Edition README
Workspace Incubator
README.md
1 # Komorebi Workspace Incubator v0.6
2
3 Read REPORT.md, RESULTS.json and checks/FAILURES.md first.
4
5 This release is an executed adaptive-development continuation, not a J-space
6 certificate. A phase-local learned retrieval mechanism improves native access
7 while original working dynamics are held fixed. A known-stress calibration
8 improves further but degrades some repeated and clean sequential accesses.
9 The analytic distributed reader is a supervised engineered control, not native
10 GRU evidence. Original no-wire/localization/partner-transfer gaps remain open.
11
12 ## What is included
13
14 - Two unchanged v0.5 parent checkpoints, and the required unchanged runtime.
15 - Four main1000-update continuations and two secondary1000-update calibrations.
16 - One full uninterrupted replay, checkpoints, optimizer/RNG state and training logs.
17 - Main512-case banks,1024-case frozen new-history confirmation, message/depth tests,
18 short native working-reuse tests, and no-reset sequential recall.
19 - The supervised distributed linear-reader control, coefficients and raw features.
20 - Twenty-four successor tests, independent integer scoring and exact replay audits.
21 - Per-command source/input hashes, actual child return codes and retained failures.
22
23 This is a278132-parameter model:171602 original parameters plus106530 copied
24 access-phase parameters. Both agents have independent64-dimensional states and
25 eight-value message wires. There is no full-state central reader in the neural
26 repair. The supplied retrieval-phase dispatch is architectural scaffolding.
27
28 ## Verify the delivered package without retraining
29
30 From this directory, using an environment with the recorded dependencies:
31
32 python verify_package.py
33 python -m pytest -q checks
34 python src/audit_results.py
35 python src/audit_additional.py
36 python checks/validate_computation_manifest.py checks/COMPUTATION_MANIFEST.json --root .
37
38 The audit scripts deterministically regenerate audit JSON. Use a copy of the
39 release to preserve the original delivery manifest. The local manifest validator
40 is explicitly not the full canonical Mathbox executable.
41
42 Recorded environment: Python3.13.5, PyTorch2.10.0+cpu, NumPy2.3.5. These scripts
43 run on CPU and do not silently move a model to a GPU. Exact cross-hardware bytes
44 are not guaranteed. No network access or API key is used. Only load checkpoints
45 whose provenance you trust; the supplied code uses torch.load(weights_only=False)
46 for optimizer and RNG records.
47
48 ## Reproduce the main matched pair
49
50 python src/access_repair.py inputs/v05_A_random.pt --out runs/new_A_access --arm access --seed 330601 --steps 1000 --batch 96 --lr 0.0003
51 python src/access_repair.py inputs/v05_A_random.pt --out runs/new_A_joint --arm joint --seed 330601 --steps 1000 --batch 96 --lr 0.0003
52
53 For lineage B use inputs/v05_B_random.pt and seed330603. Output directories must
54 not already exist. --stop N and --resume support an interrupted run. Resume
55 reproduces computation state and tensor bytes, not necessarily pickle file bytes.
56 One complete uninterrupted1000-update replay matches checkpoint bytes exactly.
57
58 python src/evaluate_access.py runs/new_A_access/checkpoint.pt --out runs/new_A_access_eval --seed 930601 --n 512
59
60 That evaluator's raw parent untrained-length flag uses24, whereas the parent
61 trained maximum is12. See checks/METADATA_CORRECTIONS.md. New main models use24.
62
63 ## Reproduce known-stress calibration separately
64
65 python src/calibrate_access.py runs/A_access/checkpoint.pt --out runs/new_A_calibrated --seed 340601 --steps 1000
66 python src/evaluate_calibrated.py runs/new_A_calibrated/checkpoint.pt --out runs/new_A_calibrated_eval
67
68 For B use the B access checkpoint and seed340603. This phase trains on the formerly
69 held-out family and length48. Do not call its length48 results OOD generalization.
70
71 The exact executed command lists and process receipts are in checks/. Campaign
72 scripts use explicit original output names and must only be run in a clean copy.
73 Limit BLAS threads with OMP_NUM_THREADS=1, OPENBLAS_NUM_THREADS=1, MKL_NUM_THREADS=1.
74 A four-CPU-equivalent quota was used. The full prior ZIP is not duplicated; its
75 hash, manifest, report, required parent weights and source are retained.
76
77 ## Current next target
78
79 notes/STABLE_ACCESS_NEXT_DESIGN.md is unexecuted. It targets preservation of
80 future access through repeated cue changes and subsequent working operations.
81 One-shot recovery, repeated-use stability and localized J-space are different
82 claims. No self-model or subjective-state inference is made here.