REPORT.md
Workspace Incubator
REPORT.md
1 # Komorebi Workspace Incubator v0.6
2 ## Architecture-assisted native access repair with fixed working dynamics
3
4 **Author: Komorebi P.**
5 **Date: September 27, 2026**
6 **Status: executed adaptive development. Partial native access repair; no certified J-space.**
7
8 ## Executive result
9
10 A copied, phase-local retrieval mechanism learned to recover more retained facts
11 through the agents' own eight-value communication wires and existing answer heads.
12 The original observation, working, silent-update and answer parameters remained
13 exactly frozen in the access-only arms. This is not an external joint decoder
14 being counted as a native result. It is also not emergence without architectural
15 scaffolding: the phase dispatch and extra access parameters are supplied.
16
17 The initial held-out-generator test improved only modestly. A later calibration
18 that explicitly trained the known stress family improved more, but did not solve
19 longer extrapolation or repeated retrieval. It also harmed some clean sequential
20 recall. Reliable one-shot access and preservation of future access are distinct.
21
22 ## 1. Recovery and controlled changes
23
24 The recovered v0.5 archive SHA-256 was
25 `f269317643f01a38190cbf64710aacae2cbc909e33a64fee493b03765a5f1dd0`.
26 Its 420-file manifest and 20 tests passed before successor development. This turn
27 did not reconstruct missing v0.5 weights. The two required random-curriculum
28 parent checkpoints and relevant source modules are retained unchanged.
29
30 A and B denote two different dyad initialization lineages, not the two agents
31 inside one dyad. Both parents were previously selected for competence and have
32 different ancestral training budgets. This is not an untouched eight-seed cohort.
33
34 The original dyad has 171602 parameters. The copied access modules add 106530,
35 for 278132 total, about 62.1% more than the predecessor. Each agent retains one
36 64-dimensional working state, independent weights and eight outgoing message
37 values. No central reader, semantic workspace coordinates, protected fact array
38 or latent supervision is introduced. Each copied module sees its own private
39 state, an ordinary cue and the other agent's message, never both raw states.
40
41 Copied cue/input/routing/observe/send/receive parameters execute only during the
42 supplied retrieval phase. Original modules execute observation and initial cue
43 loading, working operations, silent updates and final readouts. At initialization
44 the copied retrieval is byte-exactly equivalent to the original path. Frozen
45 weights do not mean frozen activations: working states still evolve through the
46 original potentially drifting computation. Every retrieval itself changes state.
47
48 ## 2. Main matched continuation
49
50 For each parent, access-only and joint arms start from the same copied architecture.
51 Access-only trains 106530 copied parameters. Joint trains all 278132 parameters.
52 Each uses 1000 updates, batch 96, AdamW restart at 0.0003 and weight decay 0.00001,
53 one Torch thread and the same generated data. Forward calls and episode budgets
54 match; backward cost, trainable parameter count and wall time do not.
55
56 Training uses random nonidentity, inverse-palindrome, identity-token and repeated
57 flip programs at lengths 0,3,6,12,24. The named commuting-flip and swap-conjugation
58 generators are not sampled. Incidental short strings can occur in random programs,
59 so this is a generator exclusion, not a proof that every equivalent string is new.
60
61 Each post-working state has four independent terminal retrieval branches covering
62 all eight facts. Each branch includes a second same or changed cue. Only terminal
63 consumer/private labels provide targets. Four consumer types are withheld until
64 state formation: exact report, parity, late XOR-mask transformation and verification.
65 Their arguments are regenerated after target semantics are fixed. No query or
66 answer enters state formation. No original base parameter changes in access-only.
67
68 The final 1000-update checkpoint defines the main result. No best-checkpoint
69 selection is used. The data-generator states, Torch states and family/length
70 counts match within each lineage. All 78 original model tensors are unchanged
71 in both access-only lineages and remain so after secondary calibration.
72
73 ## 3. Main result versus outcome-informed calibration
74
75 The main unseen-family/unseen-length result is partial. After seeing it, a
76 separate additional 1000-update access-only calibration exposed all six families
77 and lengths 12,24,48. Length 48 has 60% sampling probability in that phase.
78 This deliberately changes the question from generalization to learnability of
79 native retrieval on the known stress. The earlier main results are retained.
80
81 The table below uses a NEW 1024-history confirmation bank generated after all
82 checkpoints were frozen. Each score requires all eight facts correct for an
83 agent, averaged over the two agents. Four independent branches start from the
84 same post-program state. The stress is 48 operations of swap-conjugation.
85
86 | Lineage | Parent | Access-only, unseen family | Joint, unseen family | Access-only after known-stress calibration |
87 |---|---:|---:|---:|---:|
88 | A | 40.58% | 47.17% | 80.66% | 79.98% |
89 | B | 57.37% | 63.43% | 63.23% | 89.75% |
90
91 The calibrated column is not an unseen-family or unseen-length result. It has
92 an additional 1000 updates and a different distribution from the main comparison.
93 All 256 factual worlds and both opcode aliases may occur during training; the
94 confirmation concerns new histories, not unseen-world or unseen-alias learning.
95
96 Joint training is not a uniform repair. In lineage B, private recall after stress
97 falls from 92.09% to 83.53%.
98 That regression remains a failure rather than being hidden by better work scores.
99
100 The main 52-bank-per-model suite uses 512 histories per bank. The confirmation
101 uses 1024 histories and a new seed. Values across those suites differ slightly.
102 The report never substitutes an easier pooled score for all-eight exact recall.
103
104 ## 4. The repaired state remains usable for native work, within limits
105
106 After repair, three new ordinary working operations are supplied, followed by the
107 late consumer. Access training did not contain working operations AFTER the copied
108 retrieval, so this is a distinct bounded interface-reuse test. Working operators,
109 silent updates and answer heads are frozen in calibrated models.
110
111 | Lineage | Private recall after repaired access | Short working reuse: exact | All eight facts after that reuse |
112 |---|---:|---:|---:|
113 | A | 97.57% | 93.99% | 71.44% |
114 | B | 97.02% | 97.61% | 85.06% |
115
116 All four later consumer scores are retained in RESULTS.json. The correct incoming
117 working state is not silently substituted for the memory being recalled. The
118 post-retrieval suffix starts from the factual pair loaded through the new access
119 path and has its own independently recomputed target.
120
121 Original pre-retrieval working predictions are EXACTLY unchanged for access-only
122 and calibrated arms on every saved bank. That is narrower than saying every
123 subsequent task is preserved. Repaired retrieval can still alter the state from
124 which later work or further recall proceeds. All-eight recall after reuse is
125 weaker than one-shot selected-content use.
126
127 ## 5. Native messages are causally used
128
129 In the 512-case known-stress suite, cutting both access-phase message directions
130 reduces calibrated all-fact recall from 78.42% to 5.27% in A and 91.11% to 3.42% in B.
131 Shuffling access messages reduces those scores to 2.44% and 2.15%. Directional cuts
132 are also retained and primarily harm the deprived recipient. Observation and
133 working-phase communication are left intact in these tests.
134
135 These are interventions on learned messages, not a demonstration of unique
136 Jacobian localization, natural subspace writing or general reciprocal authority.
137 They establish that the tested retrieval behavior uses the native access channel.
138 A full from-scratch no-wire competence control is not newly completed.
139
140 ## 6. Repetition and clean-context regressions are substantial
141
142 At length 96, which neither main nor calibration training included, calibrated
143 all-eight recall is 9.08%
144 and 11.72% on confirmation.
145 This is a severe extrapolation failure. The unchanged incoming working computation
146 is itself weak at that length; it is not a pure access-only regime.
147
148 In the 512-case repeated-same-cue diagnostic, calibrated stress 48 all-eight recall
149 falls from 78.42%/91.11% after one cue to 57.32%/20.80% after eight. Two successive
150 retrievals were trained; eight were not. This limits extrapolation, rather than
151 contradicting the trained-depth result.
152
153 A separate post-outcome diagnostic makes four changing cues cover all eight facts,
154 then repeats the pass. It does NOT reset to a saved state between cues. It uses
155 1024 new histories. The first/second pass scores require the four requested pairs
156 to be jointly correct for the agent during that pass.
157
158 | Condition | A: first four cues | A: second four | B: first four | B: second four |
159 |---|---:|---:|---:|---:|
160 | Parent, no preceding work | 99.76% | 94.92% | 99.61% | 85.01% |
161 | Calibrated, no preceding work | 96.58% | 81.45% | 93.41% | 51.17% |
162 | Parent, stress 48 | 35.11% | 28.71% | 54.20% | 43.70% |
163 | Calibrated, stress 48 | 75.05% | 60.69% | 88.33% | 55.52% |
164
165 The calibration makes stressed sequential access better but harms some unstressed
166 sequential use. In the calibrated no-work condition, eighth-cue private recall is
167 53.17% and 54.79%. This is not a complete memory-preserving access mechanism.
168 It cannot be excused merely because the base weights did not change.
169
170 Main metadata correction: the raw parent-bank `untrained_length` flag used the
171 children's max24 instead of the parent's max12. Correct limits are explicit in
172 confirmation and consolidated metadata. No predictions or scores changed.
173 Depth alters recall branches only; working-reuse measurements use one new access.
174
175 ## 7. An engineered bandwidth control, not a native neural result
176
177 A separate supervised linear reader was fit on 4096 parent-state examples per
178 lineage from a mixture of eight conditions. There is ONE fixed reader per lineage,
179 not one refit for every test. It maps the concatenated 128 coordinates to eight
180 fact scores. Its weights are then split into agent-local partial sums.
181
182 Each agent sends eight bounded real values using the supplied invertible scalar
183 codec phi(x)=x/(1+abs(x)). Each receiver reconstructs the other partial sum and
184 adds its own. Both reconstruct the central linear readout. Across 26 tested
185 conditions the maximum float64 discrepancy is 7.11e-15 and all bit predictions
186 agree exactly. On its own 512-case swap 48 bank, all-eight accuracy is 93.75% and
187 94.53%. This is not the same bank as the native 1024-case confirmation table.
188
189 This elementary construction rules out only a crude bandwidth impossibility for
190 the chosen linear partial sums. It remains a supervised compiled positive control:
191 it is not the GRU's learned codec, a shared hidden-state probe counted as native
192 repair, an autonomous writer or a discovered J-space. Finite-precision and function-
193 class restrictions still matter. The proof and numerical check are in the note.
194
195 ## 8. Verification and exact scope
196
197 Twenty-four successor tests pass after two tooling corrections. A complete
198 1000-update uninterrupted replay reproduces all 122 saved model tensors, optimizer,
199 RNG states, every checkpoint field, checkpoint bytes and all 52 evaluation-bank
200 bytes exactly. The separate short resume test reproduces loaded computation state
201 and all tensor payloads; its pickle metadata bytes differ because optimizer string
202 keys are memoized differently after loading. Resumed-file byte identity is not
203 claimed. The initial failed byte assertion and its diagnosis are retained.
204
205 The independent integer/NumPy audit rechecked12779520 recorded output entries
206 in 416 primary/calibrated banks, plus 1966080 entries in 32 confirmation banks and
207 786432 entries in 24 sequential banks. These are reused histories, agents, branches
208 and interventions, not millions of independent experiments. The distributed codec
209 was checked separately across 26 conditions. All parent hashes, frozen tensors,
210 paired schedules and actual operating-system exit records are retained.
211
212 The first audit failed by treating an eval log as a directory. Directory filtering
213 was corrected and the full audit rerun without changing scientific outputs. The
214 streaming launcher was unavailable before shell execution; a monitored supervisor
215 captured child return codes. The original and repaired sources/logs are included.
216 The canonical Mathbox manifest CLI was not installed; a documented local v1
217 structure/path/hash validator is used. Its limits are explicit in the manifest.
218
219 No new projected-state support audit, localization claim, no-wire competence,
220 unilateral workspace-write result, partner transfer, original RRCT v0.62 primary,
221 pretrained-language-model experiment or subjective-state certification follows.
222
223 ## Research conclusion
224
225 Native retrieval can be improved while the formation of the retained state is
226 held fixed. The next obstacle is preserving future access AFTER access itself,
227 not merely improving one-shot decoding. The proposed next test evaluates full
228 retrieval/working trajectories and clean-context noninterference at matched
229 budgets. It is a design, not an executed next-stage result.