# The CREATE's PROPORTIONALITY BAND at 8-row resolution, and whether an INTERLEAVED ladder moves the crossover Lane `parfast-create-band-and-interleave-18sep`, items **1 and 2** of an internal note. **Status: RAN AND COMPLETE.** One sitting on intel-core-ultra-9-386h, 01:44:51Z-03:04:40Z on 18 Sep 2026, three ladders, 104 legs, every leg `rc=0`. Everything above the Result section was written and landed on origin (`df015f244`) BEFORE the sitting - the two questions, the ladder table, the ordering argument with its stated price, the guard on what the round does not license, and one candidate refuted by reading - so none of it was fitted to the numbers. The Result section says what actually happened. **In one paragraph.** Interleaving is REFUTED as the cause of the 22-row unpinned `-t16` swing: solo and interleaved landed 2 rows apart in CPU and 5 in wall against one fixture in one sitting. It is the sitting - the four unpinned `-t16` readings on this part span 30 rows, and three of the four are bracketed on the low side by a rung inside its own A/A floor, so **that crossover is not a measurable quantity at this resolution**. The unpinned `-t4` arm is the exception and replicates within 6 rows across two sittings, so the instability belongs to the FULL-BOX pool. The fine band ladder MISSED its band for exactly that reason - a band 6-35 rows wide cannot be aimed at when its location moves 30 - which makes "price the band" a task that needs a PINNED arm first. Two rungs were priced anyway, at 3.2x and 1.8x, and re-pricing the banked ladder with a wall A/A floor shows the **5.7x currently quoted in the standing wall-time rule is unresolved**, its wall gain of +1.17% sitting inside a 1.36% floor. No constant moved. ## Why these two items are one chip and one sitting They are different cells but they share a fixture, and a fixture on this box costs a ~2 min build, a 16 GiB create and a 456-745 s settle. The box is the scarce thing: intel-core-ultra-9-386h is the fleet's **only GFNI-256 part**, so no other box substitutes, and two lanes were queued behind this one. It also makes item 2 **stronger**. Its question is whether INTERLEAVING moves the crossover, and running both ladders against one fixture instance holds the fixture fixed - which is exactly the control that isolates interleaving from the "different fixture instance" candidate the round before could not separate. ## Item 1 - the proportionality band, unpriced the maintainer's wall-time rule (memory topic `nzbfast-wall-time-is-the-deciding-metric`) carries a limit: wall wins a disagreement *"unless there's something excessive and it's an extremely unfair trade, like, 10x the cpu for a little bit better wall"*. **The band between the CPU crossover and the wall crossover is the only region where the two metrics disagree**, so it is the only region where that limit can ever bite. `rounds/crpool4m-2026-09-17/` could not sample it. Its bands are 6-24 rows and its rung grid is 32, so every band fell *between* measured rungs and CPU and wall agreed at every rung it measured. **This matters for something already in the standing rule.** The 0.1x-at-384 / 5.7x-at-416 trade degradation now recorded there came off the banked **UNPINNED** `-t4` (`rounds/crg4-2026-09-16/`), whose band was 41 rows wide. Under controlled placement crpool4m read bands of 6-24, so **part of that 41 was placement and not mechanism**, and those two figures should not be quoted as a pinned result. Ladder `cband` is what replaces them. **A narrow band means small differences, so UNRESOLVED rungs are the expected case and not a failure.** "The trade cannot be priced at this resolution on this part" is publishable and is exactly what the limit needs to know. No tolerance is widened to manufacture a verdict. ## Item 2 - why the same unpinned arm read 365 and then 387 crpool4m's unpinned `-t16` read **387** CPU where `crg4`'s read **365**: 22 rows on the same configuration, binary and fixture shape, larger than the entire 16-row gap that round was chartered to investigate. Two thirds of it was settled with no box time by `rounds/crpool4m-2026-09-17/regrid-and-arm-split.py`, re-run here before taking the box and reproducing cell for cell: - **The rung grid is REFUTED.** On the common grid 320..512 every reading is identical to its own-grid reading (`crg4` `-t16` 365/393, `crg4` `-t4` 381/422, `4mc-t16` 387/411). Not a reduction artefact. - **The swing is LOCATED in the fold arm.** Between the sittings the fold is 5.8% cheaper in CPU (6.1% wall) and the force only 3.0% (2.6%). A crossover is where fold/force = 1, so symmetric noise cancels; that ~3 pp differential IS the 22 rows. What survives: `crg4` ran its two pools **INTERLEAVED** in one ladder (`-Threads 4,16`) where crpool4m ran a solo `-t16`; the fixture was a different instance; and `crg4`'s sitting was noisier. ### A FOURTH candidate, found by reading and refuted before the sitting **The two banked rounds did not run the same harness, and nobody had noticed.** `crg4` ran plib `3b0e254e` + wcomb `1ad3f260` (commit `e11512488`); crpool4m ran plib `1a714068` + wcomb `70fe0efe` (commit `e1731b0ee`). That is **878 inserted lines** across the two files - the affinity feature, `-Payload`, the create-path residency assert and a rewritten foreign-CPU accounting. A harness delta between the two sittings being invisible is exactly the shape of defect this campaign keeps finding, so it was diffed line by line rather than assumed inert. **It cannot reach an unpinned create leg:** - `Invoke-Leg`'s **only** change is the affinity block, guarded by `if ($affWant)`, and `$affWant` is `0` on an unpinned leg - PowerShell treats `0` as false, so the block does not execute and `wall`/`cpu` are produced by identical code. - The create argv is **byte-identical when `-Budget` is unset**: the newer version builds `c -q -t$threads -s$slice -c$m` and appends ` -m$budget` only `if ($budget -ne 'big')`. Neither banked round passed `-Budget`, and neither does this one. - `-Payload` defaults to `random`, which is the historical member-writing loop to the byte, and the fixture directory name is unchanged for it. - `-Residency` on the create path went from silently inert to asserted. An assert refuses a leg; it does not change the work a leg does. **So the harness is not the 22 rows**, and the three candidates the handoff names remain the three. **But one READING does not survive it.** Foreign-CPU accounting *did* change: `Get-OwnPidTree` now resolves an ancestor-walking own-pid set and `Measure-ForeignDelta` is new. So `crg4`'s "12% median" and crpool4m's "8% median" **were measured by two different instruments** and are not a like-for-like noise comparison. The handoff leans on that pair when it ranks the "noisier sitting" candidate; it should lean less hard. In this round `foreign_cpu` is quoted **within** the sitting and never across the two banked ones. ## The cell - **Box** intel-core-ultra-9-386h (Core Ultra 9 386H, GFNI-256, 4 P cores 0-3, 8 E cores 4-11, 4 LP-E cores 12-15, 31.4 GB, Windows 11). - **Binary** parfast from `06d5734b7`, gated on its **4,173,312-byte COUNT and never on a sha256** - the build embeds its own path, so the hash differs per root by construction. - **Harness** `wcomb.ps1` sha `70fe0efe` and `plib.ps1` sha `1a714068`: **crpool4m's exact pair**, chosen so ladder `csolo` differs from the reading it is trying to reproduce in nothing but the fixture instance, the sitting and its position in the sitting. - **Fixture** 16 x 1,024 MiB at 4 MiB blocks, `-c640`: n = 4,096, a 16 GiB corpus - the shape all five banked 4 MiB rounds used. n = 4,096 clears the x86 input floor of 2,048 on its own, so no rung here can fall through to the fold. - **Ladders** three, all **UNPINNED**, `-Phase create`, `fold force force2 fold2` at every rung, one rep, `-Residency resident`, `-NttBudget` 20 GiB: | # | label | threads | rungs | what | |---|---|---|---|---| | 1 | `4mcb-band` | 16 | 384,392,400,408,416 | **item 1** - 8-row spacing across `4mc-t16`'s band (CPU 387, wall 411), the widest of crpool4m's five. Builds the fixture | | 2 | `4mcb-solo` | 16 | 320,352,384,416,448,480,512 | **item 2** - the shape crpool4m read 387 from | | 3 | `4mcb-inter` | 4,16 | 320,352,384,416,448,480,512 | **item 2** - the shape `crg4` read 365/381 from | Ladders 2 and 3 run the **common grid** `regrid-and-arm-split.py` established, so both readings are comparable to 387 and 365 with no re-gridding step and no grid candidate to re-open. Every ladder carries a **distinct `-Label`**: `rowgate.py read` groups by (label, threads), and `cinter`'s `-t16` legs must not fold into `csolo`'s. ### The ordering, and what it costs All three ladders are unpinned, so `-Affinity` arms nothing and the "arm 1 must be unpinned" rule - which exists because `-Affinity` pins the **fixture create** too - is satisfied by every ordering. The ordering is therefore free, and it is chosen for item 2: **`cband` runs first and builds the fixture, so that `csolo` and `cinter` are symmetric** - both post-settle, neither carrying the 16 GiB create and its 456-745 s settle. Item 2's whole question is a contrast between those two, and the contrast is worth more than either one's tie to a banked number. **The price, stated in advance:** crpool4m's 387 was read from a ladder in position 1 that built its own fixture, and `csolo` is in position 2. A `csolo` that does not reproduce 387 therefore has **ladder position** as a residual explanation this round cannot exclude. **The ordering buys back more than it costs**, and this is the reason for it: `cband` and `csolo` are both unpinned `-t16` and **share the rungs 384 and 416**. Comparing their fold and force readings at those two rungs is a within-sitting, same-configuration estimate of exactly that position-plus-repeat term - which no other ordering provides, and which nothing in this campaign has ever measured. It is a free internal control. ## What this round does NOT license, committed to IN ADVANCE Written before the sitting, so it cannot be trimmed to fit whatever the numbers turn out to say. crpool4m recorded a commitment of this shape and it is the reason that round's conclusions held when its numbers would have allowed a stronger claim. **NO CONSTANT MOVES.** The rung decision - 384 against 416 against `create_ntt_min_rows` taking its own clause at ~372 - is **the maintainer's**, it is open, and it is explicitly waiting on wall-based inputs. This lane is the fifth to measure inputs and the fifth not to move anything. No sentence in this write-up may be read as a decision having been taken, and anything downstream citing it as "384 ships" is citing it wrongly. **A PRICED BAND IS NOT A CHOSEN RUNG.** Item 1 produces a trade ratio per rung. A ratio under 10x does not say a rung is right; it says the proportionality limit does not **exclude** it. Those are different claims and only the second is this round's. **AN UNRESOLVED BAND IS NOT A CLEAN BILL.** If the rungs inside the band come back inside their own A/A floors, the honest statement is "the trade cannot be priced at this resolution on this part" - **not** "the trade is small", and **not** "the metrics agree". **ITEM 2 CANNOT CLEAR AN INTERLEAVED LADDER, ONLY CONVICT ONE.** If solo and interleaved land together, that is evidence the mechanism is not interleaving *in this cell*, at this pool pair, on this part. It does not certify the other interleaved ladders in the campaign, which use different pools and shapes. **ONE SITTING IS NOT A REPLICATE**, and more reps cannot firm a rung: the A/A floor is a MAX over reps. **ONE PART, ONE PAYLOAD, ONE BLOCK SIZE.** Core Ultra 9 386H, random bytes, 4 MiB, n = 4,096. No other box on this fleet is GFNI-256. ## Reading rules this round is held to - **WALL DECIDES, CPU EXPLAINS.** Both crossovers are quoted on every table and any divergence is named with its size. `rowgate.py read` prints both. - **CPU-SECONDS DO NOT COMPARE ACROSS MASKS.** Inert here since all three ladders are unpinned, but `cinter`'s `-t4` and `-t16` legs are two different pools and only their CROSSOVERS compare. - **A tight A/A floor is agreement, not correctness.** Every ladder is screened with `rounds/pinaff4m-2026-09-16/ladder-monotonicity-audit.py` and the rungs BRACKETING each crossing are checked for firmness before any crossover is believed. ## Files | file | what | |---|---| | `crband-round.ps1` | the driver: load gates, class probe, extract, build, byte gate, per-ladder ARGV, rc=17 retry | | `crband-driver.log` | the driver's own log | | `logs/cband.log` | ladder 1, `4mcb-band`, unpinned `-t16`, 5 rungs. Builds the fixture | | `logs/csolo.log` | ladder 2, `4mcb-solo`, unpinned `-t16`, 7 rungs | | `logs/cinter.log` | ladder 3, `4mcb-inter`, unpinned `-Threads 4,16`, 7 rungs | ## Result **RAN AND COMPLETE.** One sitting on intel-core-ultra-9-386h, **01:44:51Z-03:04:40Z** on 18 Sep 2026, three ladders, **104 legs, every leg `rc=0`, `match=1` on the cross-arm SHA-256, `residency=resident`, and `lock_waits=0` on all three.** Nothing excluded, nothing re-measured. `cband` 20 legs (27 min, including the 16 GiB fixture create and a 989 s settle), `csolo` 28 (14 min), `cinter` 56 (33 min). Binary built on the box from `06d5734b7` and gated at 4,173,312 bytes; harness `wcomb.ps1` `70fe0efe` + `plib.ps1` `1a714068`, crpool4m's exact pair. Box released and the root deleted; 37.9 GB. ### 1. Item 2 is answered: INTERLEAVING IS REFUTED, and it is the SITTING Against **one fixture instance in one sitting**, the two shapes land on top of each other: | ladder | shape | **CPU** | **wall** | foreign median / max | |---|---|---:|---:|---| | `4mcb-solo` | `-t16` SOLO | **359** | **394** | 8% / 78% | | `4mcb-inter` | `-t16` INTERLEAVED with `-t4` | **357** | **389** | 7% / 32% | **Two rows apart in CPU and five in wall.** Interleaving is not the mechanism, and the 22 rows have to come from somewhere else. Put beside the two banked readings, all four unpinned `-t16` create crossovers on this part: | sitting | ladder | shape | CPU | wall | |---|---|---|---:|---:| | 16 Sep | `crg4` `res4m` | interleaved | 365 | 393 | | 17 Sep | `crpool4m` `4mc-t16` | solo | **387** | **411** | | 18 Sep | `crband` `csolo` | solo | 359 | 394 | | 18 Sep | `crband` `cinter` | interleaved | 357 | 389 | **The spread is 30 rows in CPU and 22 in wall, and it does not sort by shape.** The two solo readings are 28 rows apart from each other; the two interleaved readings are 8 apart; a solo and an interleaved from the same sitting are 2 apart. Shape explains none of it and sitting explains all of it. ### 2. The `-t4` arm REPLICATES, so this is a FULL-BOX problem The same interleaved ladder's other pool, against `crg4`'s, two days apart: | sitting | pool | CPU | wall | |---|---|---:|---:| | 16 Sep `crg4` | `-t4` | 381 | 422 | | 18 Sep `cinter` | `-t4` | **387** | **427** | **Six rows in CPU, five in wall** - the same order as the repair path's PINNED replication (344->349, 378->377, 398->401, within 5). So "an unpinned reading is unreliable" is too broad a conclusion: **the unpinned `-t4` create arm is as reproducible as a pinned one, and it is the unpinned FULL-BOX arm that is not.** That is a sharper statement than the handoff could make, and it is new. It also re-reads the unpinned pool "term". Within this one interleaved ladder the smaller pool crosses HIGHER by **30 rows** (`-t4` 387 against `-t16` 357), where `crg4` measured +16 and crpool4m's *controlled* arms measured the opposite sign (+26 for the bigger pool). So the unpinned pool ordering is not merely confounded by placement, as crpool4m established - **its magnitude is itself unstable by 14 rows between sittings.** ### 3. And no single one of these numbers is precise, which is the real finding Take each crossing and ask whether the rungs BRACKETING it are firm - the test crpool4m's own trap note names. Distance of `F/T` from 1 at the bracketing rung, against that rung's own A/A floor: | reading | lower bracket | upper bracket | |---|---|---| | `crg4` `-t16` 365 | 352: 3.0% vs 1.8% **firm** | 384: 4.6% vs 2.9% **firm** | | `crpool4m` `-t16` 387 | 384: 1.0% vs 3.2% **INSIDE** | 416: 9.2% vs 3.4% firm | | `csolo` 359 | 352: 1.7% vs 2.0% **INSIDE** | 384: 5.9% vs 0.5% firm | | `cinter` `-t16` 357 | 352: 1.5% vs 1.8% **INSIDE** | 384: 7.2% vs 1.6% firm | **Three of the four are bracketed on the low side by a rung whose distance from the crossing is inside that rung's own noise.** `crg4`'s 365 is the only one firmly bracketed both ways, and **the 387 is both the outlier and among the softest**, which crpool4m's README half-said about itself ("the softest reading here is `4mc-t16`'s own, bracketed at 3.2% and 3.4%"). So the honest form of the answer to item 2 is not "an unpinned create reading is worth +/- 15 rows". It is: **the unpinned full-box create crossover is not a measurable quantity at this rung resolution on this part.** Every reading of it is a three-digit number resting on at least one bracket it cannot resolve, and the four of them span 30 rows. Quoting any of them - 365, 387, 359 - to the row is quoting past the instrument. ### 4. Item 1: the fine ladder MISSED ITS BAND, for the reason item 2 just gave `cband` was positioned on crpool4m's `4mc-t16` band, CPU 387 to wall 411. In this sitting that arm's band is **[359, 394]**. The ladder's rungs 384-416 sit mostly *above* it, and its own reduction says so: **CPU crossover `<384`** (it never crossed in range) and every rung from 392 up reads `ntt` in both metrics. | m | cpu F/T | wall F/T | cost% | gain% | aa_cpu | aa_wall | verdict | |---:|---:|---:|---:|---:|---:|---:|---| | 384 | 1.043 | 0.946 | +4.3 | +5.4 | **35.5** | **33.8** | unres | | 392 | 1.100 | 1.015 | +10.0 | -1.5 | 1.2 | 1.6 | agree | | 400 | 1.104 | 1.022 | +10.4 | -2.2 | 1.1 | 0.6 | agree | | 408 | 1.113 | 1.031 | +11.3 | -3.1 | 4.0 | 4.3 | agree | | 416 | 1.138 | 1.054 | +13.8 | -5.4 | 1.7 | 1.2 | agree | **`cband`'s own wall crossover of 390 is NOT usable** and is not quoted as a result anywhere here: it is interpolated across the 384 rung, and that rung is the ruined one (section 6). **This is not bad luck, it is a structural result about the cell, and it is worth more than the table would have been.** The band on this arm is 6-35 rows wide and its LOCATION moves 30 rows between sittings. **A band cannot be pre-positioned to 8-row resolution when its position is only known to +/- 15 rows.** Pricing the create's proportionality band therefore requires a PINNED arm first - which is a concrete, checkable prerequisite the handoff did not know it needed, and which the repair path's 5-row pinned replication says is achievable. ### 5. Two rungs priced honestly anyway, and neither is near the limit `band-trade.py` (this directory) prices the trade at every rung and, unlike `rowgate.py read`, computes an A/A floor on **wall** as well as CPU - which matters because the ratio's denominator is the wall gain, a small number. | ladder | m | cost% (floor) | gain% (floor) | **ratio** | |---|---:|---|---|---:| | `csolo` `-t16` | 384 | +5.9 (0.5) | +1.8 (0.7) | **3.2x** | | `cinter` `-t4` | 416 | +4.8 (0.7) | +2.6 (0.4) | **1.8x** | | `cinter` `-t16` | 384 | +7.2 (1.6) | +0.7 (**1.4**) | unres | Both priced rungs clear both floors comfortably. Re-pricing the banked `crg4` ladder the same way gives **0.1x** at `-t4` m=384 (priced) and **1.9x** at `-t16` m=384 (priced, but marginal: +4.6 against a 2.9% floor and +2.5 against 2.3%). **And it removes a figure from the standing rule.** The **5.7x** at `crg4` `-t4` m=416 - currently a bullet in memory topic `nzbfast-wall-time-is-the-deciding-metric`, described there as "over half way to the 10x the maintainer named as excessive" - has a wall gain of **+1.17% against a 1.36% wall A/A floor on that rung**. Its denominator does not clear its own noise. Hand-checked leg by leg: fold 35.488/35.195, force 36.000/35.517. **5.7x is not a measured quantity and must not be quoted as one.** So the complete set of honestly priced create trade ratios now on record is **0.1x, 1.8x, 1.9x and 3.2x**. The largest is 3.2x. **No create rung has ever been measured at a trade approaching the 10x limit, and the figure that came closest was an artefact of an unfloored denominator.** That does not choose a rung; it says the proportionality limit does not currently exclude any of them. ### 6. The instrument finding: a ladder's FIRST rung is a warm-up ramp `cband`'s m=384 is ruined, and the four legs say exactly how - a monotone rise across the rung, with foreign CPU at 3-5% throughout, so it is not foreign load and the 989 s settle had completed (`FIXTURE-SETTLE ok=1 foreign_cpu=16.9`): ``` fold cpu=248.4 force cpu=271.5 force2 cpu=288.9 fold2 cpu=336.4 ``` `fold2` is 35% dearer than `fold`, and m=392 is already at steady state with a 1.2% floor. **The ramp is confined to the first rung**, and `fold2`'s 336 matches `csolo`'s independent 340 at the same rung. **All three ladders in this campaign that BUILT their own fixture show it, and no other ladder does:** | ladder | first rung | fold -> fold2 | aa_F | |---|---:|---|---:| | `crg4` `-t4` | 320 | 106.1 -> 116.3 | **9.6%** | | `crpool4m` `4mc-t16` | 288 | 162.6 -> 188.1 | **15.7%** | | `crband` `cband` | 384 | 248.4 -> 336.4 | **35.5%** | against 0.1-2% at a typical rung. **Why the medians survive it and the floor does not**, which is the part worth internalising: the arm order is ABBA (fold, force, force2, fold2), so a LINEAR drift of `d` per leg adds `1.5d` to both arms' medians and cancels exactly in the ratio - that is what ABBA is for. A warm-up is not linear, so only the CURVATURE leaks through, while the fold arm's A/A spans three legs of ramp and the force arm's spans one. **The floor blowing up while the verdict stays plausible is the designed behaviour, not a malfunction** - and it is why the first rung must be read as a refusal rather than as a number. **The cheap fix for any future ladder here: spend the first rung.** Put a throwaway rung below the range you care about, or build the fixture on a ladder whose lowest rung you are willing to discard. crpool4m and `crg4` both absorbed this at a rung far below their crossing and were unharmed; `cband` put its first rung *inside* the region it was measuring and lost it. ### 7. The within-sitting position control, which the ordering bought `cband` and `csolo` are both unpinned `-t16` and share rungs 384 and 416, at positions 1 and 2 of the sitting. 384 is `cband`'s ruined rung, so the control reads at 416: | | `cband` (ladder 1) | `csolo` (ladder 2) | apart | |---|---:|---:|---:| | CPU F/T | 1.138 | 1.122 | 1.4% | | wall F/T | 1.054 | 1.041 | 1.2% | Using `csolo`'s local slope (`F/T` rises 0.063 over the 32 rows from 384 to 416), **1.4% in `F/T` is worth about 8 rows.** So ladder position plus repeat inside one sitting is a real but modest term - and it is **well under the 30-row between-sitting spread**, which is the comparison the control exists to make. The stated price of the ordering (section "The ordering, and what it costs") is therefore paid and bounded: `csolo` sitting at position 2 rather than 1 can account for roughly 8 of the 28 rows between it and crpool4m's 387, not for all of them. ### 8. Screening `ladder-monotonicity-audit.py` over all three logs - **and the audit was blind when this lane first ran it; see below**: ``` 4 4mcb-inter NO 1.6% logs/cinter.log full-box (threads>=16) ladders: 3 non-monotone: 0 (0%) sub-box (threads< 16) ladders: 1 non-monotone: 1 (100%) ``` All three full-box ladders are monotone. `cinter`'s `-t4` is not: its dip is 480 -> 512 (`F/T` 1.160 -> 1.142) at floors of 0.5% and 1.2%, **above** its crossing, which is bracketed by 384 (floor 1.0%) and 416 (floor 0.7%). Same shape as every previous sitting's non-monotone arm, and its 387 stands on the stated test. Foreign CPU by ladder: `cband` median 7% / max 43%, `csolo` 8% / 78%, `cinter` `-t4` 9% / 81%, `cinter` `-t16` 7% / 32%. Both 78-81% maxima are on `fold` legs at m=320, below every crossing. ### A tool that was reporting a clean bill over nothing `rounds/pinaff4m-2026-09-16/ladder-monotonicity-audit.py` - the screen this round and every round in the campaign is required to pass - **matched nothing and printed `full-box ladders: 0 non-monotone: 0` with exit 0** on every log, from `9cb6c5fbf` (2026-09-17T21:35Z) when `rowgate.py read`'s table header gained a `create|repair` phase token that the audit's regex did not follow. Found here by running the mandated pre-round screen over `crg4` and getting zero ladders out of the very log whose non-monotone `-t4` ladder crpool4m's README quotes a result for. Fixed in this round's work, and the fix is a pointer repair rather than a new behaviour - it restores **two independently banked screen outputs byte for byte** (`crg4`'s `4 res4m NO 9.6%` and crpool4m's `4 4mc-p4 NO 1.2%` with four of five arms monotone). Reducing zero ladders out of a non-empty file list now REFUSES with exit 2: failing to find is failing, and a screen that cannot locate its subject must not read as a screen that passed. ### What this licenses, and what it does not **It moves no constant and recommends no rung**, per the guard committed before the sitting. Every bullet there stands. In particular the round did NOT price the band as a curve, and nothing here says 384, 416 or ~372 is right or wrong. What it establishes for the decision that is waiting: - **Interleaving is refuted** as the cause of the 22-row swing (2 rows, direct, one fixture, one sitting), and the `cinter` ladder is therefore evidence that the campaign's other interleaved ladders do not need re-reading *for this reason*. It does not certify them at other pools or shapes. - **The unpinned full-box create crossover is not measurable at this resolution.** Four readings span 30 rows and three of the four rest on a bracket inside its own floor. **Every unpinned figure in this campaign inherits that** - including the 381/365 pair the 384 recommendation was argued from, whose 16-row gap is half the spread of one of its own terms. - **The unpinned `-t4` arm is the exception and replicates within 6 rows**, so the instability is a property of the full-box pool, not of unpinned measurement. - **Pricing the proportionality band needs a PINNED arm.** This is now a measured prerequisite and not a preference: an 8-row ladder cannot be aimed at a band located to +/- 15 rows. - **The largest honestly-priced create trade on record is 3.2x**, and the 5.7x that reads as "over half way to excessive" is not a measurement. ### Stated limits - **ONE SITTING**, and more reps cannot firm a rung: the A/A floor is a max over reps. The solo-vs-interleaved contrast is one comparison on one fixture. - **`cband` did not answer item 1 as chartered.** What it produced is a reason the question cannot be asked that way on an unpinned arm, plus one ruined rung that turned into section 6. Two band rungs were priced from the *other* two ladders, opportunistically, not from the ladder built for it. - **ONE PART, ONE PAYLOAD, ONE BLOCK SIZE.** Core Ultra 9 386H, random bytes, 4 MiB, n = 4,096. No other box on this fleet is GFNI-256. - **The position control reads at one rung**, 416, because the other shared rung was the ruined one. An 8-row position term is one measurement. - **Item 2's "different fixture instance" candidate is retired only for the solo-vs-interleaved contrast**, which shared a fixture by construction. The between-sitting comparisons in section 1 still differ in fixture instance as well as sitting, and this round cannot separate those two. ### Owed - **A pinned band ladder.** Pin an arm (the repair path's `e4`-style mask), find its crossovers, then put 8-row rungs across the band it actually has. The pinned arms replicate within 5 rows, so the band can be aimed at. - **Why the unpinned full-box arm is unstable where the unpinned `-t4` arm is not.** 4 P + 8 E + 4 LP-E all in play against four threads the scheduler places consistently is the obvious hypothesis and it is not tested here. - **Spend the first rung**, in every future ladder on this rig.