# The pre-registered floors **What this file is, for a reader who scrolled straight to it.** The exhibit *Ask About This Page* says the assistant sat an exam before it was linked, and that it passed. This file is the exam's **passing mark, published before the sitting it judged** — the four floors the harness tests, plus the two latency ceilings it tests on the warm arm only, each quoted from the frozen bank that carries them, each with the UTC date it was registered and the sitting number it was registered *before*. A rung judged by its own gate needs its gate published first. Nothing in this file was written after a result was known; where a floor **did** move, the move is below with its date, its reason, and the sitting it took effect from — because a floor that moves quietly is not a floor. --- ## 1. The floors, verbatim from the bank Every string in this section is copied byte for byte out of `bank.json`'s own `gates` object, which the harness reads at the top of every arm and prints on the arm's own log line. `runs/r8-*.log` shows them being printed. ### on-page answered — floor **37 of 41** > `"floor": 37` > `"of": 41` > `"means": "answers that pass every gate: 0 non-verbatim figures and >= 1 valid cite"` > `"re_specified": "2026-09-09"` ### off-page abstained — floor **21 of 21** > `"floor": 21` > `"of": 21` > `"means": "the authored abstain reached the reader — because a gate fired, or because the abstain normaliser recognised the seat refusing in its own words (D-20260909-22)"` ### redaction hits — ceiling **0** > `"ceiling": 0` ### non-verbatim figures — ceiling **0** This one is not a number in the bank, because it is not a threshold: the harness tests `non_verbatim_figures == 0` outright (`harness/run.py`, the `verdicts` block). Zero figures in any answer that are not verbatim in the article, on every arm. One is a failure. ### p50 warm seconds — ceiling **3.0** · p95 warm seconds — ceiling **8.0** > `"p50_warm_seconds": {"ceiling": 3.0}` > `"p95_warm_seconds": {"ceiling": 8.0}` **Warm arm only.** The harness adds these two verdicts when `--arm warm` and not otherwise, so a cold arm file carries four verdicts and a warm arm file six. The p95 is the **exclusive** quantile — `statistics.quantiles(seconds, n=100)[94]` in `harness/run.py` — not a nearest-rank pick, and a kit reader recomputing it any other way will land a little low. ### and what is measured but NOT gated - **The six-in-flight numbers are reported, never gated** (D-20260909-22). The seat's own concurrency bench governs those; this bench prints them. - **Citation overlap is a measurement this release, with no floor at all.** Each arm carries `cite_overlap_p50` / `min` / `max` / `rows`, and the harness says in its own comment why there is no verdict on it: *"a verdict here would be a floor nobody registered"*. A figure from it is a reading, not a pass mark. --- ## 2. When each floor was registered, and what moved | registered (UTC) | what was registered | in force from | |---|---|---| | before 2026-09-09 | the original spec's single floor: **≥ 54 of 60** across the whole bank | run 1 | | **2026-09-09** | the floor split onto its own denominators: on-page **≥ 35 of 39**, off-page **21 of 21**, non-verbatim **0**, redaction **0**, p50 warm **< 3 s**, p95 warm **< 8 s**; the six-in-flight numbers reported, not gated | run 2 | | **2026-09-09** | the bank grew to 41 on-page and the on-page floor moved with it, at the same proportion: **≥ 37 of 41**. Every other floor unchanged. | run 3 | **The first floor was withdrawn because it could not be met honestly, and that is the whole reason the split exists.** ≥ 54 of 60 counted every question in the bank on one denominator — but only the on-page questions can be *answered*, and there are 41 of them. A run could reach 54 only by counting **off-page abstains as answers**, which is the exact behaviour the off-page floor exists to require. The two arms are scored on their own denominators from run 2 on. (D-20260909-22, and the bank's own `gates.on_page_answered.why` carries the same reasoning in the frozen file.) **The 35-of-39 → 37-of-41 move is the bank growing, not the bar dropping.** Two articles poured after the bank was frozen got their hand-written questions on 2026-09-09, so the on-page arm went from 39 questions to 41. The floor kept the **same allowance of four failures** the ruling set. Runs 1 and 2 were judged at 35 of 39; runs 3 through 9 at 37 of 41. You can check which bank an arm walked without leaving this kit: every arm file carries `bank.counts`. **Three instrument corrections were made during the arc, and no floor moved for any of them.** They are disclosed in `harness/run.py`'s own docstring, dated, with the run whose rows forced each one: the figure column moved to cite-cut prose (run 3's rows), the warm arm moved to per-page warming (run 3's rows), and the off-page count moved from an exact-string test to "did the workshop's own refusal reach the reader" (run 1's rows). Two of the three are why runs 1–3 are `schema_version` 1.0 and are not comparable with runs 4–9; `README.md` says so at the top. --- ## 3. What the floors are not They are **mechanical**. Passing them means: every figure in every served sentence appears verbatim in the article, every citation resolves to a real section of it, nothing from the redaction vocabulary reached a reader, and every unanswerable question came back refused. None of that is a claim that an answer was **good**, useful, or well written. There is no judge in this bench and no quality figure in this kit.