Method: no tuning, a stress test for every claim, and bit-for-bit reproduction
Every result is governed by four rules: no tuning (inherited numbers are principles, never fitting targets, so new_tuned_constants = 0), a stress test designed to break each new claim before it earns a grade, an explicit grading vocabulary ([V] verified / [L] cited / [O] open), and bit-for-bit reproduction — twelve deterministic sims, each pinned by a SHA-256 digest.
A claim earns [V] only by passing a test built to falsify it; a break is recorded and the line restarts with the break applied (the inherited Stress Principle). Honest negatives are kept, never hidden or tuned away. The frozen L0 substrate is read-only — every upper layer reuses it unchanged — and each version’s result is reproduced bit-for-bit from a fixed SEED (verify with python3 repro/check_completeness.py).
The method is what makes the program falsifiable rather than illustrative. Four invariant rules run through all twelve sessions, and each is enforced by the reproduction harness, not by narration.
No tuning
No constant is fit to make a result come out. Numbers inherited from the two source whitepapers are used as principle and provenance, never as targets to hit — the brain’s order parameter R = 0.39 and working-memory ~7 are cited as directions, not reproduced by adjustment. Every session reports new_tuned_constants = 0, and a sweep accompanies each claim so the reader sees the whole curve, not a single favourable point.
The Stress Principle
A new claim earns [V] or [L] only by passing a test specifically designed to break it. When a test breaks a hypothesis, the break is recorded and that line restarts with the break built in — so honest negatives drive the work forward instead of being discarded. Examples carried on the books: L1’s shallow binding depth (which motivated L3), L4’s capacity⊥derivability trade-off (closed in L5), and L6’s prediction-sufficiency limit (resolved in L7).
The grading vocabulary
| grade | meaning | example in this paper |
|---|---|---|
| [V] | verified mechanism, reproduced in-sim under a breaking test | one-shot recall 1.0; gate derived from a raw cue |
| [L] | locked to a cited input / principle, not re-derived here | inherited B1–B5, P1–P3; the brain R, WM ~7 |
| [O] | open, or an honest negative, with the obstacle named | O(1) physical parallelism; exact arithmetic needs a digital hand-off |
Bit-for-bit reproduction
Each version’s simulation is deterministic from a fixed SEED and pinned by a SHA-256 digest over its result file. Re-running the module reproduces the digest exactly; the canonical artifact is this HTML, and every number shown is loaded from that verified result. The audit python3 repro/check_completeness.py checks all files are present, re-runs all twelve sims, and compares digests.
hebbian_field, relax, overlap, global_R, pattern_to_phase, corrupt_phase) unchanged — even the axiom-audit knock-outs are broken substitutes built beside L0, never edits to it. The reproduction anchors: v0.1 e9fdd3bc… through v0.12 7f59ced6….