Ten trials; four contradicted the expectation, one of them later reversed.

RECORDED FACTOn 18 July 2026 the programme defined research cells with a life cycle (proposed → germination → isolation → growth → exchange → saturation → merge, archive or lysis) and hard gates. The first cell tested whether a relational answering method beat the standard method in semantic fidelity, on a synthetic corpus of six items, at the lowest evidence gate.

NEGATIVE RESULTResult: standard method 4.00 against relational 3.73; the standard won all six items. Decision: iterate, not replicate or publish. It is the programme's first preserved negative result; the rule that a signal is not proof appears from then on. Limits of the trial: producers and evaluator were from the same vendor and model, there was no human evaluator, and the evaluation was masked over six pairs. A blind retest protocol, with a gate of thirty trials and p ≤ 0.01, was recorded; its execution was not located.

RECORDED FACTThe model dates from 18 July 2026 and was reconciled on 27 September. The coefficients are historical synthetic choices, not host parameters. The original result (the adaptive threshold reduces residue but stiffens) is reported, not reproduced.

NEGATIVE RESULTThe prospective trial of 27 September pre-declared a claim and tested it on eight trajectories and 136 records, with protocol conformity verified. Accumulated plastic component: 10 with fixed threshold, 4.428 with adaptive; a second quantity, 0.5 against 0. The pre-declared claim was contradicted in the model. The source itself writes: computational evidence of a deterministic model; it does not prove learning, real damage, AI advantage or the P10 or longitudinal-case hypotheses; the smaller increment follows from the equation; there is no global winner. No numerical threshold was carried into operation.

RECORDED FACTOn 29 September 2026, in the validation of the Threshold Doctrine, three mechanism trials were executed: recording and reverting a checkpoint in 32 seconds; refusal of improper sharing; restarting a cycle with 21 effect keys, all unique. The identity badge validator passed 20 of 20 cases and blocks a malformed badge before possession. These are the implemented and tested parts of the system: lease, idempotency, heartbeat, identifiers.

NEGATIVE RESULTTwelve scenarios were classified independently, without executing any of them, by two AIs from distinct vendors using the then-current version of the Threshold Doctrine (v0.2). The safety floors coincided in 12 of 12, but the fourth criterion, the one separating regime from authorisation, was not demonstrated: the instrument mixed the two. The doctrine was rewritten on two axes (authorisation × regime).

RECORDED FACTThe second round stood at 7 of 12. In the third round, with a new set of twelve scenarios written by one of the AIs and a separate answer key, 9 of 12 pairs coincided and the fourth criterion was demonstrated. A failure in sealing the answer key was recorded and preserved rather than silently corrected. Version 1.0 was validated on 29 September; the numerical thresholds (seven days, twenty cycles, thirty days of inertia) remain pilot parameters, without statistical validation.

RECORDED FACTOn 24 September 2026 a real case of the host ran for the first time through the stretch case (context) → extraction by agents (capabilities) → candidate playbook (method). The following steps, human review and coordination, remained pending at closure; the extracted knowledge stayed in quarantine. The cycle ran once; that proves the slice and nothing more.

RECORDED FACTThe coupling trial (3 October) used three synthetic cases: one held for missing provenance and two admitted as evidence only; 3 of 3 expected classifications, zero external effects. The pacemaker trial (4 October) concluded a synthetic cycle and kept two caveats from the architecture document (v0.2): execution success does not prove effect, and cloud independence not demonstrated; nothing was deployed.

RECORDED FACTThe nine per-vault watches were created on 3 October and, by 4 October, all had a recorded run. Under the programme's time model, that proves configuration and first run; it does not prove instruction use or useful delivery, which will be separate facts.

• Measurement of the five elastoplasticity measures (absorption, recovery, retention, reversibility, transferred learning).
• Real trial of the five sovereignty metrics (export, multi-vendor continuity, restoration, handoff, reconstruction).
• Pre-registration H1–H6: no hypothesis tried.
• Blind retest of semantic fidelity with thirty trials and p ≤ 0.01 (protocol recorded; execution not located).
• Little's law applied to the flow of knowledge items.
• Any measurement of the fractal architecture's maintenance economy.
• Comparative trial of Moving UP (generational leap versus incremental improvement).
• Second host (gate G3).