# AMENDMENT-1 · PREREG-QUANT · a second board on the bus, G-9 when the anchor cannot be judged, one path *What this file is, for a reader who scrolled straight here.* An amendment to `PREREG-QUANT.md`, the pre-registration of the quantization-ladder bench (Mistral NeMo 12B Instruct at five quantization levels on benchbox's GeForce RTX 3090 24 GB), frozen 2026-10-14T20:52:44Z with freeze line `withheld-2026-09-29`. Written by lane `quant-ladder-fold` (Opus 5.5) at 2026-10-14T22:15:53Z (UTC, `date -u`) from the kit audit (1 BLOCKING, 4 IMPORTANT, 5 NICE), BEFORE any scored call: benchbox did not answer ssh at 2026-10-14T21:59:28Z and no smoke receipt exists. PREREG section 9 asks for an amendment below the FROZEN line; the fold's brief rules that the frozen file is never edited and that an amendment is its own file the kit pins. So this file sits beside the PREREG and its sha256 is a line of `kit/SHA256SUMS`: the KIT paste (`kit/push.sh`) refuses to copy anything to the box unless this file and the PREREG verify. INTERNAL, not public-bound. ## A1 · G-1 for the tasks step on a box with two boards (the audit's BLOCKING B1) A 2026-10-09 recon reads two GeForce RTX 3090s in benchbox. The stage's card rule (PREREG section 4) already picks the x16 seat when two qualify, hides the other from the bench's Ollama, and the speed step runs under the one-model posture and the watcher. But the cost-of-intelligence instrument's posture gate refused unless exactly ONE board was on the bus, so on a two-board benchbox every level's tasks step would refuse (rc 4) and the ladder would halt at Q4_K_M with no six-way or doorman row read. **Amended rule, for the tasks step:** the named card (the UUID the stage picked) is listed once on the bus; no compute process runs on any other board at each posture read; other boards stay idle, watched by `qwatch` (G-2: no other board grows by 200 MiB or more). 300.00 W and persistence Enabled on the named card are unchanged. This is the fourth documented edit to `coi_quant.py` (`kit/INSTRUMENT-DELTA.md`); no measuring line moves. **One consequence, stated so no page claims otherwise:** a WALL figure (the UPS's whole-box watts) includes any idle second board, so WALL figures are not compared with the 2026-09-23 anchor's (read on a one-board box). The board figures, the anchor check (P4) and every prediction read the named card only; the reading takes each level's growth from the fit rule's named-card figure, never from a sum over boards. ## A2 · G-9 when the anchor cannot be judged G-9 reads "P4 fails". An UNDECIDABLE P4 (the Q4_K_M tasks were not tested, an item count is not the frozen 108 or 48, or the per-item join with the anchor does not hold) has not shown the replication either. So G-9 also fires then, labelled "replication not shown"; a REFUTED P4 keeps the frozen label, "instrument drift". ## A3 · a path Section 2.1 cites `tasks/INSTRUMENT-DELTA.md`; the file is `kit/INSTRUMENT-DELTA.md`. ## Not amendments: where the reading now follows the frozen text (listed so a reader can check them) - P1: each adjacent pair's gap in decode tok/s at 4,096 must exceed the larger of the two rungs' run spreads, read as the max − min of each rung's three scored runs in tok/s (the first reading dropped this clause). - P2: the ±15 % band is around the five-level median of decode tok/s × store bytes (the first reading used four). - P4 needs n = 108 and 48 in the anchor join; P5 and P6 need n = 108 and 48 at every level; "the lowest of the five" is strict, so a tie is not the lowest. --- AMENDED 2026-10-14T22:15:53Z · sha256 of PREREG-QUANT.md's bytes followed by every byte above this line: withheld-2026-10-16 ## A4 · the card rule reads the link each board trained at *What this entry is, for a reader who scrolled straight here.* The fourth entry of this amendment to `PREREG-QUANT.md`, the pre-registration of the quantization-ladder bench (frozen 2026-10-14T20:52:44Z), written by lane `ql-leadfix` (Opus 5.5) at 2026-10-15T02:18:44Z (UTC, `date -u`), BEFORE any scored call: nothing has run on benchbox (no stage receipt, no smoke receipt). It changes one clause: the card rule of PREREG section 4, "the x16 seat when two qualify" (A1 above cites the same words). **Why.** The lead read benchbox at about 2026-10-15T02:2xZ (read-only): two GeForce RTX 3090 24 GB boards, both reporting `pcie.link.width.max` 16. The board on bus withheld-2026-10-16 reads `pcie.link.width.current` 4 (250 W, persistence off); the board on bus withheld-2026-10-16 reads current 16 (300 W, persistence off). The kit's rule looked for "the x16 seat" by the max width, found two and refused, so the stage would have refused on the box as it stands. **Amended rule.** Exactly the name `NVIDIA GeForce RTX 3090` with at least 24,000 MiB, as before. When more than one board qualifies, the board with the larger `pcie.link.width.current` is the card; a tie at the top refuses (rc 4) and the refusal names every board's current and max width. The stage records the card's current and max width in the box-local `state/card.env` and in its receipt (`link=x/x` on the STAGE OK line; every board's pair in the receipt file). On the read above, the rule picks the board on bus withheld-2026-10-16 (x16 now, x16 max). G-1 is unchanged: every later stage re-reads the card by UUID, name and bus id. **What it does not change.** No measuring line, prediction, gate or posture moves. 300.00 W and persistence Enabled are still read on the picked card at every stage; on the read above the picked card has persistence off, so the stage will report `posture=NOT-READY` and an operator's posture paste (SMOKE-READY.md step 2b) comes before the smoke. *Beside this entry, not an amendment:* every `ql-*` unit's output (the smoke, speed and tasks units, the ladder unit and the bench's own Ollama) now passes an in-unit scrub into the kit's own logs, never benchbox's user journal (systemd 259 sends a unit's output there unless `--pipe` is given). No PREREG clause says where logs go. --- AMENDED 2026-10-15T02:18:44Z · sha256 of PREREG-QUANT.md's bytes followed by every byte above this line: withheld-2026-10-16