# Amendment 1 to the final scattering-side design: comparison A comes out of the pilot **Draft, October 15, 2026, for Colson's approval.** It amends the locked plan, plan-final.md (SHA-256 0e3f8f114ed44eade7e211317217f23e27c1075008e561c2b07ea11645823121), locked by lock.json (SHA-256 3c7f70973247a327067a1817766b0dae3e0d72154f1de7785df7213ff271b23d). Neither file is edited; both stay word for word. Nothing has run for this amendment. Nothing runs until he approves it and, separately, gives his go for step 3. His words, on the results of steps 1 and 2 (he chose "Draft an amendment for comparison A"): "Professor Squirrel, please draft a dated amendment for comparison A only. My preference is to remove comparison A from the upcoming fresh-start pilot rather than replace its unsuccessful moment-by-moment matching with average matching. Preserve the 10/20 matching result and the original 18/20 feasibility requirement as part of the experimental record. Keep comparisons B and B′ unchanged, including the 0.6 and 0.4 intervention levels, outcome definitions, margins, and counter-rotating rhythm detection criteria. Please explicitly preserve the old-start results as feasibility observations, not prospective confirmation. Before any pilot, clarify whether the existing sample size and uncertainty rules are sufficient to evaluate the predicted difference between the two faster-half levels. Show me the amendment and revised computational budget for approval. Do not execute fresh starts 135–146, alter the website, or modify any previous experimental locks. ~ Draft amendment only." ## Why Step 2, the window-only check on old starts 16 and 17, asked whether arm U's search could match the raised pull's coherence at R within a relative 1% at 18 or more of the 20 moments. It matched at 10. Of the 10 unmatched moments: - at 8, all 8 runs were used without coming within 1%; - at 2, the secant pointed to c above 1, so the search stopped. At one of them, the raised pull itself ended below untouched at R. Coherence at R jumps when c changes by tiny amounts. At start 16, moment 6, c = 0.994807 and c = 0.994876 left arm U's coherence at R 0.0013 above and 0.0087 below the raised pull's, against a tolerance of 0.0004 (1% of the raised pull's coherence at R). The locked plan says that if step 2 shows the pilot can't work as written, a dated amendment is proposed for Colson's go before the pilot. This is that amendment. ## What changes: comparison A only 1. **Comparison A is removed from step 3**, the exploratory pilot on fresh starts 135 to 146. Arm U and its search aren't run in the pilot, and comparison A's readings, labels and margins aren't computed or reported there. 2. **No other matching takes its place.** Average matching, or any other replacement for the moment-by-moment search, is not adopted. 3. **The pilot therefore runs four arms at each moment, not five:** the untouched rerun, the raised pull, arm F (c_f = 0.6) and arm F′ (c_f = 0.4). Nothing the plan records for arm U alone is recorded, and the numbers it records "for the raised pull and arm U" are recorded for the raised pull only. The raised pull stays, because B1 is read as a share of its gain. Nothing else changes. ## What stays on record - **Comparison A as planned.** Its section in plan-final.md stays word for word: the measures, margins and labels, the expected response, and the pilot's rule that more than 1 unmatched moment in 10 makes its readings inconclusive. None of them is changed; the comparison just isn't run in this pilot. - **The original feasibility requirement:** arm U matching within 1% at 18 or more of the 20 moments (plan-final.md, step 2). - **The result: 10 of 20, requirement not met.** It stays in window_check.txt and window_check.json, with every unmatched moment, its reason and its closest run, as sealed by change-2026-10-15-3. - **Comparison A isn't scheduled anywhere else.** Bringing it back in any form would be a new proposal for Colson's go. ## What doesn't change: comparisons B and B′, and everything else - **The two faster-half levels,** c_f = 0.6 and c_f = 0.4, and the theory's expectation at each: no edge order at 0.6, and an edge rhythm at 0.4 (a counter-rotating pair turning at about ±0.96). - **The outcome definitions:** B1, the gain at R as a share of the raised pull's gain; B2, the faster half's own coherence at R, arm minus untouched; and the rhythm readings E-pair, E-any and E-steady. - **The margins and labels:** - ±0.25 for B1 and ±0.005 for B2; - 0.05 and 0.10 for the rhythm readings; - the labels for every reading that has them. B1 at 0.4 and E-pair at 0.6 are described, not labelled, and E-steady is used only inside the combined readings; - the combined readings at each level, checked in their fixed order. - **The counter-rotating rhythm detection criteria:** - L ≥ 1.5, measured against each arm's own natural rates, in intervals 0.03 wide inside the band 0.80 to 1.30, each direction read on its own; - S ≥ 0.057; - the read over the window's last 50 units, (T + 150, R]; - the rate range 0.90 to 1.10 for the predicted rhythm. - **The uncertainty rules:** starts are the units. The bar is the larger of the resampled bar (400 resamples, seed 20261016) and SD/√n, and shares take the resampled bar. - **The sample:** fresh starts 135 to 146 (seeds 134 to 145), with 10 moments each by the pilot's rule and draws: 120 moments, every start counting equally. - **The arms that remain, and the model:** - the raised pull: K × 1.005 for 200 units from T, then 500 units at 1.9; - the untouched rerun, and arms F and F′; - each run with both copies for 700 units; - the model, its settings and the 200-unit window. - **The step-1 code checks:** all 11 stay as locked. They run again with the pilot's code before the pilot, as the commands require. The checks on arm U stay too: they test machinery the other arms share, and they cost seconds. - **The ground rules:** a toy model; histories, not memory. Nothing leaves this folder, there's no change to the website, and nothing is published. Every earlier lock and record stays exactly as it is: lock.json, the Cause Test, the medium pilot and the Fluidity Replication. ## The old-start results are feasibility observations Everything step 2 produced on old starts 16 and 17 is a **feasibility observation, not prospective confirmation** of either level's prediction. That covers: - arm U's 10 of 20; - the first look at B1 and B2; - the rhythm rule's numbers: no rhythm in untouched; none at 0.6; rhythms at 0.4 in 68 of 80 directions, with a median |Ω*| of 0.964. They carry no labels, are never pooled with the fresh starts, and never enter the pilot's readings. The plan and the lock already say so; this amendment repeats it on purpose. Only the fresh-start pilot can suggest anything about the predictions, and only a locked replication could confirm. ## Before the pilot: can 12 starts evaluate the predicted difference between 0.6 and 0.4? This is a clarification, not a change. **How the plan reads the difference.** B and B′ are each read against the same untouched reruns at the same 120 moments, and they're never pooled. The predicted transition shows as the two combined readings side by side: - at 0.6: "No additional edge synchronization, as the theory expects"; - at 0.4: "The predicted counter-rotating edge rhythm". Both levels share the same starts and moments, so the side-by-side reading compares like with like. There's no separate test of 0.4 minus 0.6. It isn't needed for that reading, and none is added. **What each labelled reading needs, against the spread the old starts suggest.** The old-start spreads come from two starts, projected to 12, and are for feasibility only: two starts estimate spread poorly. | Reading | What its labels need from the bar | Old-start spread, projected to 12 starts | Resolves? | | --- | --- | --- | --- | | B1 at 0.6 (±0.25) | "as expected": the reading ± 2 bars inside ±0.25, so a bar under 0.125 | 0.016 (between starts), 0.046 (between moments) | Yes, unless the reading lies between about 0.16 and 0.25 either way | | B2 at 0.6 (±0.005) | "as expected": a bar under 0.0025 | 0.0005 and 0.0010 (the plan's random-phase estimate: about 0.001) | Yes, unless the reading lies between about 0.003 and 0.005 either way | | E-any at 0.6 (0.05), and E-steady at 0.6 (used only in the combined readings) | "as expected" for E-any, and "inside" for E-steady: the reading ± 2 bars inside ±0.05 | 0 (no rhythm and no steady turning at 0.6 or in untouched at the old starts) | Yes, unless rhythms or steady turning at 0.6 are common enough to put a reading near 0.05 | | B2 at 0.4 | "as expected": more than 3 bars above zero and at least 0.005. "Contradicts": inside ±0.005, so a bar under 0.0025 | 0.0075 and 0.0030 | "As expected" yes, for a clear rise. "Contradicts" not at this spread (see below) | | E-pair and E-any at 0.4 (0.10) | "as expected": more than 3 bars above zero and at least 0.10. "Contradicts": inside ±0.05 | E-pair about 0.04 (0.60 and 0.80 at the two starts); E-any about 0 (1.00 at both) | Yes, unless a reading lies between about 0.05 and 0.12 | | Median \|Ω*\| at 0.4 (0.90 to 1.10) | a range, no bar | every rhythm at the old starts within 0.948 to 0.987 | Yes | | B1 at 0.4 | described, not labelled, so no bar requirement | 0.081 and 0.058 | Not applicable | **The one limit is B2 at 0.4.** If the faster half's coherence varies between starts as much as it did at the two old starts (+0.074 and +0.038), its bar, about 0.0075, is too wide for "contradicts". That needs the reading ± 2 bars inside ±0.005. This doesn't block the predicted outcome's label: "as expected" needs more than 3 bars above zero, and a rise of a few hundredths clears that. The opposite outcome would bring its own, smaller spread. If no order formed at 0.4, B2 would sit at the level random phases give, where the bar is about 0.001, and "contradicts" could be reached. Only a modest rise that varies widely between starts would come out unresolved, and unresolved is the honest reading for that case. **The uncertainty rules with 12 starts.** With 12 units: - "± 2 bars" covers about 93% rather than 95% (95% would need 2.20 bars, from t with 11 degrees of freedom); - "more than 3 bars" corresponds to about 1.2%, two-sided. That suits an exploratory pilot, which the plan says can suggest, not confirm. Confirmation would come from a locked replication (step 4), its own decision. **Answer.** Yes: the existing sample (12 fresh starts, 120 moments) and uncertainty rules are enough to resolve each labelled reading that the predicted difference depends on. Each reading resolves except in a narrow zone near its own margin, which is what the "unresolved" label is for. Removing comparison A doesn't touch B's or B′'s sample or rules. No change to the sample size, the bars or the labels is proposed. ## Revised computational budget A copy-unit is one copy of the model stepped through one time unit. | Step | What runs | Computing | Commands | Saved | | --- | --- | --- | --- | --- | | Checks again | the 11 locked step-1 checks, with the pilot's code | about 30 seconds | 1 | under 1 MB | | Pilot base runs | 12 fresh starts (seeds 134 to 145), up to 7 side by side | about 6 minutes | 2 | about 64 MB | | Pilot moments | 120 moments, each with four arms of 2 copies × 700 units (5,600 copy-units) | about 20 minutes | about 6 to 7 | about 92 MB | | In all | | about 27 minutes | about 10 | about 160 MB | - **Timing basis: measured, not assumed.** - Step 1 timed one pilot arm directly: 2 copies, 700 units, with every reading, in 16.7 seconds. Four arms make about 67 seconds per moment on one worker, so 120 moments on 7 workers take about 20 minutes. - Step 2's base runs took 149 seconds each (two side by side). The medium pilot's took about 172 seconds with 7 side by side. The 12 starts fit in 2 pieces, about 6 minutes. - **Storage basis:** step 2's saved files. A base run is 5.3 MB. Each arm's readings come to about 192 KB at 700 units: every pulled oscillator's turning rate over the read, and the readings every unit. - **Against the locked plan's estimate** (about 37 minutes in about 11 pieces, about 160 MB): - Removing arm U and its search saves about 14 minutes by the same measure: about 40 minutes with arm U, against about 27 without. Arm U's search averaged 6 one-copy window runs per moment in step 2, about 32 seconds. - Storage stays about 160 MB (it would have been about 180 MB with arm U). The plan underestimated each arm's saved readings, as step 2 showed (21.7 MB against "about 12"). - **How it runs:** every command runs in the foreground with a hard limit under 5 minutes and up to 7 worker processes that end with it, and leaves nothing running. Saved runs go to generated-artifacts/scattering-side/, kept out of the project's history and the published app. ## What approval would mean - **If Colson approves this amendment,** a dated record seals it with his words, beside plan-final.md and lock.json, and it becomes part of the plan the pilot follows. plan-final.md and lock.json stay word for word. - **Approving it doesn't start the pilot.** Step 3 needs its own go. Under that go: 1. a dated pilot lock fixes the pilot's settings before its code is written: the fresh starts and their seeds, the four arms, the error-bar resamples and their seed, and the commands; 2. the pilot code is added; 3. the 11 step-1 checks run again with it, and if any fails, everything stops and the failure is reported; 4. the pilot commands run, and their results come back to Colson.