# Context-to-dynamics transfer V1: transfer with conventional equality

**Question.** Can one frozen, compact contextual rule grow executable heterogeneous spatial circuits that preserve useful temporal computation across independently changed module sizes and separation, and what—if anything—does this achieve beyond an equally informed conventional builder?

**Primary result.** D and B each achieved **100.00% unconditional correct-and-on-time query accuracy** over the seven held-out settings. The equally weighted paired difference was **D−B = 0.00 percentage points**, in each of twelve seed groups. All 576 constructor/condition cases completed. The two-point exploratory screening margin was not exceeded. This is a ceiling-limited comparison of two fixed hand-authored algorithms.

The three conclusions are separate:

- **Construction transfer: supported within this supplied substrate family.** The same local rule constructed functional circuits for every independent memory/routing growth and separation combination. Four logical addresses stayed fixed. The conventional builder transferred equally well.
- **Causal mechanism: generic transport-metric calibration and timing dependence supported.** Replacing the construction metric reduced held-out accuracy to 8.08%; swapping unequal control/memory delays reduced it to 0%. Both effects were identical for D and B. The interventions alter rendezvous timing, not task identities or topology. Fixture retiming restored performance, showing feasibility. This is a property of the programmed coincidence circuit, with no demonstrated development-specific causal benefit.
- **Comparative utility: none demonstrated.** Every paired executable circuit hash was identical, including under interventions. Cells, execution edges, storage, task messages and outputs were equal. D incurred developmental offers; B used global loops. Construction times were close and too short for a meaningful speed claim. Neither accuracy nor a measured resource tradeoff favors development here.

Execution is complete. Publication status is maintained separately by the existing publisher receipt. This is a locally frozen exploratory experiment, not public preregistration, peer review or independent reproduction.

## What was supplied and what was tested

The no-learning task is spatial-last-write-select-v1. Four public addresses accept signed writes and explicit queries, with no unknown routing permutation or optimizer. A query asks for the latest preceding externally issued write. Four initialization writes precede randomized overwrites and queries; 48–58 tick gaps make all transactions nonoverlapping and writes settled. Four 64-event episodes per seed/setting/condition contain at least 32 queries each. One correct response in ticks 1–44 after the query is required; missing, wrong, duplicate or late outputs fail. Targets come exclusively from external history. The independent analytical reference verifies physical settlement and response feasibility from actual positions.

Each address has a fork, m serial latches, r serial control relays, and a control/value interface; one shared input relay broadcasts through real delayed edges. Every latch stores a sign and contributes its local value to a bounded running read sum. The last emits sign(sum). Thus all memory sites participate, while replicated storage does not increase logical task capacity. Relays forward bounded messages; the interface compares public addresses and joins exactly one control/value transaction at a single tick. Geometry sets strictly positive integer edge delays by ceil(Euclidean length / fixed unit speed). A faster path receives an explicit buffered hold computed during construction. Positions are genuinely noncoplanar. This does not test whether three dimensions are necessary.

The supplied guide-bond corridors, forward port offers, material/role cues, public addresses, latch, bounded sum/sign decoding and coincidence primitive are strong hand-authored priors. D locally replicates into adjacent offered sites, differentiates, forms connections and passes cumulative transport metrics. A newborn acts on the next development tick. B globally fills the same public substrate and computes holds with a topological pass. Its global access is permitted and its actual work counted. D does not call B. Neither constructor sees evaluator targets or future events. Construction finishes before execution. No discovery, learned specialization, learning economy, biology, exact fractality, lifelong rewiring or inherited trained state is claimed.

The in-range dimensions are m=r=2,d=1; held-out dimensions are m=r=4,d=5. m and r count active serial sites per address; d is the memory corridor's x offset. Across four addresses, memory/routing occupancy grows independently from 8 to 16 each. For group 0, the memory x extent grows from about 1.10 to 3.11 spatial units and routing from about 1.14 to 3.14; exact three-axis extents for every circuit are in construction.json. Increasing d changes genuine port distances and transport, leaving logical capacity fixed. The supplied corridor topology is a considerable limitation: these are untested dimensions of a programmed family, not arbitrary new spatial organizations.

The completed earlier assessment and conventional learning pilot were read from local artifacts after web-tool retrieval of the two supplied public URLs failed. That pilot is background only; it was not rerun, retuned, treated as a developmental comparator, or used to equate sign accuracy and squared loss. Earlier artifacts and the separate primary experiment were untouched.

## Complete matrix and failure classes

Entries below equally average the twelve seed groups. **D=B in every entry.** Per-seed/setting/condition denominators, scores and validity are in aggregate.csv; all responses and failures are in queries.csv. The first setting is the in-range reference. The other seven carry equal primary weight.

| Setting | Cells / execution edges | Intact | Context mismatch | Timing reassignment |
| --- | ---: | ---: | ---: | ---: |
| m2-r2-d1 | 25 / 28 | 100.00% | 21.36% | 0.00% |
| m2-r2-d5 | 25 / 28 | 100.00% | 0.00% | 0.00% |
| m2-r4-d1 | 33 / 36 | 100.00% | 41.72% | 0.00% |
| m2-r4-d5 | 33 / 36 | 100.00% | 0.00% | 0.00% |
| m4-r2-d1 | 33 / 36 | 100.00% | 0.00% | 0.00% |
| m4-r2-d5 | 33 / 36 | 100.00% | 0.00% | 0.00% |
| m4-r4-d1 | 41 / 44 | 100.00% | 14.83% | 0.00% |
| m4-r4-d5 | 41 / 44 | 100.00% | 0.00% | 0.00% |

There were no disconnected/invalid constructions, omitted cases, model-budget failures, wrong emitted signs, duplicates or late emissions. Across both arms/all settings, intact had 27,856 successes. Context mismatch had 2,712 successes and 25,144 missing answers; timing reassignment had 27,856 missing answers. Missing answers count as deadline failures, but these were lost rendezvous events rather than answers that arrived late. Correctness without a deadline equals the correct-and-on-time score because every emitted response was correct and timely. Intact response latency ranged from 11 to 26 ticks. Larger separation genuinely changed latency even though the generous, prespecified feasible deadline left intact accuracy at ceiling.

The context intervention replaces every length/speed-derived construction metric by one while retaining neighborhood, direction, material and public task identities. Both constructors receive the same reduced view; no alternative coordinates are supplied to them. Counts and some metrics remain inferable. Execution uses unchanged true positions. This tests the supplied calibration cue used for holds; it neither removes every possible spatial cue nor demonstrates that contextual morphology is necessary.

The timing intervention exchanges the first unequal control/memory delay pair per address in edge-ID order, retaining topology, operators, holds, initial state and the delay multiset. Every exchange changed relevant arrival sums; no identity/path-invariant permutations occurred. It is an explicit counterfactual transport law. The strongest counterexample to an unrestricted transfer claim is that **every timing-intervened circuit failed every query**, despite remaining connected with valid ports and physically feasible ideal retiming. The exact-tick coincidence primitive is deliberately brittle. Harm does not establish a development-specific timing advantage.

## Twelve outer groups and descriptive uncertainty

Queries, four episodes, size variants and cloned challenges are nested observations, not independent discovered algorithms. The primary weighting first averages setting fractions within each group, then the twelve groups. The following shows every outer difference; N is the four-episode query denominator per setting/arm/condition. Intervention drops are each arm's intact-minus-intervened held-out score; their paired D−B effect differences are zero in every group.

| Seed group | N | Held-out D | Held-out B | D−B | Context drop, both arms | Timing drop, both arms |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| 0 | 146 | 100% | 100% | 0.00 pp | 90.12 pp | 100 pp |
| 1 | 139 | 100% | 100% | 0.00 pp | 91.37 pp | 100 pp |
| 2 | 147 | 100% | 100% | 0.00 pp | 97.08 pp | 100 pp |
| 3 | 142 | 100% | 100% | 0.00 pp | 88.63 pp | 100 pp |
| 4 | 140 | 100% | 100% | 0.00 pp | 92.96 pp | 100 pp |
| 5 | 150 | 100% | 100% | 0.00 pp | 87.05 pp | 100 pp |
| 6 | 144 | 100% | 100% | 0.00 pp | 96.63 pp | 100 pp |
| 7 | 163 | 100% | 100% | 0.00 pp | 88.26 pp | 100 pp |
| 8 | 146 | 100% | 100% | 0.00 pp | 96.48 pp | 100 pp |
| 9 | 143 | 100% | 100% | 0.00 pp | 89.21 pp | 100 pp |
| 10 | 138 | 100% | 100% | 0.00 pp | 92.96 pp | 100 pp |
| 11 | 143 | 100% | 100% | 0.00 pp | 92.31 pp | 100 pp |

The sample SD and SE of the twelve paired differences are zero; the prespecified descriptive t11 calculation consequently gives [0.00,0.00] pp. This degeneracy reflects identical circuits and observed ceiling performance. It provides no powered equivalence guarantee, broad upper bound on developmental benefit, or uncertainty estimate over other rules/tasks. The two-percentage-point margin is an exploratory screen, not a statistical-power claim. Mean held-out context and timing drops were respectively 91.92 pp and 100 pp for both constructors; paired differences of these effects were 0 pp.

## Query-blind challenge and exact replay

Four selected pairs per group/setting/condition/constructor start from a settled shuffled [1,−1,1,−1] bank. Each pair clones complete state, clock, sequence, outputs and queues, then queries the first positive and first negative addresses at the same time. Settled queues are empty in this challenge; cloning preserves that fact. Targets are opposite, so a deterministic query-blind response cannot get both. These selected dependent observations remain separate from ordinary test accuracy.

Held-out challenge results below are equal for D and B, averaged by setting and seed:

| Condition | Individual correct/on-time | Pairs with both correct/on-time |
| --- | ---: | ---: |
| intact | 100.00% | 100.00% |
| context | 6.99% | 1.49% |
| timing | 0.00% | 0.00% |

Across all eight settings, intact challenge accuracy and both-correct fraction were also 100%; context was 9.38% individual and 2.08% both-correct, timing 0% for both. These are 2,304 dependent pairs/4,608 queries across both arms, not 2,304 outer replications.

All **1,152 D ordinary episodes** passed exact compiled replay after actual JSON serialization of structure, physical positions, parameters/holds, initial latch state, clock, buffers and schedule. Full mutable-state trajectories and final states agreed exactly; integer operations need no numerical tolerance. The full state is compared at injection, active delivery/commit and transaction-end boundaries; omitted idle ticks change only the deterministic clock. Construction history supplied no additional execution benefit to an identical executable system.

An illustrative full trace is in trace.json. In the fully held-out group-0 case, query event 4 for address 2 arrives at tick 212. Its control reaches G2 at tick 225, waits nine ticks in an explicit buffer, and joins the retrieved −1 at tick 234. The single response is −1 with latency 22, inside the fixed window. Failed timing cases and their missing outputs are retained in the same per-query evidence, rather than being omitted in favor of this successful illustration.

## Attempts, resources and supplied information

Three tiny-fixture executions preceded the freeze. Their outcomes are retained separately. An audit added bounds/accounting and then made intermediate latch states contribute to the read sum; no score-informed change used evaluation seeds. The final fixtures detected memory erasure, query blindness, broken transport, reset/clone failures, duplicate/late scoring and replay/timing defects. No fixture or numerical evaluation failed. One later read-only shell startup failed before analysis began; it was rerun with approved filesystem access and conservatively charged one second. No score-dependent repair, added arm, seed extension or second campaign followed the matrix.

The frozen matrix body used 39.253 CPU seconds and 39.319 wall seconds. Including fixture/freeze/author-rerun jobs, monitoring and the one-second startup reservation, the new job ledger charges 47.732 CPU and 43.148 executed-job wall seconds. The separate content audit used 0.426 CPU seconds and is covered by the setup/analysis allowance. Conservatively carrying prior pilot costs and both 60-second allowances gives 248.178 CPU and 214.930 wall seconds against 3,600/5,400-second ceilings. Highest new sampled process-tree RSS was 37,961,728 bytes (about 36.2 MiB).


The ordinary matrix executed 147,456 input events and 83,568 query exposures; exact D replay added 73,728 input events. Challenges added 13,824 events, including 4,608 dependent query observations. Ordinary execution sent 3,455,664 messages, charged 82,935,936 logical payload bytes. Both arms had identical task communication. The largest circuit had 41 cells/44 edges, below the tighter 64-cell/512-edge ceilings. Development completed in five or seven ticks, with 28–44 offers per D circuit (896–1,408 logical bytes); B sent no developmental offers. B's explicit global construction makes two edge visits per edge, while D makes one offer per edge; cue/role and common-validation work are disclosed rather than collapsed into a claimed FLOP saving.

The largest queued buffer was eight messages/192 logical payload bytes; 64 messages/1,536 payload bytes were reserved per model. Complete serialized mutable state maxima ranged from 469 to 1,280 bytes across episodes; Python-state object samples ranged from 4,076 to 11,180 bytes. These include different things from the 8–16 bytes of logical latch payload. Representative compiled artifacts, including physical positions, were 4,718–7,460 bytes. Python/interpreter and trajectory-audit overhead are covered by process RSS. Watchdog sampling is not instantaneous hard RSS enforcement or a measurement of host-wide interference. Logical episodes used 3,329–3,435 ticks, below 4,096.

Across 288 constructions per arm, mean construction wall time was D 0.1598 ms and B 0.1569 ms. Total ordinary execution wall time was D 11.259 s and B 10.859 s; Python scheduling/audit overhead is included and the circuits are identical. Such tiny construction timings and small execution-time differences do not support a resource-efficiency claim.


D's executable develop/local-rule source segments total 1,780 UTF-8 bytes; B's build function is 1,181 bytes, and shared Runtime is 4,959 bytes. The 419-byte RULE list is a prose descriptor, not a separately interpreted genome. These counts exclude common cell/finish/validation functions and the substantial supplied corridor/cue description. Full source bytes and representative environment bytes are recorded in resources.json. Source length is not an information-theoretic entropy bound. No trained weights, optimizer, model inference, third-party dependency, GPU, cloud job, paid service, extra agent or scheduler was used. Python 3.14.8 ran one numerical worker, with numerical-library thread variables set to one. Remote publication CPU is not measured.

## Evidence and reproduction

Frozen manifest SHA-256: `13c86845959d0f0a076155fca988eba28ffdb38f7a37f52a3bec1b0e6d2f7cbe`, sealed at 2026-10-14T20:33:21.337119+00:00, before evaluation began. Protocol/configuration, executable source, three fixture records and development history are hash-bound. The post-run independent reaggregation source is clearly separate from frozen experiment code. verification.json checks all 83,568 unique ordinary rows against external-history targets and independently recalculates scoring/primary weights; it also checks opposite challenge targets, complete pairs, paired circuit hashes and recorded replay agreement.

All public files are scoped, project-authored static artifacts. events.csv stores the shared external streams; queries.csv includes input time/address, target, responses/times, retention age, validity and failure class. challenge.csv stores the selected bank and both queries. summary.json/aggregate.csv retain every seed/setting score. construction.json gives costs/extents and deterministic circuit hashes; representative-circuits.json gives complete in-range and fully held-out D/B artifacts in all conditions. Source/configuration regenerates the others deterministically. attempts.json, resources.json, author-rerun.json and trace.json complete the evidence. No whole repository, raw chat, unrelated log, private path or primary-trial state is distributed.

Download/flatten the listed source/configuration and evidence files into a disposable directory. Use Python 3.14.8 for byte-for-byte freeze checks; no external packages are needed. Keep all frozen files unchanged. For a small paired held-out reproduction with interventions:

```sh
python3 run_job.py smoke
```

For a complete author-style reproduction, first copy frozen-status.json to status.json in that disposable directory, with no existing results/ directory, then run:

```sh
python3 run_job.py full
```

To recompute published scores without model execution, run `python3 verify_results.py`. The author actually reran group 0/m4-r4-d5 intact and both interventions for both constructors in a disposable directory: all six cases, 24 ordinary episodes, challenge scores, D trajectories and circuit hashes matched. The directory was removed automatically. This is an author rerun, not independent reproduction.

Content/source and outcome inspection, full execution, exact replay, independent score reaggregation and the disposable author rerun were performed. A visual/browser inspection of this new report is not yet claimed at release preparation. Publication verification is a separate static delivery check, not scientific replication. The frozen protocol sentence saying no summation refers to the interface join; latch retrieval explicitly uses the bounded sum/sign operation described above and in the frozen source. No frozen bytes were edited to make that wording clearer after evaluation.

The closest single next scientific action would be an independently specified transfer test with less prescriptive substrate cues and a strong context-aware builder. It is not executed or authorized by this result. No further campaign follows this report.