Start with crr.md (the positively stated canon, every claim labelled by epistemic layer).
Everything it cites is in this folder and was re-run on 8 Sept 2026; the logs/ folder holds the output of that run.
| file | what it verifies | headline (reproduced) |
|---|---|---|
crr_memory_checks.py |
Fisher-age kernel closed forms, Bose–Einstein form, occasion coarse-graining lemma (sympy + numeric + quadrature) | 51/51 checks pass |
crr_recurrence_test.py |
surplus vs recency vs uniform regeneration kernels on recurrence worlds A/B | A: surplus −0.77 vs recency −1.37; B: recency −0.73 vs surplus −4.06; old-dominates = 0 for Ω ≤ 1.2 |
crr_tech_bench.py |
T1 streaming memory, T2 replay budget, T3 change-point locality, then T1-EXT (adaptive Kalman, ablations, SNR sweep, Kalman identity) and T1-OPT (Ω = 1 variants) | exact-gain CRR at Ω = 1 beats tuned adaptive Kalman at every SNR; K(1) = 1/φ |
t1_ext.py, t1_opt.py |
standalone versions of the T1-EXT / T1-OPT sections (T1-OPT imports from t1_ext) | D ratios 0.83–0.99 (1–17%) |
crr_cl_bench.py |
Split-MNIST online class-incremental continual learning; CRR-CLS vs ER/MIR/EWC++/DER++/CLS-ER/GDumb | CRR-CLS 0.874 / 0.884 (5 seeds); fetches MNIST on first run |
crr_cl_checks.py |
stress tests on the CRR-CLS result: fresh seeds, depth-ratio sweep, Ω on recurring stream, buffer size, adaptivity ablation (python3 crr_cl_checks.py [1-5]) |
margin holds; d = e is the optimum; Ω = 1 again |
crr_bit.py |
Bernoulli manifold: FR length = π, rupture = bit flip; P1 (passive Bayes never ruptures) holds; P2 fails (kept) | Ω = 1/π exact |
crr_actor.py |
actor-learner; scalar cut C_n ≥ C*; bounded-mass κ regeneration | cut + κ ≈ 5 within 0.7% of tuned forgetting |
crr_blackhole.py |
pre-registered P1–P4 on thermal and microstate families | Page time is a CRR rupture (pre-Page coherence 8.6%) |
pip install numpy scipy sympy scikit-learn
python3 crr_memory_checks.py # ~1 min
python3 crr_recurrence_test.py # ~1 min
python3 crr_bit.py && python3 crr_actor.py && python3 crr_blackhole.py
python3 crr_tech_bench.py # ~20 min CPU (all sections)
python3 crr_cl_bench.py smoke # quick check; full table ~1 h CPU; see docstring for modes
All runs are seeded and deterministic; re-running reproduces the logs.
Scripts recovered verbatim from the 7–8 Sept 2026 sessions (including the in-session patches), re-executed here,
and checked against the recorded numbers. Only additions: smoke mode and a portable MNIST cache path in
crr_cl_bench.py, a portable import path in t1_opt.py, and RECORD notes in the docstrings of the
pre-registered scripts stating which predictions held and which failed.