Edition README
RRCT multispace convergence protocol
README.md
1 # RRCT v0.61 — Multi-Space Convergence Protocol
2
3 Status: `PROTOCOL_ALPHA_1_WITH_RUNNABLE_SMOKE`
4
5 This package extends RRCT v0.60 without changing it. It addresses a circularity risk in calling a Jacobian-derived readout a workspace before independent geometry has converged on the same mechanism.
6
7 The central rule is now:
8
9 > No single lens, probe, basis, frame, or visualization is allowed to define the workspace by itself.
10
11 The frozen model is examined through independently constructed candidates:
12
13 1. an active verbal frame derived from the corpus-averaged Jacobian transport
14 2. a future-update predictive space
15 3. a cross-layer persistent space
16 4. an empirical write space
17 5. a downstream structural read space
18 6. a balanced write/read space
19 7. an independently fitted fuzzy formation field
20 8. a relational binding residual that tests structure beyond a bag of words
21
22 Only family-balanced convergence, followed by matched causal interventions and mediation, can support a workspace claim. Derived estimators are not counted as independent votes. In particular, B is constructed from C and O, so C, O, and B form one causal-geometry family rather than three discoveries.
23
24 ## What Subtext contributes
25
26 Subtext demonstrates a practical continuous path for reading a pre-fitted Jacobian lens at every input and generated token. Its hook-plus-cache implementation is useful engineering evidence that temporal readouts can be collected during both reading and generation.
27
28 Subtext does not provide an independent workspace discovery method. It visualizes Jacobian-lens coordinates. RRCT v0.61 adopts the temporal-trace idea but treats those traces as one measured view among several.
29
30 ## Important mathematical correction
31
32 The J-space in the source paper is not the range of a Jacobian matrix and is not generally a fixed linear subspace. Token-indexed J-lens vectors are overcomplete and may span the full residual stream. The paper defines J-space, at sparsity level k, as the union of cones generated by sparse nonnegative combinations of at most k active J-lens vectors.
33
34 The runnable 0.5B smoke computes an `active_verbal_frame` proxy from token directions selected by lens readouts. It does not claim to reproduce the paper's full gradient-pursuit cone decomposition. That full decomposition is required for the primary campaign.
35
36 ## Profiles
37
38 `verify` validates package coverage, hashes, configuration, schemas, and protocol invariants.
39
40 `synthetic` runs planted-positive and independent-null numerical checks of the multi-space estimators.
41
42 `qwen-smoke` additionally analyzes the existing frozen Qwen2.5-0.5B model and locally fitted lens. It produces descriptive geometry only. Scientific interpretation is forbidden.
43
44 ## Qwen smoke outputs
45
46 The Qwen runner stores:
47
48 - `multispace_results.json`
49 - `overlap_matrix.csv`
50 - `subtext_like_readouts.json`
51
52 It reports provisional alignment between two nonverbal estimator families, split-half stability, random-rotation nulls, and a future-row permutation control. Two families are insufficient for the primary convergence gate. It does not establish reportability, causal mediation, fuzzy formation, relational binding, or emergence during training.
53
54 ## Claim language
55
56 Until all required primary gates pass, use:
57
58 > Candidate multi-space geometry was measured in a frozen model.
59
60 Do not use:
61
62 > A J-space emerged.
63
64 or:
65
66 > The model has a genuine workspace.
67
68 ## Hardware target
69
70 The smoke profile is designed for the user's existing Qwen2.5-0.5B FP16 model and six-layer lens on an RTX 3070 Ti Laptop GPU with 8 GB VRAM. Matrix estimation is performed on CPU after bounded GPU trace collection.
71
72 ## Contents
73
74 - `protocol/MULTISPACE_DISCOVERY.md` gives the theory and causal design.
75 - `protocol/CAUSAL_PILOT.md` freezes the strongest feasible 0.5B follow-up experiment.
76 - `protocol/SUBTEXT_INTEGRATION.md` separates reusable instrumentation from scientific evidence.
77 - `schemas/candidate_spaces.json` defines every candidate and its information boundary.
78 - `schemas/success_gates.json` defines the revised claim gates.
79 - `configs/` contains smoke and primary configurations.
80 - `tools/multispace_core.py` implements the geometry estimators.
81 - `tools/run_synthetic_validation.py` validates estimators against planted truth and null data.
82 - `tools/run_qwen_multispace_smoke.py` performs the frozen-model smoke.
83 - `Run_RRCT_v061.ps1` is the Windows entry point.
84
85 ## Boundaries
86
87 This is a functional and mechanistic experiment. It does not test subjective experience, phenomenal consciousness, sentience, valence, welfare, desire, or moral status.