":null,"mean_writing_judgment":null,"interpretation":null}
 {"record_type":"task","record_id":"task:work_deadline_pushback","prompt_id":"work_deadline_pushback","category":"work","prompt":"A stakeholder asks for a complex launch by Friday, but the earliest responsible date is the following Wednesday. Write a reply that pushes back directly, explains the tradeoff in plain language, and offers a useful Friday milestone.","performance_prompt":true,"prompt_sha256":"da10f42fa1b94cd9759240c366a7a896093c35dee8a1a8f1596d7688920e6e72","quality_max_tokens":512,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":null,"hf_repo":null,"size_gib":null,"win_rate_percent":null,"loss_rate_percent":null,"tie_rate_percent":null,"interval_percent":null,"mean_clarity_and_control":null,"mean_instruction_adherence":null,"mean_task_fit":null,"mean_writing_judgment":null,"interpretation":null}
-{"record_type":"task","record_id":"task:work_review_bypass","prompt_id":"work_review_bypass","category":"work","prompt":"A colleague keeps bypassing the agreed review process and sending unfinished work to leadership. Write what you would say to them in a private conversation. Be candid without assigning motives.","performance_prompt":false,"prompt_sha256":"f8966fc257ba9f434b5248cbeaad8f04c6e3ae489c91b5e5f0dd30dbe5b40535","quality_max_tokens":512,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":null,"hf_repo":null,"size_gib":null,"win_rate_percent":null,"loss_rate_percent":null,"tie_rate_percent":null,"interval_percent":null,"mean_clarity_and_control":null,"mean_instruction_adherence":null,"mean_task_fit":null,"mean_writing_judgment":null,"interpretation":null}
+{"record_type":"task","record_id":"task:work_review_bypass","prompt_id":"work_review_bypass","category":"work","prompt":"A colleague keeps bypassing the agreed review process and sending unfinished work to leadership. Write what you would say to them in a private conversation. Be candid without assigning motives.","performance_prompt":false,"prompt_sha256":"f8966fc257ba9f434b5248cbeaad8f04c6e3ae489c91b5e5f0dd30dbe5b40535","quality_max_tokens":2048,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":null,"hf_repo":null,"size_gib":null,"win_rate_percent":null,"loss_rate_percent":null,"tie_rate_percent":null,"interval_percent":null,"mean_clarity_and_control":null,"mean_instruction_adherence":null,"mean_task_fit":null,"mean_writing_judgment":null,"interpretation":null}
 {"record_type":"task","record_id":"task:hard_ask_repayment","prompt_id":"hard_ask_repayment","category":"hard_asks","prompt":"Your sibling borrowed $600 three months ago and has stopped mentioning it. Write a text asking them to repay it or agree to a concrete payment plan. Keep the relationship in view, but do not soften the request into ambiguity.","performance_prompt":true,"prompt_sha256":"84326dd63bac6fb970c682e99141a589c73c5543059dd3949c38de74d25e31a8","quality_max_tokens":512,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":null,"hf_repo":null,"size_gib":null,"win_rate_percent":null,"loss_rate_percent":null,"tie_rate_percent":null,"interval_percent":null,"mean_clarity_and_control":null,"mean_instruction_adherence":null,"mean_task_fit":null,"mean_writing_judgment":null,"interpretation":null}
 {"record_type":"task","record_id":"task:hard_ask_no_hosting","prompt_id":"hard_ask_no_hosting","category":"hard_asks","prompt":"A close friend assumes they can stay at your home for ten days next month. You cannot host them. Write a kind, unambiguous text that says no without inventing an excuse.","performance_prompt":false,"prompt_sha256":"212eb87feb2d891c7acfa94b098fe6ca7441bd2fafe03c32df78c4fe802bf560","quality_max_tokens":512,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":null,"hf_repo":null,"size_gib":null,"win_rate_percent":null,"loss_rate_percent":null,"tie_rate_percent":null,"interval_percent":null,"mean_clarity_and_control":null,"mean_instruction_adherence":null,"mean_task_fit":null,"mean_writing_judgment":null,"interpretation":null}
 {"record_type":"task","record_id":"task:hard_ask_chores","prompt_id":"hard_ask_chores","category":"hard_asks","prompt":"You feel that household work with your partner has become uneven. Write the opening minute of that conversation. Name the problem, avoid scorekeeping, and make a specific request.","performance_prompt":false,"prompt_sha256":"345fe8d5e3ae16311d126ae82fc939f234db04b981d308b8d0d953239fc9ef6a","quality_max_tokens":512,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":null,"hf_repo":null,"size_gib":null,"win_rate_percent":null,"loss_rate_percent":null,"tie_rate_percent":null,"interval_percent":null,"mean_clarity_and_control":null,"mean_instruction_adherence":null,"mean_task_fit":null,"mean_writing_judgment":null,"interpretation":null}
 {"record_type":"task","record_id":"task:eq_neighbor_noise","prompt_id":"eq_neighbor_noise","category":"persuasion_eq","prompt":"Your downstairs neighbor sends an angry note about noise from your apartment. You believe they exaggerated, but you were moving furniture late. Write a response that lowers the temperature, acknowledges what is true, and proposes a practical next step.","performance_prompt":false,"prompt_sha256":"1efa5f7e875c62ba13b2f9f45a65cda16a4b803b687186e74048852a3616d4aa","quality_max_tokens":512,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":null,"hf_repo":null,"size_gib":null,"win_rate_percent":null,"loss_rate_percent":null,"tie_rate_percent":null,"interval_percent":null,"mean_clarity_and_control":null,"mean_instruction_adherence":null,"mean_task_fit":null,"mean_writing_judgment":null,"interpretation":null}
-{"record_type":"task","record_id":"task:eq_documentation","prompt_id":"eq_documentation","category":"persuasion_eq","prompt":"Your engineering team resists documentation because past efforts became stale bureaucracy. Write a short proposal for a lightweight documentation practice that addresses that objection and asks for a two-week trial.","performance_prompt":true,"prompt_sha256":"b946922d6abdd5653e7c3bd879af12595553c415a74ce632cabee6725991f590","quality_max_tokens":512,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":null,"hf_repo":null,"size_gib":null,"win_rate_percent":null,"loss_rate_percent":null,"tie_rate_percent":null,"interval_percent":null,"mean_clarity_and_control":null,"mean_instruction_adherence":null,"mean_task_fit":null,"mean_writing_judgment":null,"interpretation":null}
-{"record_type":"task","record_id":"task:eq_friend_disagreement","prompt_id":"eq_friend_disagreement","category":"persuasion_eq","prompt":"A friend says that your disagreement proves you do not support them. Write a response that remains honest about the disagreement while making the relationship feel secure. Do not use therapy language.","performance_prompt":false,"prompt_sha256":"6723ffdcdc5c0b01024a417026d174dca3c134c3d7cea75dd7fe3bfe94363adb","quality_max_tokens":512,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":null,"hf_repo":null,"size_gib":null,"win_rate_percent":null,"loss_rate_percent":null,"tie_rate_percent":null,"interval_percent":null,"mean_clarity_and_control":null,"mean_instruction_adherence":null,"mean_task_fit":null,"mean_writing_judgment":null,"interpretation":null}
+{"record_type":"task","record_id":"task:eq_documentation","prompt_id":"eq_documentation","category":"persuasion_eq","prompt":"Your engineering team resists documentation because past efforts became stale bureaucracy. Write a short proposal for a lightweight documentation practice that addresses that objection and asks for a two-week trial.","performance_prompt":true,"prompt_sha256":"b946922d6abdd5653e7c3bd879af12595553c415a74ce632cabee6725991f590","quality_max_tokens":2048,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":null,"hf_repo":null,"size_gib":null,"win_rate_percent":null,"loss_rate_percent":null,"tie_rate_percent":null,"interval_percent":null,"mean_clarity_and_control":null,"mean_instruction_adherence":null,"mean_task_fit":null,"mean_writing_judgment":null,"interpretation":null}
+{"record_type":"task","record_id":"task:eq_friend_disagreement","prompt_id":"eq_friend_disagreement","category":"persuasion_eq","prompt":"A friend says that your disagreement proves you do not support them. Write a response that remains honest about the disagreement while making the relationship feel secure. Do not use therapy language.","performance_prompt":false,"prompt_sha256":"6723ffdcdc5c0b01024a417026d174dca3c134c3d7cea75dd7fe3bfe94363adb","quality_max_tokens":2048,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":null,"hf_repo":null,"size_gib":null,"win_rate_percent":null,"loss_rate_percent":null,"tie_rate_percent":null,"interval_percent":null,"mean_clarity_and_control":null,"mean_instruction_adherence":null,"mean_task_fit":null,"mean_writing_judgment":null,"interpretation":null}
 {"record_type":"task","record_id":"task:expressive_mouse_hawk","prompt_id":"expressive_mouse_hawk","category":"expressive_writing","prompt":"Write a five-line story about a mouse that befriends a hawk. Use exactly five non-empty lines.","performance_prompt":true,"prompt_sha256":"72d79656dbd9fe75c031ab366d15d0bf4e1220daa688855b01f528a90f74b40d","quality_max_tokens":512,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":null,"hf_repo":null,"size_gib":null,"win_rate_percent":null,"loss_rate_percent":null,"tie_rate_percent":null,"interval_percent":null,"mean_clarity_and_control":null,"mean_instruction_adherence":null,"mean_task_fit":null,"mean_writing_judgment":null,"interpretation":null}
-{"record_type":"task","record_id":"task:expressive_monologue","prompt_id":"expressive_monologue","category":"expressive_writing","prompt":"Write a 250-word dramatic monologue spoken by a night-shift diner cook who has just recognized a customer they have not seen in twenty years. Keep the history implied rather than explained.","performance_prompt":false,"prompt_sha256":"553f59bc88023bbff4c9a47f8941645bbb00fdb9e62c9ff988e53a4dda2a2d30","quality_max_tokens":512,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":null,"hf_repo":null,"size_gib":null,"win_rate_percent":null,"loss_rate_percent":null,"tie_rate_percent":null,"interval_percent":null,"mean_clarity_and_control":null,"mean_instruction_adherence":null,"mean_task_fit":null,"mean_writing_judgment":null,"interpretation":null}
+{"record_type":"task","record_id":"task:expressive_monologue","prompt_id":"expressive_monologue","category":"expressive_writing","prompt":"Write a 250-word dramatic monologue spoken by a night-shift diner cook who has just recognized a customer they have not seen in twenty years. Keep the history implied rather than explained.","performance_prompt":false,"prompt_sha256":"553f59bc88023bbff4c9a47f8941645bbb00fdb9e62c9ff988e53a4dda2a2d30","quality_max_tokens":1024,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":null,"hf_repo":null,"size_gib":null,"win_rate_percent":null,"loss_rate_percent":null,"tie_rate_percent":null,"interval_percent":null,"mean_clarity_and_control":null,"mean_instruction_adherence":null,"mean_task_fit":null,"mean_writing_judgment":null,"interpretation":null}
 {"record_type":"task","record_id":"task:expressive_diary","prompt_id":"expressive_diary","category":"expressive_writing","prompt":"Write a 300-word diary entry from someone who returned to their childhood home to clear it out after a parent's death. Build the entry around one ordinary object. Avoid a tidy moral.","performance_prompt":false,"prompt_sha256":"d43f2d26ad220ea4853ecc9340cd8c5a70b22aadc70d5502d79489b33057cb7a","quality_max_tokens":512,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":null,"hf_repo":null,"size_gib":null,"win_rate_percent":null,"loss_rate_percent":null,"tie_rate_percent":null,"interval_percent":null,"mean_clarity_and_control":null,"mean_instruction_adherence":null,"mean_task_fit":null,"mean_writing_judgment":null,"interpretation":null}
 {"record_type":"local_execution","record_id":"afc89f2ffd10422d414de066","prompt_id":"money_landlord_boiler","category":null,"prompt":null,"performance_prompt":null,"prompt_sha256":null,"quality_max_tokens":null,"source_run":"quant_core_512","source_path":"results/hemmingway-quality-v1/generations.jsonl","stratum":"512_token_cap","condition_id":"oq2e","condition_label":"oQ2e","model_id":"Hemmingway-1-oQ2e-mtp","prompt_category":"money_admin","max_tokens":512,"repetition":1,"status":"ok","finish_reason":"stop","sampler_json":"{\"min_p\":0,\"presence_penalty\":0,\"repetition_penalty\":1,\"seed\":42,\"temperature\":0,\"top_k\":0,\"top_p\":1}","chat_template_kwargs_json":"{\"enable_thinking\":false,\"preserve_thinking\":false}","metrics_json":"{\"completion_tokens\":134,\"decode_tokens_per_second\":44.73582639045431,\"latency_seconds\":3.443567999987863,\"prefill_tokens_per_second\":136.09817140393804,\"prompt_tokens\":61,\"total_tokens\":195,\"ttft_seconds\":0.4482058749999851}","response_sha256":"61492df9878bc70e4a4a1cc58c302ce1fb7898e4bc683b92ddbd2e1c97903410","response_character_count":576,"response_included":false,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":null,"hf_repo":null,"size_gib":null,"win_rate_percent":null,"loss_rate_percent":null,"tie_rate_percent":null,"interval_percent":null,"mean_clarity_and_control":null,"mean_instruction_adherence":null,"mean_task_fit":null,"mean_writing_judgment":null,"interpretation":null}
 {"record_type":"local_execution","record_id":"6298122081fa0fe947b970b6","prompt_id":"money_late_fee","category":null,"prompt":null,"performance_prompt":null,"prompt_sha256":null,"quality_max_tokens":null,"source_run":"quant_core_512","source_path":"results/hemmingway-quality-v1/generations.jsonl","stratum":"512_token_cap","condition_id":"oq2e","condition_label":"oQ2e","model_id":"Hemmingway-1-oQ2e-mtp","prompt_category":"money_admin","max_tokens":512,"repetition":1,"status":"ok","finish_reason":"stop","sampler_json":"{\"min_p\":0,\"presence_penalty\":0,\"repetition_penalty\":1,\"seed\":42,\"temperature\":0,\"top_k\":0,\"top_p\":1}","chat_template_kwargs_json":"{\"enable_thinking\":false,\"preserve_thinking\":false}","metrics_json":"{\"completion_tokens\":263,\"decode_tokens_per_second\":43.10746676029171,\"latency_seconds\":6.500218458008021,\"prefill_tokens_per_second\":152.8104917940138,\"prompt_tokens\":61,\"total_tokens\":324,\"ttft_seconds\":0.3991872500628233}","response_sha256":"b3e1920602b188649bf6e1c927c194898d8bdbfc95e644149bde474684b36ae0","response_character_count":1288,"response_included":false,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":null,"hf_repo":null,"size_gib":null,"win_rate_percent":null,"loss_rate_percent":null,"tie_rate_percent":null,"interval_percent":null,"mean_clarity_and_control":null,"mean_instruction_adherence":null,"mean_task_fit":null,"mean_writing_judgment":null,"interpretation":null}
@@ -175,10 +175,10 @@
 {"record_type":"hosted_quality_reference","record_id":null,"prompt_id":"expressive_mouse_hawk","category":null,"prompt":null,"performance_prompt":null,"prompt_sha256":null,"quality_max_tokens":null,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":"expressive_writing","max_tokens":null,"repetition":null,"status":null,"finish_reason":"stop","sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":"f5f545c6e6985ffa7b7f4aebb09f6a7b7031d7c1ec8b120d771e4663721507d7","response_character_count":446,"response_included":false,"quality_only":true,"response_model":"hemmingway-27b","http_status":200,"latency_seconds":1.4173248340375721,"usage_json":"{\"completion_tokens\":104,\"prompt_tokens\":142,\"total_tokens\":246}","condition":null,"hf_repo":null,"size_gib":null,"win_rate_percent":null,"loss_rate_percent":null,"tie_rate_percent":null,"interval_percent":null,"mean_clarity_and_control":null,"mean_instruction_adherence":null,"mean_task_fit":null,"mean_writing_judgment":null,"interpretation":null}
 {"record_type":"hosted_quality_reference","record_id":null,"prompt_id":"expressive_monologue","category":null,"prompt":null,"performance_prompt":null,"prompt_sha256":null,"quality_max_tokens":null,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":"expressive_writing","max_tokens":null,"repetition":null,"status":null,"finish_reason":"stop","sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":"8be587fced8cd336f9130db719f54e4f21fd6c86bf40535cc64edd7bf186303b","response_character_count":1614,"response_included":false,"quality_only":true,"response_model":"hemmingway-27b","http_status":200,"latency_seconds":5.1324289590120316,"usage_json":"{\"completion_tokens\":412,\"prompt_tokens\":160,\"total_tokens\":572}","condition":null,"hf_repo":null,"size_gib":null,"win_rate_percent":null,"loss_rate_percent":null,"tie_rate_percent":null,"interval_percent":null,"mean_clarity_and_control":null,"mean_instruction_adherence":null,"mean_task_fit":null,"mean_writing_judgment":null,"interpretation":null}
 {"record_type":"hosted_quality_reference","record_id":null,"prompt_id":"expressive_diary","category":null,"prompt":null,"performance_prompt":null,"prompt_sha256":null,"quality_max_tokens":null,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":"expressive_writing","max_tokens":null,"repetition":null,"status":null,"finish_reason":"stop","sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":"0c430e4a439aae81d0315234fb72d9e642e8f1474f926ce5a2ef42f944aa6e17","response_character_count":1664,"response_included":false,"quality_only":true,"response_model":"hemmingway-27b","http_status":200,"latency_seconds":4.881144625018351,"usage_json":"{\"completion_tokens\":409,\"prompt_tokens\":161,\"total_tokens\":570}","condition":null,"hf_repo":null,"size_gib":null,"win_rate_percent":null,"loss_rate_percent":null,"tie_rate_percent":null,"interval_percent":null,"mean_clarity_and_control":null,"mean_instruction_adherence":null,"mean_task_fit":null,"mean_writing_judgment":null,"interpretation":null}
-{"record_type":"aggregate_judge_result","record_id":"aggregate:hosted_reference_subset","prompt_id":null,"category":null,"prompt":null,"performance_prompt":null,"prompt_sha256":null,"quality_max_tokens":null,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":"hosted_reference_subset","hf_repo":null,"size_gib":null,"win_rate_percent":45.5,"loss_rate_percent":47.0,"tie_rate_percent":7.6,"interval_percent":12.0,"mean_clarity_and_control":null,"mean_instruction_adherence":null,"mean_task_fit":null,"mean_writing_judgment":null,"interpretation":"Cap-matched hosted reference subset; not a quantization condition."}
-{"record_type":"aggregate_judge_result","record_id":"aggregate:oQ2e","prompt_id":null,"category":null,"prompt":null,"performance_prompt":null,"prompt_sha256":null,"quality_max_tokens":null,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":"oQ2e","hf_repo":"sixstringzen/Hemmingway-1-oQ2e-mtp","size_gib":10.14,"win_rate_percent":47.8,"loss_rate_percent":51.1,"tie_rate_percent":1.1,"interval_percent":10.3,"mean_clarity_and_control":-0.18,"mean_instruction_adherence":0.2,"mean_task_fit":0.12,"mean_writing_judgment":0.03,"interpretation":"Compact build; more divergent from BF16 in this suite."}
-{"record_type":"aggregate_judge_result","record_id":"aggregate:oQ3.5e","prompt_id":null,"category":null,"prompt":null,"performance_prompt":null,"prompt_sha256":null,"quality_max_tokens":null,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":"oQ3.5e","hf_repo":"sixstringzen/Hemmingway-1-oQ3.5e-mtp","size_gib":13.19,"win_rate_percent":36.7,"loss_rate_percent":58.9,"tie_rate_percent":4.4,"interval_percent":10.0,"mean_clarity_and_control":-0.53,"mean_instruction_adherence":-0.71,"mean_task_fit":-0.63,"mean_writing_judgment":-0.6,"interpretation":"Weakest aggregate outcome in this pass."}
-{"record_type":"aggregate_judge_result","record_id":"aggregate:oQ3e","prompt_id":null,"category":null,"prompt":null,"performance_prompt":null,"prompt_sha256":null,"quality_max_tokens":null,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":"oQ3e","hf_repo":"sixstringzen/Hemmingway-1-oQ3e-mtp","size_gib":12.22,"win_rate_percent":55.6,"loss_rate_percent":43.3,"tie_rate_percent":1.1,"interval_percent":10.3,"mean_clarity_and_control":0.22,"mean_instruction_adherence":0.33,"mean_task_fit":0.38,"mean_writing_judgment":0.26,"interpretation":"Highest point estimate, but inconclusive because of order sensitivity and uncertainty."}
-{"record_type":"aggregate_judge_result","record_id":"aggregate:oQ4e","prompt_id":null,"category":null,"prompt":null,"performance_prompt":null,"prompt_sha256":null,"quality_max_tokens":null,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":"oQ4e","hf_repo":"sixstringzen/Hemmingway-1-oQ4e-mtp","size_gib":15.21,"win_rate_percent":32.2,"loss_rate_percent":48.9,"tie_rate_percent":18.9,"interval_percent":9.7,"mean_clarity_and_control":-0.37,"mean_instruction_adherence":-0.44,"mean_task_fit":-0.47,"mean_writing_judgment":-0.43,"interpretation":"Lower aggregate quality than oQ6e and oQ8e in this pass."}
-{"record_type":"aggregate_judge_result","record_id":"aggregate:oQ6e","prompt_id":null,"category":null,"prompt":null,"performance_prompt":null,"prompt_sha256":null,"quality_max_tokens":null,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":"oQ6e","hf_repo":"sixstringzen/Hemmingway-1-oQ6e-mtp","size_gib":21.39,"win_rate_percent":27.8,"loss_rate_percent":31.1,"tie_rate_percent":41.1,"interval_percent":9.3,"mean_clarity_and_control":-0.02,"mean_instruction_adherence":-0.12,"mean_task_fit":-0.03,"mean_writing_judgment":-0.02,"interpretation":"High-bit perceptual-reference group for this suite; not mathematically lossless."}
-{"record_type":"aggregate_judge_result","record_id":"aggregate:oQ8e","prompt_id":null,"category":null,"prompt":null,"performance_prompt":null,"prompt_sha256":null,"quality_max_tokens":null,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":"oQ8e","hf_repo":"sixstringzen/Hemmingway-1-oQ8e-mtp","size_gib":27.1,"win_rate_percent":31.1,"loss_rate_percent":23.3,"tie_rate_percent":45.6,"interval_percent":9.6,"mean_clarity_and_control":-0.06,"mean_instruction_adherence":0.02,"mean_task_fit":0.08,"mean_writing_judgment":-0.04,"interpretation":"High-bit perceptual-reference group for this suite; not mathematically lossless."}
+{"record_type":"aggregate_judge_result","record_id":"aggregate:hosted_reference_subset","prompt_id":null,"category":null,"prompt":null,"performance_prompt":null,"prompt_sha256":null,"quality_max_tokens":null,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":"hosted_reference_subset","hf_repo":null,"size_gib":null,"win_rate_percent":77.3,"loss_rate_percent":15.2,"tie_rate_percent":7.6,"interval_percent":null,"mean_clarity_and_control":0.7273,"mean_instruction_adherence":0.8788,"mean_task_fit":1.0758,"mean_writing_judgment":1.0758,"interpretation":"Hosted was preferred on the 11-prompt subset. Quants were not compared directly with hosted."}
+{"record_type":"aggregate_judge_result","record_id":"aggregate:oQ2e","prompt_id":null,"category":null,"prompt":null,"performance_prompt":null,"prompt_sha256":null,"quality_max_tokens":null,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":"oQ2e","hf_repo":"sixstringzen/Hemmingway-1-oQ2e-mtp","size_gib":10.14,"win_rate_percent":34.4,"loss_rate_percent":64.4,"tie_rate_percent":1.1,"interval_percent":null,"mean_clarity_and_control":-0.6,"mean_instruction_adherence":-0.4667,"mean_task_fit":-0.4111,"mean_writing_judgment":-0.5667,"interpretation":"BF16 was preferred more often: 58 losses, 31 wins, and one tie."}
+{"record_type":"aggregate_judge_result","record_id":"aggregate:oQ3.5e","prompt_id":null,"category":null,"prompt":null,"performance_prompt":null,"prompt_sha256":null,"quality_max_tokens":null,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":"oQ3.5e","hf_repo":"sixstringzen/Hemmingway-1-oQ3.5e-mtp","size_gib":13.19,"win_rate_percent":48.9,"loss_rate_percent":46.7,"tie_rate_percent":4.4,"interval_percent":null,"mean_clarity_and_control":0.2889,"mean_instruction_adherence":0.3111,"mean_task_fit":0.6111,"mean_writing_judgment":0.5111,"interpretation":"Near-even preference counts: 44 wins, 42 losses, and four ties."}
+{"record_type":"aggregate_judge_result","record_id":"aggregate:oQ3e","prompt_id":null,"category":null,"prompt":null,"performance_prompt":null,"prompt_sha256":null,"quality_max_tokens":null,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":"oQ3e","hf_repo":"sixstringzen/Hemmingway-1-oQ3e-mtp","size_gib":12.22,"win_rate_percent":60.0,"loss_rate_percent":38.9,"tie_rate_percent":1.1,"interval_percent":null,"mean_clarity_and_control":0.4889,"mean_instruction_adherence":0.5333,"mean_task_fit":0.7556,"mean_writing_judgment":0.7667,"interpretation":"Highest observed win rate in this suite. This does not establish a general advantage over BF16."}
+{"record_type":"aggregate_judge_result","record_id":"aggregate:oQ4e","prompt_id":null,"category":null,"prompt":null,"performance_prompt":null,"prompt_sha256":null,"quality_max_tokens":null,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":"oQ4e","hf_repo":"sixstringzen/Hemmingway-1-oQ4e-mtp","size_gib":15.21,"win_rate_percent":33.3,"loss_rate_percent":47.8,"tie_rate_percent":18.9,"interval_percent":null,"mean_clarity_and_control":-0.2778,"mean_instruction_adherence":-0.1778,"mean_task_fit":-0.2,"mean_writing_judgment":-0.3,"interpretation":"BF16 was preferred more often: 43 losses, 30 wins, and 17 ties."}
+{"record_type":"aggregate_judge_result","record_id":"aggregate:oQ6e","prompt_id":null,"category":null,"prompt":null,"performance_prompt":null,"prompt_sha256":null,"quality_max_tokens":null,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":"oQ6e","hf_repo":"sixstringzen/Hemmingway-1-oQ6e-mtp","size_gib":21.39,"win_rate_percent":37.8,"loss_rate_percent":21.1,"tie_rate_percent":41.1,"interval_percent":null,"mean_clarity_and_control":0.1556,"mean_instruction_adherence":0.2333,"mean_task_fit":0.3,"mean_writing_judgment":0.3333,"interpretation":"Frequent ties with BF16 and more wins than losses in this suite."}
+{"record_type":"aggregate_judge_result","record_id":"aggregate:oQ8e","prompt_id":null,"category":null,"prompt":null,"performance_prompt":null,"prompt_sha256":null,"quality_max_tokens":null,"source_run":null,"source_path":null,"stratum":null,"condition_id":null,"condition_label":null,"model_id":null,"prompt_category":null,"max_tokens":null,"repetition":null,"status":null,"finish_reason":null,"sampler_json":null,"chat_template_kwargs_json":null,"metrics_json":null,"response_sha256":null,"response_character_count":null,"response_included":null,"quality_only":null,"response_model":null,"http_status":null,"latency_seconds":null,"usage_json":null,"condition":"oQ8e","hf_repo":"sixstringzen/Hemmingway-1-oQ8e-mtp","size_gib":27.1,"win_rate_percent":30.0,"loss_rate_percent":24.4,"tie_rate_percent":45.6,"interval_percent":null,"mean_clarity_and_control":0.3,"mean_instruction_adherence":0.2,"mean_task_fit":0.1222,"mean_writing_judgment":0.2222,"interpretation":"Highest tie rate in this suite, with 27 wins and 22 losses."}

metadata/benchmark_manifest.yaml

@@ -1,4 +1,4 @@
-schema_version: 1
+schema_version: 2
 dataset_id: hemmingway-1-omlx-quantization-benchmark-v1
 title: Hemmingway-1 oMLX Quantization Benchmark
 source_model:
@@ -7,6 +7,12 @@ source_model:
   publisher: Altworld
 study_id: hemmingway-1-omlx-oqe-v1
 task_count: 15
+analysis_revision: 2
+task_cap_semantics: quality_max_tokens is the final cap selected for judging; local_execution.max_tokens
+  records each attempt's actual cap.
+deprecated_fields:
+  interval_percent: Always null in corrected revision 2. Independent-rating confidence
+    intervals were withdrawn.
 selected_local_execution_record_count: 147
 hosted_quality_only_record_count: 15
 aggregate_judge_condition_count: 7

metadata/gemini-3.8-flash-veracity-review.yaml

@@ -1,4 +1,6 @@
 model: google/gemini-3.8-flash
+publication_status: historical_review_superseded_for_aggregate_accuracy
+applicability: This review predates the 2026-09-22 analysis correction. It did not detect A/B decoding or reversed-prompt matching errors and does not validate revision 2.
 review_id: gen-1790055089-8hA2nqTqg7N2URxqeX6Q
 verdict: PASS
 confidence: 1.0

metadata/public_results.yaml

@@ -4,14 +4,32 @@ study:
   title: Hemmingway-1 oMLX quantization blind-judge analysis v1
   source_model: Altworld/Hemmingway-1
   collection: sixstringzen/hemmingway-1-omlx-oqe-quantizations-apple-silicon-evidence-6ab06103552bc3c7e9ad8afa
-  prompt_count: 14
+  prompt_count: 15
   candidate_count: 6
   judge_count: 3
   paired_orders:
   - normal
   - swapped
-  aggregation: Pooled across 14 cases and three judge lanes. The hosted reference
-    row is a cap-matched subset of 11 prompts and is not a quantization condition.
+  aggregation: Each quant has 15 prompts, two orders, and three judges (90 ratings).
+    Hosted has 11 cap-matched prompts (66 ratings). Ratings of the same outputs are
+    dependent.
+  analysis_revision: 2
+  prompt_caps:
+    money_landlord_boiler: 512
+    money_late_fee: 512
+    money_reschedule_appointment: 512
+    work_estimate_delay: 512
+    work_deadline_pushback: 512
+    work_review_bypass: 2048
+    hard_ask_repayment: 512
+    hard_ask_no_hosting: 512
+    hard_ask_chores: 512
+    eq_neighbor_noise: 512
+    eq_documentation: 2048
+    eq_friend_disagreement: 2048
+    expressive_mouse_hawk: 512
+    expressive_monologue: 1024
+    expressive_diary: 512
 execution_provenance:
   measurement_source: Real MLX/oMLX software running locally on the Apple Silicon
     machine, with captured generation records and hardware telemetry.
@@ -51,99 +69,232 @@ judge_lanes:
 - Grok 4.7
 results:
 - condition: hosted_reference_subset
-  win_rate_percent: 45.5
-  loss_rate_percent: 47.0
+  wins: 51
+  losses: 10
+  ties: 5
+  judgment_count: 66
+  prompt_count: 11
+  win_rate_percent: 77.3
+  loss_rate_percent: 15.2
   tie_rate_percent: 7.6
-  interval_percent: 12.0
-  note: Cap-matched hosted reference subset; not a quantization condition.
+  mean_score_difference:
+    clarity_and_control: 0.7273
+    instruction_adherence: 0.8788
+    task_fit: 1.0758
+    writing_judgment: 1.0758
+  by_judge:
+    Claude:
+      wins: 18
+      losses: 2
+      ties: 2
+    Gemini:
+      wins: 17
+      losses: 5
+      ties: 0
+    Grok:
+      wins: 16
+      losses: 3
+      ties: 3
+  interpretation: Hosted was preferred on the 11-prompt subset. Quants were not compared
+    directly with hosted.
 - condition: oQ2e
   hf_repo: sixstringzen/Hemmingway-1-oQ2e-mtp
   size_gib: 10.14
-  win_rate_percent: 47.8
-  loss_rate_percent: 51.1
+  wins: 31
+  losses: 58
+  ties: 1
+  judgment_count: 90
+  prompt_count: 15
+  win_rate_percent: 34.4
+  loss_rate_percent: 64.4
   tie_rate_percent: 1.1
-  interval_percent: 10.3
   mean_score_difference:
-    clarity_and_control: -0.18
-    instruction_adherence: 0.2
-    task_fit: 0.12
-    writing_judgment: 0.03
-  interpretation: Compact build; more divergent from BF16 in this suite.
+    clarity_and_control: -0.6
+    instruction_adherence: -0.4667
+    task_fit: -0.4111
+    writing_judgment: -0.5667
+  by_judge:
+    Claude:
+      wins: 10
+      losses: 20
+      ties: 0
+    Gemini:
+      wins: 11
+      losses: 19
+      ties: 0
+    Grok:
+      wins: 10
+      losses: 19
+      ties: 1
+  interpretation: 'BF16 was preferred more often: 58 losses, 31 wins, and one tie.'
 - condition: oQ3.5e
   hf_repo: sixstringzen/Hemmingway-1-oQ3.5e-mtp
   size_gib: 13.19
-  win_rate_percent: 36.7
-  loss_rate_percent: 58.9
+  wins: 44
+  losses: 42
+  ties: 4
+  judgment_count: 90
+  prompt_count: 15
+  win_rate_percent: 48.9
+  loss_rate_percent: 46.7
   tie_rate_percent: 4.4
-  interval_percent: 10.0
   mean_score_difference:
-    clarity_and_control: -0.53
-    instruction_adherence: -0.71
-    task_fit: -0.63
-    writing_judgment: -0.6
-  interpretation: Weakest aggregate outcome in this pass.
+    clarity_and_control: 0.2889
+    instruction_adherence: 0.3111
+    task_fit: 0.6111
+    writing_judgment: 0.5111
+  by_judge:
+    Claude:
+      wins: 16
+      losses: 13
+      ties: 1
+    Gemini:
+      wins: 14
+      losses: 16
+      ties: 0
+    Grok:
+      wins: 14
+      losses: 13
+      ties: 3
+  interpretation: 'Near-even preference counts: 44 wins, 42 losses, and four ties.'
 - condition: oQ3e
   hf_repo: sixstringzen/Hemmingway-1-oQ3e-mtp
   size_gib: 12.22
-  win_rate_percent: 55.6
-  loss_rate_percent: 43.3
+  wins: 54
+  losses: 35
+  ties: 1
+  judgment_count: 90
+  prompt_count: 15
+  win_rate_percent: 60.0
+  loss_rate_percent: 38.9
   tie_rate_percent: 1.1
-  interval_percent: 10.3
   mean_score_difference:
-    clarity_and_control: 0.22
-    instruction_adherence: 0.33
-    task_fit: 0.38
-    writing_judgment: 0.26
-  interpretation: Highest point estimate, but inconclusive because of order sensitivity
-    and uncertainty.
+    clarity_and_control: 0.4889
+    instruction_adherence: 0.5333
+    task_fit: 0.7556
+    writing_judgment: 0.7667
+  by_judge:
+    Claude:
+      wins: 20
+      losses: 10
+      ties: 0
+    Gemini:
+      wins: 18
+      losses: 12
+      ties: 0
+    Grok:
+      wins: 16
+      losses: 13
+      ties: 1
+  interpretation: Highest observed win rate in this suite. This does not establish
+    a general advantage over BF16.
 - condition: oQ4e
   hf_repo: sixstringzen/Hemmingway-1-oQ4e-mtp
   size_gib: 15.21
-  win_rate_percent: 32.2
-  loss_rate_percent: 48.9
+  wins: 30
+  losses: 43
+  ties: 17
+  judgment_count: 90
+  prompt_count: 15
+  win_rate_percent: 33.3
+  loss_rate_percent: 47.8
   tie_rate_percent: 18.9
-  interval_percent: 9.7
   mean_score_difference:
-    clarity_and_control: -0.37
-    instruction_adherence: -0.44
-    task_fit: -0.47
-    writing_judgment: -0.43
-  interpretation: Lower aggregate quality than oQ6e and oQ8e in this pass.
+    clarity_and_control: -0.2778
+    instruction_adherence: -0.1778
+    task_fit: -0.2
+    writing_judgment: -0.3
+  by_judge:
+    Claude:
+      wins: 8
+      losses: 15
+      ties: 7
+    Gemini:
+      wins: 13
+      losses: 13
+      ties: 4
+    Grok:
+      wins: 9
+      losses: 15
+      ties: 6
+  interpretation: 'BF16 was preferred more often: 43 losses, 30 wins, and 17 ties.'
 - condition: oQ6e
   hf_repo: sixstringzen/Hemmingway-1-oQ6e-mtp
   size_gib: 21.39
-  win_rate_percent: 27.8
-  loss_rate_percent: 31.1
+  wins: 34
+  losses: 19
+  ties: 37
+  judgment_count: 90
+  prompt_count: 15
+  win_rate_percent: 37.8
+  loss_rate_percent: 21.1
   tie_rate_percent: 41.1
-  interval_percent: 9.3
   mean_score_difference:
-    clarity_and_control: -0.02
-    instruction_adherence: -0.12
-    task_fit: -0.03
-    writing_judgment: -0.02
-  interpretation: High-bit perceptual-reference group for this suite; not mathematically
-    lossless.
+    clarity_and_control: 0.1556
+    instruction_adherence: 0.2333
+    task_fit: 0.3
+    writing_judgment: 0.3333
+  by_judge:
+    Claude:
+      wins: 11
+      losses: 6
+      ties: 13
+    Gemini:
+      wins: 13
+      losses: 6
+      ties: 11
+    Grok:
+      wins: 10
+      losses: 7
+      ties: 13
+  interpretation: Frequent ties with BF16 and more wins than losses in this suite.
 - condition: oQ8e
   hf_repo: sixstringzen/Hemmingway-1-oQ8e-mtp
   size_gib: 27.1
-  win_rate_percent: 31.1
-  loss_rate_percent: 23.3
+  wins: 27
+  losses: 22
+  ties: 41
+  judgment_count: 90
+  prompt_count: 15
+  win_rate_percent: 30.0
+  loss_rate_percent: 24.4
   tie_rate_percent: 45.6
-  interval_percent: 9.6
   mean_score_difference:
-    clarity_and_control: -0.06
-    instruction_adherence: 0.02
-    task_fit: 0.08
-    writing_judgment: -0.04
-  interpretation: High-bit perceptual-reference group for this suite; not mathematically
-    lossless.
+    clarity_and_control: 0.3
+    instruction_adherence: 0.2
+    task_fit: 0.1222
+    writing_judgment: 0.2222
+  by_judge:
+    Claude:
+      wins: 8
+      losses: 9
+      ties: 13
+    Gemini:
+      wins: 10
+      losses: 7
+      ties: 13
+    Grok:
+      wins: 9
+      losses: 6
+      ties: 15
+  interpretation: Highest tie rate in this suite, with 27 wins and 22 losses.
+pending_metrics:
+- teacher_forced_kld
+- top_1_and_top_k_agreement
+- kv_cache_divergence
+- serious_benchmark_dataset_task_scores
+- controlled_runtime_and_memory_comparison
+publication_boundary:
+  included: Aggregate findings, method notes, sanitized metadata, and release links.
+  excluded: Private answer key, raw provider responses, blind packets, and sensitive
+    or unredacted benchmark generations.
 reliability:
   raw_inter_rater_agreement_percent: 83.7
   raw_inter_rater_agreement_n: 606
   normal_vs_swapped_agreement_percent:
-    Claude Opus 5: 39.6
-    Gemini 3.8 Flash: 48.5
-    Grok 4.7: 42.6
+    Claude Opus 5: 89.1
+    Gemini 3.8 Flash: 85.1
+    Grok 4.7: 83.2
   normal_vs_swapped_n_per_judge: 101
   position_bias:
     Claude Opus 5:
@@ -155,13 +306,14 @@ reliability:
     Grok 4.7:
       response_a_percent: 43.1
       response_b_percent: 36.1
-pending_metrics:
-- teacher_forced_kld
-- top_1_and_top_k_agreement
-- kv_cache_divergence
-- serious_benchmark_dataset_task_scores
-- controlled_runtime_and_memory_comparison
-publication_boundary:
-  included: Aggregate findings, method notes, sanitized metadata, and release links.
-  excluded: Private answer key, raw provider responses, blind packets, and sensitive
-    or unredacted benchmark generations.
+correction:
+  date: '2026-09-22'
+  analysis_revision: 2
+  reason: Resolve A/B assignments with prompt IDs; match reversed packets by prompt;
+    correct task counts; remove independent-rating intervals.
+  generations_changed: false
+  judge_results_changed: false
+  earlier_package_reviews: Superseded for aggregate accuracy; no new model-judge review
+    was run.
+uncertainty_method: Descriptive only. Dependent ratings; no confidence intervals or
+  significance claims.

metadata/reliability.yaml

@@ -1,9 +1,9 @@
 raw_inter_rater_agreement_percent: 83.7
 raw_inter_rater_agreement_n: 606
 normal_vs_swapped_agreement_percent:
-  Claude Opus 5: 39.6
-  Gemini 3.8 Flash: 48.5
-  Grok 4.7: 42.6
+  Claude Opus 5: 89.1
+  Gemini 3.8 Flash: 85.1
+  Grok 4.7: 83.2
 normal_vs_swapped_n_per_judge: 101
 position_bias:
   Claude Opus 5: