RRCT v0.60 Persistent Fuzzy-to-J Workspace Bridge
RRCT 0.60.0.dev1 turns the proposed fuzzy-workspace/J-space connection into a bounded, falsifiable experiment.
The new family is:
experiment.family = "persistent_fuzzy_j_workspace"
It tests a serial causal hypothesis:
learned fuzzy membership
→ moving Jacobian-discovered workspace
→ sparse nonnegative accessible content
→ causal broadcast
→ provenance-conditioned planning
→ later consequence and recurrence
The package preserves the v0.59 reliability control plane, adds stricter resume and sweep-integrity checks, and retains the canonical RRCT/Komorebi release archive.
What is new
- evaluator-blind synthetic controller inputs; hidden event labels and terminal targets are withheld until post-commit scoring/feedback
- task-independent event schedules to prevent event-type leakage
- learned fuzzy gates optimized from terminal consumer loss rather than membership labels
- a public NumPy Jacobian-frame API with exact JSON resume
- moving J-lens estimation, fixed-rank projector discovery, and sparse nonnegative codes
- same-rank PCA and seeded random-subspace controls
- measured membership shuffle, J-space cut, broadcast cut, provenance cut, planner cut, and feedback cut
- longest-supported-prefix reporting, so a late link cannot hide an earlier failed link
- an optional local-only Hugging Face adapter for residual capture, target-token VJPs, norm-matched patching, and projector ablation
- stricter completed-run and sweep-resume digest validation
Package scope
- 308 configuration keys
- 39 major feature switches
- 33 intervention controls
- 9 scientific/engineering families plus one all-family smoke matrix
- 15 presets and 8 sweep templates
- 14 canonical prior release archives
- Windows PowerShell and Bash launchers
- no required network access
This is an engineering and bounded functional-research platform. Its metrics are not probabilities of consciousness, subjective experience, sentience, or moral status.
Fast start
Run the new bounded synthetic bridge:
.\RUN_FUZZY_JSPACE.ps1
Run a shorter diagnostic:
.\RUN_FUZZY_JSPACE.ps1 -Set "jspace.steps=48","jspace.calibration_steps=12","run.name=fuzzy_j_local_trial"
Verify the package, run tests, and execute the all-family smoke matrix:
.\RUN_V060_LOCAL_CHECKS.ps1
Causal protocol
The synthetic bridge separates controller-visible task cues from evaluator-only event truth and terminal targets. A terminal outcome is revealed only after the controller commits its prediction and plan. Evaluation reports every link separately:
- formation: learned fuzzy membership predicts useful access
- geometry: the moving Jacobian frame captures task-relevant sensitivity
- access: sparse content remains decodable and beats matched controls
- broadcast: intervention changes downstream report/inference outputs while preserving upstream code
- planning: provenance and workspace state affect committed actions
- recurrence: later access changes as a consequence of earlier outcomes
The result also reports the longest consecutive prefix that passes its declared diagnostic gates. Those gates are smoke diagnostics unless a freeze and replicated holdout cohort are run.
The bundled single-seed preset presently passes the first three diagnostic links—formation, Jacobian-space discrimination, and selective broadcast—and fails planner necessity and recurrence specificity. This negative boundary is recorded in VALIDATION_REPORT.md; it is not smoothed into a closure score.
See docs/FUZZY_JSPACE_CAUSAL_PROTOCOL.md for endpoints, controls, and falsifying outcomes.
Resumable sweeps
Rank/sparsity frontier:
.\RUN_SWEEP.ps1 -Config .\sweeps\07_jspace_rank_sparsity_frontier.toml
Causal intervention ladder:
.\RUN_SWEEP.ps1 -Config .\sweeps\08_jspace_intervention_ladder.toml
Sweep cell identities now bind the resolved base configuration. A ledger from a different base configuration is rejected instead of silently reused.
Optional local transformer probe
The adapter accepts only an existing local model path. It forces local_files_only=True and trust_remote_code=False; it does not train, download, or call a network API.
.\RUN_LOCAL_JSPACE_PROBE.ps1 `
-Model C:\models\small-local-causal-lm `
-Prompts .\my_prompts.jsonl `
-Output .\runs\local_jspace_probe.json `
-TargetTokenId 123
PyTorch, Transformers, and compatible local weights are optional and are not bundled. The adapter is a measurement path, not evidence that the synthetic bridge has already been found inside a pretrained transformer.
Reliability commands
.\CHECK_CONTRACT.ps1 -Config .\configs\14_persistent_fuzzy_j_workspace.toml
.\RESUME_RUN.ps1 -RunDirectory .\runs\persistent_fuzzy_j_workspace_seed600101
.\AUDIT_RUN.ps1 -RunDirectory .\runs\persistent_fuzzy_j_workspace_seed600101
.\FREEZE_CONFIG.ps1 -Config .\configs\14_persistent_fuzzy_j_workspace.toml -Output .\validation\fuzzy_j_freeze.json
.\VERIFY_FREEZE.ps1 -Freeze .\validation\fuzzy_j_freeze.json
.\RUN_FROZEN.ps1 -Freeze .\validation\fuzzy_j_freeze.json
Checkpoints bind the stable configuration digest and complete serialized experiment state. A completed run is skipped only after its resolved configuration, summary, run state, and checkpoint digests are all verified.
Scientific correction to v0.59
An audit found that the legacy continuous_open_set_runtime repair path directly copied evaluator-side latent state into its estimates. It remains available as an oracle-assisted engineering scaffold and now emits that limitation explicitly. Its attribution and repair numbers must not be cited as evaluator-blind evidence.
The new persistent_fuzzy_j_workspace family was built specifically to remove that shortcut. Evaluator truth is retained only for post-commit scoring.
Safety and interpretation boundary
This offline build hard-blocks network access, external-tool autonomy, autonomous persistence, self-preservation reward, valence and negative welfare-like states, and unrestricted file writing.
The new experiment can support or falsify a synthetic causal workspace chain. It does not establish a natural transformer J-space, open-world awareness, consciousness, sentience, or moral status.