# Test 1 pilot: the gold, and a review written after the run Written 2026-10-14 14:39 Halifax, after the pilot report of 13:11 Halifax. Saved at Colson's request: "save it all." Colson, on reading the results: "I never really fully digested how meaningful these results actually are in scale and towards what we are trying to discover." - Everything below was written after the results were known. Any interpretation in it is after the fact. - The record is pilot_results.txt, pilot_results.json and pilot_runlog.json, under the freeze in medium_freeze.json. This note doesn't change or replace them. If anything here disagrees with them, they win. - A toy model: structure inside the model, not evidence about cells. - Exploratory: it can suggest, not confirm. - Histories, not memory: every medium-versus-push difference is a difference between two intervention histories and the states they leave, not evidence of an independent memory variable. - The Cause Test FAIL stands as registered. ## The gold > Exploratory pilot, oscillator model, 12 fresh starts: a 0.5% change in the pull for 200 time units, then set back > exactly, shifted later growth in the natural correlation's direction (about 1.3 times the benchmark; 12 of 12 starts). > A push matched to the same coherence did not reproduce it. Same coherence, same settings, different future: what came > next depended on how the state was reached, not only on its coherence. In short: same coherence, different future. Keep this wording whenever the line is used: - "was not reproduced by the one push method tested", not "could not be reproduced". The pilot tested one pull (K), one size, one window length and one push method. - "suggests", not "indicates" or "confirms". - The future that differed is later growth. How fast the extra coherence faded did not clearly differ. ## Why this yes counts The same model, with the same kind of controls, has already said no. The Cause Test pushed coherence directly, and its up/down difference S came to +0.00031 ± 0.00015 against a half-benchmark line of 0.00104. Its FAIL stands. A test that has shown it can say no is one whose yes carries weight. ## The scale, in three numbers 1. **A small lever.** Raising the pull by 0.5% for 200 time units brought coherence at the restoration point R to 1.81 times the untouched runs (0.0457 against 0.0252). Lowering it brought coherence to 0.65 times (0.0165). The model sits right at its threshold (K + ε = 2Δ), where a small change in conditions has a large effect. 2. **A clean lean.** S = +0.00303 ± 0.00036 (8.4 error bars), with 12 of 12 starts above zero. - Suppose there were no direction-dependent effect, the starts were independent, and each was equally likely to lean either way. Then all 12 leaning the same way (either way) would happen by chance about once in 2,000 tries (2 × 0.5¹² = 1 in 2,048). - Its size is about 1.3 times the benchmark of 0.00234, which is the natural correlation's slope applied to this coherence difference. - On the same kind of scale, the Cause Test's pushes reached about 0.15 (causal slope −0.0238 against the benchmark slope −0.1602). The designs differ, so this comparison is rough and made after the fact. 3. **The gold.** Same coherence at R, the same K = 1.9 afterwards, different later growth. Medium against its matched push, like for like: S difference +0.00181 ± 0.00051 (3.6 bars; 10 of 12 starts above zero); P difference −0.00126 ± 0.00040 (3.1 bars). ## The arc across three tests 1. **Test 17:** higher coherence goes with slower later growth. The correlation replicated. 2. **The Cause Test:** pushing the coherence number directly didn't carry that link (FAIL, standing). 3. **This pilot:** changing how the parts interact for a while did carry it, and a push to the same coherence didn't. Colson chose to change the conditions of interaction rather than push the state again, and inside this model that is where the effect appeared. The coherence number alone didn't decide what came next; the arrangement behind it, shaped by how it was reached, did. For the harm-reduction lens, the parallel inside the model is this: at a tipping point, gently changing the conditions did what forcing the state didn't. It is a statement about the model, not about cells. ## Keeping it exact - **Most of the effect came from raising.** D_raised = −0.00541 ± 0.00088 (−6.1 bars); D_lowered = +0.00066 ± 0.00067 (not clear). Lowering had little room: in calibration, coherence bottomed out near 0.62–0.66 of untouched at every larger size. - **Effort was unequal.** The raised arm delivered about 1.75 times the extra pull (0.0660 against 0.0378) and carried most of the effect, so this design can't separate direction from amount. - **Later growth differed; persistence did not clearly differ.** Of the raised-minus-lowered coherence difference at R, 0.350 was still there over (R + 200, R + 500] for the medium, against 0.298 for the push; the difference is 1.6 bars. - **No lasting switch.** The medium's coherence difference was still there over (R + 400, R + 500] but clearly declining. - **The pushes repeated the Cause Test's pattern.** Against their shams they were non-monotonic (P −6.9 bars); against untouched, they had no clear effect. The shams themselves raised growth (+4.0 bars). - **Counts.** The medium arms ran at all 120 moments. Matched pushes and shams were built at 119 of 120 (start 113, moment 2, lowered: not built). - **A pilot.** Twelve starts, exploratory. A separate registered replication would be what confirms it; this pilot is not to be reread as one. ## Which features of the state differ (descriptions, read after the fact) These come from the state descriptions at R in pilot_results.txt, averaged over the groups. They were reported as descriptions and not used for any reading. - **Where the extra coherence sits.** - Raised: the medium put nearly all of it in the slower-turning half of each group, the oscillators closest to the group's own rhythm. That half reached 0.0931 against 0.0490 untouched, while the faster half stayed unchanged (0.0203 against 0.0200). - Raised: the matched push spread it across both halves, 0.0690 in the slower half and 0.0338 in the faster. - Lowered: medium 0.0304 slower and 0.0199 faster; push 0.0405 and 0.0224. - A change of pull mostly moves the oscillators it can draw in; a push moves every oscillator. - **The timing between groups A and B.** Raising the pull drew the groups closer (0.3682 against 0.7891 untouched); lowering pushed them apart (1.1910). The pushes left the timing alone (0.7825 and 0.7806). - **What doesn't differ.** The higher-order measures |z2|, |z3| and |z4| stay between 0.0118 and 0.0135 in every state, so there's no sign of new clustering. - **The phase distance** between medium and push states is rms 0.726 raised and 0.421 lowered, after lining up the timing. It's mostly what chaos does to the drifting oscillators over any 200-unit change. It shows the states differ, not which difference matters. ## Candidate explanations (hypotheses, after the fact; none tested) - **Slowly built coherence.** - Natural coherence already sits mostly in the slower half (0.0490 against 0.0200 untouched). The follow-up found the natural link is carried by coherence over the past 100 to 500 time units. - The medium builds coherence the same slow way, through the system's own dynamics, so it may take on the natural direction. - The push builds a different kind, partly in faster oscillators that drift off. Its coherence faded a little faster (0.298 against 0.350), though not clearly. - The plan named this possibility before the run ("It might reach what a push didn't"). The mechanism is a hypothesis. - **The timing between the groups.** If the groups stay closer after K is restored, the faint link between them keeps adding a little pull. The past change would keep acting through the state. - **The constraint on both.** In this model, once K is back at 1.9, the phases are the whole state. Any lasting difference has to be carried by the phase arrangement; nothing is stored anywhere else. ## What the saved results could check, read-only All three checks were run on Oct 14 at Colson's go; their results are in the next sections. At his request, check 3 was read as a fluidity question. The saved results include, for every arm, each group's coherence and rhythm at every time unit and the gap growth per 5 units, plus the state descriptions at R for each moment. Without running the model, they could show: 1. whether the A–B timing difference lasts into the growth window (R + 200 to R + 500) or fades before it; 2. whether moments with bigger slower-half or timing differences show bigger medium-versus-push growth differences; 3. whether naturally high-coherence moments in the untouched runs also carry their coherence in the slower half. Each would be after the fact and labelled exploratory. The states of the individual oscillators after R weren't saved. Going deeper would mean re-running arms, which is a new step and Colson's decision. ## Check 1 result (2026-10-14 14:47 Halifax; after the fact, exploratory) Colson's go: "yes begin with the first check." The full output is check1_timing.txt and check1_timing.json, made by check1_timing.py. It read only the 120 saved moment files, each matching its recorded hash, and ran no model. Its measure reproduces the report's A–B timing at R exactly. The split is half of (the lowered arm's timing gap minus the raised arm's). Positive means the raised arm's groups sit closer in timing. - **Answer: the medium's timing shift lasts.** Its split was +0.411 at R (12.1 bars) and +0.257 over the growth window (16.6 bars; 12 of 12 starts), about 62% of its size at R. It fades slowly and is still clear at R+400 to R+500 (+0.212, 11.7 bars). For comparison, about 35% of the coherence difference was left over the same window. - **The push catches up.** At R the pushes had no timing shift (−0.001), as built. Within the first 100 units they developed most of it on their own (+0.266 against the medium's +0.379 in that block), and +0.190 over the growth window (9.6 bars). - **The difference that's left is small.** Medium against matched push, like for like: +0.407 at R, +0.111 in the first 100 units, +0.065 over the growth window (4.4 bars; about a sixth of its size at R), and +0.041 at R+400 to R+500 (1.8 bars, not clear). - **No common part and no sham effect.** Changing the pull either way didn't shift the gap on average (−0.038 over the growth window, 1.5 bars), and the shams showed no split (+0.018, 1.5 bars). Reading, after the fact: - **The timing between the groups looks mostly like a follower of coherence level.** - When each group is more coherent, the two groups' rhythms fall into step within about 100 time units. When they are less coherent, the rhythms drift apart, partly because noisier rhythms average a wider gap. - The medium had a 200-unit head start, and the push caught up most of the way. - The medium-versus-push coherence difference shrinks over the same period. For the raised arms it went from +0.0043 in the first block after R to +0.0003 in the last (from the report's curves). - **Timing is now a weaker candidate for what the push didn't reproduce.** The push reproduced most of it, only later. It isn't ruled out, because a small medium-specific part remains over the growth window. - **The two courses don't line up simply.** - The medium-versus-push timing difference is largest right after R and shrinks steadily. - The growth difference between them is smallest in the first 100 units (+0.00107, 2.2 bars) and largest from R+100 to R+300 (+0.00219 and +0.00231). It is still about +0.0017 at the end (2.3 bars). - A delayed effect of timing could fit this, and so could a slower feature. - **Something slower may be at work.** By the last block, the two arms' coherence and timing have nearly converged, while their growth difference hasn't clearly shrunk. That points, untested, toward something slower than the current coherence and timing, such as which oscillators hold the coherence or how long coherence was held. This check doesn't touch the slowly-built-coherence candidate. ## Check 2 result (2026-10-14 15:05 Halifax; after the fact, exploratory) Colson's go: "run it." The full output is check2_moments.txt and check2_moments.json, made by check2_moments.py. It read only the 120 saved moment files, each matching its recorded hash, and ran no model. Before reading on, it reproduced the report's growth difference, the size of the coherence change at R, the state descriptions at R and check 1's timing difference at R. Each difference is medium minus matched push, moment by moment, over the 119 moments with both pushes built. The main measure is the correlation with the growth difference within each start, averaged over the 12 starts. Its error bar came out near 0.1, as expected, so only a correlation of about 0.3 or more could be clear. - **Answer: not clear for either.** - Slower half: +0.064 ± 0.104 (0.6 bars; 8 of 12 starts above zero). In the raised arm alone, where most of the effect sits, −0.094 (0.9 bars). - Timing: +0.168 ± 0.083 (2.0 bars; 7 of 12 starts above zero). In the raised arm alone, +0.222 ± 0.076 (2.9 bars, just under the line). - Each start's half of moments with the biggest difference, against its half with the smallest, showed no clear growth difference for either (−0.4 and +0.3 bars). - **The slower-half difference mostly tracks the size of the change.** Within starts it correlates +0.832 with the size of the coherence change at R (24.5 bars). It also goes with timing (+0.337, 4.5 bars), and timing goes with the size of the change (+0.327, 6.8 bars). - **The growth difference didn't clearly scale with the size of the change either** (+0.136, 1.1 bars). Reading, after the fact: - **This check neither supports nor rules out either candidate.** Each moment's growth difference is noisy, so only strong moment-to-moment links could have shown. - **The medium's slower-half shift looks like a steady signature.** It grows with the size of the change rather than varying on its own from moment to moment, which leaves this check little independent variation to test it with. - **Timing leans the expected way without reaching clear, and its halves comparison shows nothing.** Check 1's finding stands: the push reaches most of the timing shift within 100 units. ## Check 3 result (2026-10-14 16:01 Halifax; after the fact, exploratory) Colson's go: "run it once accommodated for." At his request the check was read as a fluidity question: order that holds while everything keeps moving. The full output is check3_fluidity.txt and check3_fluidity.json, made by check3_fluidity.py. It read only the 12 untouched base runs and the 120 saved moment files, each matching its recorded hash, and ran no model. Before reading on, it reproduced the base runs' coherence record, the report's state descriptions at R and the report's persistence (largest difference 3.3e-16). The natural states are the untouched run's saved states at T and at R of every moment: 20 per start, both groups, 480 group states. A group's coherence splits exactly into three contributions along its own rhythm: the slower half of the pulled oscillators, the faster half and the 128 free turners. The three questions and their readings were fixed in the script before any numbers were seen. - **1. Where natural order sits: more in the slower half than an even spread.** - When a group is more coherent on its own, the slower half supplies 0.810 ± 0.029 of the extra, the faster half 0.181 ± 0.033 and the free turners 0.008 ± 0.007. - An even spread would give the slower half 0.484. Its excess over that, +0.326 ± 0.029, is 11.4 bars (12 of 12 starts). - On average the slower half makes up 0.853 of a group's coherence. - **2. The same relationship across routes: like the medium.** - The slower half's share of the extra, from each half's own coherence: natural rises 0.928 ± 0.028, the medium 0.994 ± 0.005, the push 0.714 ± 0.024. - Natural rises differ clearly from the push (+0.213, 10.5 bars) and not clearly from the medium (−0.066, 2.5 bars). The medium is a little more concentrated still, leaning but under the line. - Per unit of extra group coherence, the slower and faster halves gained +1.53 and +0.12 in natural rises, +2.15 and +0.01 with the medium, and +0.98 and +0.39 with the push. - Against untouched, the medium raised and lowered only the slower half (the faster half moved by 0.0003 and 0.0001). The push raised both halves, and the lowering push left the faster half slightly more coherent than untouched (+0.0024, 3.1 bars). - **3. Which order lasts: the slower half's.** - Of each part's contribution at a moment, this much was still in the group's coherence over the next 100 units (the three parts read together): slower half +0.380 ± 0.030 (12.8 bars), faster half +0.056 ± 0.049 (1.1 bars). Slower minus faster: +0.324 ± 0.039 (8.3 bars; 12 of 12 starts). - The slower half's part stays clear through the third block, to t + 300 (+0.260, then +0.165 at 4.3 bars), and fades after it (+0.073, then +0.027). The faster half's part isn't clearly there in any block. - A natural rise as a whole fades faster than either route's excess. About 0.32 of it is left over the next 100 units and 0.07 over t + 200 to t + 500. The medium's excess kept 0.877 and 0.350 over the same spans, the push's 0.676 and 0.298. Reading, after the fact: - **The system's own order and the medium's sit in the same place.** In this model, natural rises in coherence come mostly from the slower half: the oscillators whose natural rates are closest to the group's own, which can stay in step while still moving. The medium put its extra there too, even more concentrated. The push spread more of its extra into the faster half. - **Order held in the slower half lasts; order in the faster half doesn't clearly last.** In the fluidity picture, the order that survives the system's own fluctuations is the order held by the oscillators that keep moving with the group. - **Where order sits isn't the whole story of how long it lasts.** Natural rises sit mostly in the slower half, yet even the slower half's natural part kept only 0.380 over the next 100 units, against the medium's 0.877. The designs differ: natural rises compare moments with each other, while the routes compare arms at the same moment. One reading, untested: a single moment's natural coherence includes fast chance fluctuations from the finite number of oscillators, while the medium built and held its excess for 200 units. That fits the follow-up's finding that the natural link is carried by coherence over the past 100 to 500 time units. - **What check 3 doesn't show.** It describes where order sits and how long it lasts; it doesn't test what carries the growth difference. The slowly-built-coherence candidate is consistent with all three answers and remains a hypothesis. The push's excess faded somewhat faster than the medium's in every block, but no test of that difference was fixed here; over the growth window the report found it not clear (0.298 against 0.350). ## Wording notes for anyone summarizing this pilot - **Direction was a hypothesis.** "Same direction" means the natural correlation's direction, one of five shapes fixed before the run, not a predicted direction. - **119 of 120** counts the moments with both matched pushes and shams. The medium arms ran at all 120. - **"The outcomes differed"** means later growth. Persistence did not clearly differ. - **The pull was restored** to exactly K = 1.9. - **Memory.** Don't write that a memory field is "not necessarily" required; in this model there is none. Once K is back, the phases are the whole state. - **"Much larger than the Cause Test"** holds in benchmark units (about 1.3 against about 0.15), across different designs. - **"A remarkably small change"** holds because the model sits at its threshold.