Checks before the lock: Fluidity Replication (plan.md, revision 1), 2026-10-14 17:33 Halifax.
No fresh start was touched. Equality and mechanics only: no outcome of any intervention is read, shown or kept. The pilot's own numbers are used only to check the code.

0. The runner's settings match the plan's numbers (tolerances, starts, moment rule, seeds, size, verdict rule), and the pilot's code and Check 3 define the same slower half: yes
1. The check runs finished (1 piece, 146 s), all under this plan and this code: yes
   a. Pilot start 103's base run, re-run unbroken now (145 s, one piece), against the saved file (which matches its recorded fingerprint), every array (start, z, logs, trace, kept_t, kept_u, kept_f): identical to the last digit: yes
   b. Its moment 1 (T = 2055), run now by the pilot's own code and by the new runner from the saved base run, against the saved file, every array the pilot saved (38): pilot code = saved yes; new runner = saved yes; new runner = pilot code yes; identical to the last digit: yes
2. The added recording.
   a. Non-invasive: the new runner's arrays are identical to the pilot code's (1b), and on old start 1 each arm with the recording matches the same arm without it to the last digit (6 arms: a zero change, a changed pull, two pushes, two shams): yes
   b. Adds up: in every arm of moment 1, at every recorded time, the three parts add up to the group's coherence (largest difference 3.3e-16; at most 1e-12), likewise on old start 1; every recorded array has its planned shape and is finite where built: yes
   c. Correct: at R, the recorded parts against Check 3's own calculation from the states (largest difference 1.1e-16; at most 1e-12), likewise on the stand-in: yes
3. Mechanics on old start 1 (short runs, as in the pilot's checks): a zero change reproduces the untouched run exactly; the pull is exactly as scheduled and back at exactly 1.9 after the change; both pushes built, on target, shams in band, sizes the same, gap kept; the replication's moment task on a stand-in moment gives every array the pilot's does, identically, with the pull restored at R: yes
4. The report code reproduces the pilot's records.
   a. Check 3's calculations, pointed at the pilot's files: check3_fluidity.json's verdicts and numbers exactly (296 numbers; largest difference 0.0e+00): yes
   b. The pilot's analysis function, called the way the runner calls it (its starts and cap set for the call and put back after): pilot_results.json's readings exactly (505 numbers; largest difference 0.0e+00): yes
5. The verdict code gives the planned verdict in each of the plan's examples and each validity condition (16 cases): yes
6. The whole report on made-up numbers (written to /tmp/fluidity-replication-dry only; 93 lines; 9 of 200 moments not like for like; 12 s; meaningless numbers): ran end to end: yes
7. Nothing earlier changed: the pilot's freeze record verifies (its protected files, plan, code and calibration results); the pilot's saved data match their fingerprints (132 files); the cause test's lock holds; the pilot's records (22 files) fingerprinted for the lock: yes
8. No replication output exists: yes

All checks passed: yes
Time: the check runs took 146 s in one piece; these comparisons 16 s (the report code on the pilot's files 3 s).
Run under plan.md SHA-256 50aaebfbb4492f455744be7eb7ce2cced685198fc38b0ed4746eb534749d2d23 and replication.py SHA-256 499f5648a6aeed7282fc71918b16d543b92c6e1d29c6d7f5d5e9c0473f70dee3.