# Status of the two pre-lock checks

## October 13, 2026, about 13:35 Halifax: one-push check not run; E2 check held; planning paused

- **One-push check: not run**, by Colson's choice. Its M4 shortfall (about 0.6 natural spreads against the 1 needed)
  stays a pre-run estimate from numbers already on record, not a measured failure. The check has no verdict.
- **E2 fresh-gap check: held, not run.** It is planned in e2_check_plan.md, with C1 reported only and C2 (stability),
  C3 (agreement) and C4 (κ_E2) keeping their pass marks. No E2 number has been computed.
- **Paused:** further experimental planning and running wait on Colson's review. The cause test's FAIL, the
  calibration FAIL, both pinned holds and H2 (NOT FEASIBLE at ±25%) stay on record as written. The separate
  future-hypotheses backlog is untouched.

## October 13, 2026, about 13:20 Halifax: C1 chosen; the one-push check planned, not run

**Colson's answer (about 13:00).** C1 is reported only (filled in in e2_check_plan.md); C2, C3 and C4 keep their pass
marks. The one-push design check is prepared for review and not run. H2 stays NOT FEASIBLE at ±25%, with every
unsuccessful attempt and the first check's pinned-hold failure on record. The one-push design is a new question, not a
success of the sustained-hold hypothesis, and nothing from the H2 outcomes is used to tune it. The cause test's FAIL is
unchanged, the A/B/C comparison stays a separate future study, and no RGMC backlog concept enters the protocol. No
fresh starts, no lock, no publication.

**The plan (one_push_check_plan.md).** The draft experiment compares a ±25% push at T − 500 with the same push at T,
each with up, down and sham, beside untouched. Its primary contrast is the timing difference Δ = S_E − S_L, judged
against a threshold from the early push's achieved history change; M1, M4 (unchanged), M5 and M6 must hold. The check
on old starts 13 to 18 measures the pushes, the history, the coherence left at T and the effort, with no growth.

**Expected shortfall, stated before any run.** By the first check's numbers, one ±25% push at T − 500 would change the
500-unit history by about 0.0034 (about 0.6 natural spreads), against M4's 0.0057. So the check would very likely end
NOT FEASIBLE. At ±50% the cause test's numbers give about 0.0052, still short; it is not proposed. M4 is kept as
written.

**Nothing has been run since the H2 check.** The E2 check and the one-push check both wait on Colson's approval.

## October 13, 2026, about 12:55 Halifax: Colson's choices, the pause test, the H2 check and the E2 plan

**Colson's choices.** Outcome E4: the carried gap (E1) and a fresh gap made at T + 200 and counted over
(T + 300, T + 500] (E2) must agree for a decisive result, each conditional on passing its own pre-lock validity check.
Hold H2: correct only the slow drift. No exploratory early look at the fresh gap. The revised checks run on old starts
only: no fresh starts, no lock and no publication. If a measurement or the slow-drift hold remains infeasible, a
simpler alternative goes to him rather than forcing the design through.

**Pause test (11:35 to 11:37; pause_test_results.txt).** Old start 7, with the nudge test's run code unchanged, was
run in one go and with one pause at t = 180. The paused run parts from the unbroken one at the first put-back after
the pick-up (t = 185). So the pauses are enough to explain the departures of starts 1 to 12; that they caused them is
not shown, since the pause points were not recorded. The corrections that followed are at the end of
prelock_amendment.md.

**H2 check (plan: hold_check_plan.md, about 12:00; run 12:01 to 12:48): NOT FEASIBLE at ±25% as specified.** No
candidate λ met the criteria for the 500-unit hold on old starts 13 to 15, so by the plan's rule the confirmation on
starts 16 to 18 was not run.

| | needs | λ = 0.05 | λ = 0.10 | λ = 0.20 |
|---|---|---|---|---|
| F1 moments carried out | 29 of 30 | 26 | 24 | 17 |
| F2 history achieved | 90% of moments | 4% | 21% | 82% |
| F3a difference sustained | 90% of moments | 65% | 83% | 94% |
| F3b held evenly | every block | not met | not met | met |
| F4a and F4c sham effort | every moment | met | met | met |

- Firmer correction held the level better but more often asked a kick for more than one kick can do. Every moment
  lost had a kick that could not move a group's coherence as far as asked (at λ = 0.20, 11 of the 13 in the down arm).
- At λ = 0.20, on the 17 moments carried out: ΔM = 0.00661 against 0.00648 planned (1.2 natural SDs of 500-unit
  history means). The sham matched up's and down's effort (Σ Δθ² and Σ |Δθ|, ratio 1.000, the same kick times), and
  its history stayed near neutral (+0.05). Its effort was Σ Δθ² ≈ 6, against 216 for the first check's pinned hold.
- The 100-unit hold was not judged: the plan judges it on the confirmation. On starts 13 to 15 it was carried out in
  29 of 30 moments at every λ, but its history fell short (A_up 0.69 to 0.81, A_down 0.54 to 0.73; F2 at most 21%).
- The dry run's numbers (one moment on old start 1, seen before the check) count toward nothing and are listed in the
  results. The base runs and the λ stage took 39 minutes of running (47 in all), against the plan's 30 to 45 minutes
  for everything, confirmation included.
- Results: hold_check_results.txt and .json. Every part matches its run log; every piece ran in one numerical
  environment; the cause test's locked files and the protected folders are unchanged.

**E2 check: planned, not run (e2_check_plan.md).** Old starts 1 to 18, untouched, with new moments and new directions.
C2, C3 and C4 are needed (C4: κ_E2 against the carried gap over the same 200 units, between 0.75 and 1.25). C1 (lining
up) waits on Colson's choice: reported only (recommended), or the calibration's pass mark. The plan also sets E4's
agreement rules and makes a failed calibration or manipulation check inside the experiment INCONCLUSIVE.

**What follows.** By the plan, H2 does not go ahead, and a simpler alternative goes to Colson, not run: one ±25% push
at T − 500 (the first check's pulse, moved earlier; it was carried out in 60 of 60 moments, with the sham's effort
matched exactly), then untouched, with E4 as the outcome. It gives up a held level: in the first check, roughly half
the change remained 200 to 500 units after a push. Whatever is chosen gets its own check on old starts before any
lock. No new start has been run, and nothing has been published.

## October 13, 2026, about 11:15 Halifax: both checks finished

**Calibration (old starts 1 to 18, untouched only): FAIL**, read with the revised exactness check (V1-revised,
prelock_amendment.md). Read with V1 as preregistered, it gives no reading (6 of 18 starts reproduce their saved runs).
C1 not met: 22.6% of fresh gaps lined up within the window (needs 90%). C2 met: ratio 0.30. C3 met: r = 0.873 ± 0.007.
C4 not met: κ = 0.579 ± 0.024 (needs 0.75 to 1.25). Eleven pieces, 08:29 to 10:00, about 35 minutes of running.
Results: calibration_results.txt and .json.

**Feasibility (old starts 13 to 18, ten moments each, ±25%): neither hold length is feasible.** Its base runs met V1
as preregistered (all 6 reproduce their saved runs exactly) and V1-revised.

- 500-unit hold: F1, F2 and F3 not met (50 of 60 moments carried out; 82%; 8%). Both sham checks (F4) not met.
- 100-unit hold: F1, F2 and F3 not met (52 of 60; 81%; 81%). F4a not met; F4b met.
- Pulse: carried out in 60 of 60 moments.
- Fourteen pieces, 10:01 to 10:59, about 39 minutes of running against the plan's "about 15 minutes". Results:
  feasibility_results.txt and .json.

Both checks ran in one numerical environment throughout. Every part matches its run log. The cause test's locked
files and both earlier folders are unchanged.

**What follows, by the plan and Amendment 2.** The fresh gap is not the primary outcome. The carried gap is not put
in its place automatically. Neither hold goes ahead as checked. hold_vs_pulse_revised.md sets out the alternatives,
with their measurement-history biases, for Colson's review. It is not locked and not run, and whatever is chosen gets
its own check on old starts before any lock. No new start has been run, and nothing has been published.

**Checked after the results** (from the saved base runs, with no new simulation): all 17 kicks that could not reach
their targets fell where the untouched run's coherence had dipped almost to zero. They fall at 13 distinct instants,
with values of 0.0003 to 0.0019, the lowest 0.4% of the run's values (typical: about 0.023).

## Correction to the 09:00 status below

Four statements in it were wrong. The text is kept as it was written.

- **"The cause is the machine."** Starts 13 to 18 later reproduced their saved runs exactly on this machine; starts 1
  to 12 did not. The departures follow how the saved runs were made: the two batches made in paused pieces depart,
  and the batch made in one go does not. The most likely cause is the pauses, but it is not settled (see "The cause"
  in prelock_amendment.md, corrected at about 10:25). The same claim stood in the description of
  prelock_machine_check.py until 10:34, when it was corrected too.
- **"The saved runs were made on October 10."** All three files were made on October 11, in three batches (table in
  prelock_amendment.md).
- **"The same code reproduced them on October 12, evidently on a different machine."** The October 12 run repeated
  starts 13 to 18 only, and those reproduce today as well, so it never showed a change of machine. The difference in
  speed (about 214 s per start then, 157 s today) does not show one either.
- **"They will almost surely depart in the same way"** (feasibility's base runs). They did not: all six reproduced
  their saved runs exactly.

What it said under "What held" stands: the re-runs agree to the last digit with the unchanged code run on the same
machine, now for all 18 starts.

## Earlier record: status at about 09:00 Halifax (as written then)

**Calibration: stopped after 7 of 18 starts (4 pieces).** V1 as written ("every re-run reproduces its saved run
exactly") fails for starts 1 to 7. By the plan's rule, the calibration gives no reading as it stands. No calibration
result (G, W, overlap, alignment time or κ) has been computed for any start. The only such numbers seen so far are the
dry run's, printed by design for start 1 at t = 200 to 550, outside the calibration's moments.

**The cause is the machine.** The cause test's own base-run code (ct.base_run), unchanged and run in one go today,
departs from the saved runs at exactly the same points as the calibration's re-runs (machine_check.txt). The first
gap difference comes at t = 320 to 2,625. The first coherence difference (at the saved complex64 precision) comes at
t = 1,834 to 3,479. The saved runs were made on October 10, and the same code reproduced them on October 12, evidently
on a different machine: it took about 214 s per start then and 157 s today. Today's machine is an Intel Xeon Platinum
8581C running numpy 2.4.3. Rare last-digit rounding differences grow in this chaotic model.

**What held.** For all 7 starts, the calibration's re-runs agree to the last digit, over all 6,000 units, with that
unchanged code run on the same machine: coherence, every gap log growth, and the settling trace. So the fresh copies
left the base run's arithmetic untouched, and stopping and resuming (4 pieces per start) changed nothing. That is
what V1 was meant to confirm.

**Feasibility: not started.** Its base runs (starts 13 to 18) must also reproduce their saved runs exactly. They will
almost surely depart in the same way, since its moments need states up to about t = 5,500.

**Waiting on Colson.** Either replace V1 in both checks with "agrees to the last digit with the cause test's unchanged
base-run code, run on the same machine", recorded in the plan as a change made after V1 failed and before any
outcome was computed, or keep V1 as written, in which case neither check gives a reading.