Reports

• 2026-09-15

  b5_run_compare_v1_8_2026-09-15.json

  The live detector and the commercial detector we compared it with, on 51 heavily rewritten documents.

• 2026-09-15

  unseen_human_1200_v1_8_2026-09-15.md

  Twelve hundred human-written documents nobody had scored before, read by the live detector.

• 2026-09-14

  sensitivity_deploy_2026-09-14.md

  Shipping the three settings, and checking the default did not move in the deploy.

• 2026-09-14

  sensitivity_points_2026-09-14.md

  Where the three settings sit, and what each one costs, measured on the live detector.

• 2026-09-13

  v1_8_ship_decision_2026-09-13.md

  Which of the candidate runs went live, and what the choice cost.

• 2026-09-13

  v1_8_seed2_score_2026-09-13.md

  Test scores for the candidate runs, side by side.

• 2026-09-13

  seed_interval_2026-09-13.md

  How much of a one-document difference comes from the training run alone.

• 2026-09-12

  v1_8_train_2026-09-12.md

  A training pass for the candidate runs, and the margin it left.

• 2026-09-12

  v1_8_build_2026-09-12.md

  How the training set was put together, and which documents were kept out because they are test evidence.

• 2026-09-11

  beating_the_humanizer_2026-09-11.md

  What rewriting tools do to a reading, on an earlier run.

• 2026-09-10

  decoder_vs_data_2026-09-10.md

  Whether a gain on an earlier run came from the scoring arithmetic or from the training documents.

• 2026-09-10

  whitepaper_evidence_2026-09-10.md

  Claims traced back to per-document rows, for the two runs before the live one.

• 2026-09-09

  compute_rates_2026-09-09.md

  What a training pass and a scoring pass cost to run.

• 2026-09-09

  container_ceiling_2026-09-09.md

  A ceiling on the serving setup in its busiest hour, written down and tested.

• 2026-09-09

  light_edit_signal_2026-09-09.md

  Whether an earlier run could place a light polish edit inside a paper.

• 2026-09-09

  placement_whole_papers_2026-09-09.md

  Where the marked passages fall, checked on whole papers, for an earlier run.

• 2026-09-09

  serve_latency_2026-09-09.md

  Response times on one document, from calls that had already happened. No load test.

• 2026-09-09

  threshold_sweep_oos_2026-09-09.md

  False flags at many lines, repeated on human documents an earlier run had never scored.

• 2026-09-09

  truncation_artefact_2026-09-09.md

  A measurement error that raised scores on human writing, found and corrected.

• 2026-09-09

  trust_set_scores_2026-09-09.md

  Scores on documents kept out of training on purpose, for an earlier run.

• 2026-09-09

  unseen_human_600_2026-09-09.md

  Six hundred human documents nobody had scored before, read by an earlier run.

Paths are relative to the detector project’s root. The live detector is run v1_8_s2_seed303. Reports dated before 2026-09-12 describe earlier runs. Only a report that names the live run carries its figures.