The writing loop ledger ledger · 2026-09-13 22:39 · 20 loops

One row a loop, generated by python exam/ink_loop.py ledger from <ink_dir>/loop/ledger.json. The plan and the reading of these numbers: the writing loop. Win = the share of blind pairs (both orders) the critic preferred the Ink piece over a length-and-class-matched major, with its 95% range in brackets — 24 verdicts a loop is a small sample, so two loops whose ranges overlap cannot honestly be told apart (added 09-11, the outside research §7 item 1); score = the critic's raw 0–10 for the Ink pieces and for the majors in the same sitting (the gap is the number; the majors are not all 9s — the bank's plain band carries academic Conversation pieces and a joke DMCA notice); spotted = how often the critic guessed which piece was the machine (50% would mean it cannot tell); agree = how often the two orders of the same pair agreed; over-used = words · sentence shapes still over-used against the bank at z ≥ 3 (the target is 0 · 0); tells = negation titles % · short flat「X is Y」sentences % · contractions /100 · words a paragraph (the majors: 5–10% · 3–3.5% · 1.7–2.5 · 50–69). The models line under each recipe (added 09-13) = the writer model id the loop ran on and the critic's own model id as written into its verdicts — a win rate is read against a writer version and a critic version, never across one; two loops on different models are not a before-and-after (the industry scan §4: a rented model's habits move with each release). Loops before 09-13 show — for both.