# Context-to-dynamics transfer V1 — frozen exploratory protocol 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? This is one no-learning, hand-authored constructor comparison. It separately tests construction transfer, causal dependence on context/timing, and comparative utility. No algorithm discovery, learned specialization, learning economy, biological fidelity, fractal property, necessity of 3D, lifelong rewiring or trained-state inheritance is claimed. The prior conventional learning pilot is background, not a comparison arm. Its sign accuracy and squared loss remain distinct. No original pilot is rerun. ## Local provenance and execution boundary The web retrieval tool failed on both supplied assessment/baseline URLs. Methods and limitations were read from completed local assessment, handoff, baseline report, amendment, frozen configuration, actual builder/task, final resources and attempts. The new assignment explicitly authorizes this separate directory beyond the old completed milestone's directory restriction. All earlier files remain untouched. No separate primary experiment, Git state, environment, agent or private state is accessed. No extra agent, scheduler, installation, GPU, paid/cloud computation or service is used. Only Python's standard library is required. ## Supplied substrate, rules and primitives There are four logical addresses throughout. The substrate supplies two directed port-marked corridors per address, an input relay P, a fork F and an interface G. Neighborhood consists of the undirected supplied guide bonds, each within 12 spatial units. A forward port offer gives direction, material, public address and a metric cue equal to ceil(Euclidean length / one spatial unit per tick). Positions have three coordinates and nonzero tetrahedral volume. The material/corridor cues encode substantial intended structure; morphology and primitive roles are supplied priors. They are not inferred from task feedback. Both constructors have identical cues. Memory size m is the number of serial active latch sites per address; routing size r is the number of serial active control relays per address. Each latch stores the most recent signed write locally and forwards writes/accumulating read requests; all added latches participate in communication and storage. These repeated latches provide redundancy and distance, not increased logical capacity. d shifts the memory corridor along x; it changes fork-to-memory and memory-to-interface distances. The routing corridor, memory corridor and interface positions are explicit. Source/configuration and actual module cell counts/extents accompany the results. No inert padding exists. D starts with P. The same local rule replicates into adjacent offered vacant sites, differentiates from local material, connects to the parent, and forwards a cumulative metric in bounded developmental offers. A newborn acts on the next development tick. At G, two local arrivals specify the control and memory path sums. Construction ends before execution. Global deterministic scheduling, occupancy/resource checks and validation are disclosed; they do not select task-solving wiring. D does not call B. B fills all supplied sites with the same roles and uses ordinary global loops and a topological path-sum pass to set holds. It has stronger global access to that view. Both use the same validation; realized work, communication and sizes are measured. A message is four bounded integer fields: (kind, transaction ID, public address, bounded signed value/partial sum in [-4,4]). The logical wire payload ledger conservatively charges 24 bytes. Kinds are W write, Q query, C control, R read request, V retrieved value. P broadcasts external events over four physical edges. F sends a matching W into its latch corridor; Q splits into C and R on both corridors. Relays forward. Each latch has state h in {-1,0,+1}, initially 0, and commits h'=v on W, otherwise h'=h. W propagates to the last latch; each R adds the local h to its carried bounded partial sum, and the last latch emits V=sign(sum), with sign(0)=0. Every memory site therefore contributes its stored value to retrieval; the redundant copies provide no extra logical addresses. The supplied sum/sign operations are disclosed prior information. Thus no bank lookup occurs inside the runtime. G matches its address against C/V and joins exactly one of each with the same transaction ID at the same local commit tick, then emits V once. Unmatched arrivals are discarded at that tick. No summation or implicit normalization is used. For path metric Lc,Lm the internal holds are max(Lc,Lm)-Lc and max(Lc,Lm)-Lm. These explicit delays occupy charged finite buffers. The supplied coincidence gate is intentionally timing-sensitive. A more tolerant protocol is not tested. All primitive implementations are prior information, not learned algorithms. Execution edge delay is a strictly positive ceil of true physical length. At each active tick, messages deliver in (tick,destination,edge,sequence) order, latch writes commit, outgoing messages are scheduled, internal holds due now are delivered, and interfaces join/emit. Zero internal hold is local computation; intercell same-tick delivery is forbidden. Empty ticks change only the clock; complete mutable state is recorded at active commits, injections and transaction ends for exact replay. Episode reset reinstantiates the serialized zero states, empty queues, clock and schedule. The four-lane input/output port identities are public semantics. There is no unknown permutation, optimizer, teacher forcing or evaluator access by either runtime. ## Task, reference and scoring Spatial-last-write-select-v1 has four addresses. A write becomes symbolically authoritative when issued externally, and queries request the latest preceding external write for that address. Nonoverlapping transactions with 48–58 tick gaps ensure physical settlement before any subsequent event. Each episode has 64 events: four initialization writes then randomly interspersed writes (probability .4) and queries, with a deterministic quota of at least 32 queries. Signs and addresses are random, repeated overwrites allowed. Retention age is recorded in ticks. The oracle scans external events only. An independent analytical reference computes terminal-write settlement and query control/data path latencies from environment positions, without generated topology, learned parameters or model outputs. It must agree with oracle values and fit the fixed 1–44 tick response window. Analytic bounds from m,r<=4 and d<=5 give at most 6 input-edge ticks, at most 23 memory-path ticks and 12 control-path ticks, below the fixed deadline with ideal retiming. Even the sum of both path totals plus input edge is at most 41 ticks. Interventions preserve all ports and positive transport; fixture retiming restores performance. They do not make the oracle target intrinsically physically impossible. No deadline is chosen from achieved model scores or latency. Exactly one correct response in [input tick+1,input tick+44] is required. Any duplicate response for that transaction also fails, including one outside the window. Missing, wrong, duplicate and late responses fail; correctness without deadline is separate. All responses are observed through the transaction's end (next event minus one; last event plus 47). The analytical maximum transport/hold sum is below that end. No target-dependent best instant or two-sign output can win. A selected shortcut challenge settles a bank obtained by shuffling [1,-1,1,-1]. For each of four pairs, complete state including clock, buffers, sequence and outputs is cloned. The first positive and first negative addresses receive different queries at identical times in the clones. Individual accuracy and both-correct pair fraction are separate outcomes. These intentionally selected dependent observations are not ordinary samples or extra independent replications. ## Fixtures, freeze, matrix and interventions Only (m,r,d)=(1,1,1),(1,2,1),(2,1,2),(2,2,1) are debugging fixtures. They must detect memory erasure, query blindness, broken transport, duplicate/late scoring, reset or clone defects, and compiled replay differences. The delay exchange must alter relevant arrival sums; physical retiming is a fixture-only diagnostic, not a scored third arm. Brief manual changes are recorded in development-history.md. No learning, search or calibration campaign occurs. Before any scored generation, freeze source hashes, this protocol, rule data, primitive/schedule definitions, config, fixtures and deterministic seed table in a timestamped local manifest. This is a locally frozen exploratory experiment, not public preregistration. The publisher has no existing registration endpoint used here. No registry is built. No evaluation-informed repair or replacement is allowed. The complete matrix is m in {2,4}, r in {2,4}, d in {1,5}. (2,2,1) is in range; seven cells are held out, including independent growth/separation changes. Twelve fresh task/layout seed groups each nest all settings, both constructors and all three conditions. Four ordinary 64-event episodes and four paired-query challenges are scheduled per constructor/condition/setting/group. Identical task streams are reused across settings and arms. config.json freezes seed derivation, counts, interleaved order, denominators and ceilings. Layout jitter never consults targets. Context mismatch replaces all construction length/speed metric cues by 1 while retaining the same neighborhood, material, task address and direction cues. The true execution geometry remains unchanged. Path counts remain inferable, and some actual metrics can coincide with the false cue. Absolute coordinates are not a redundant constructor input in this protocol. This tests calibration of the supplied transport metric, not every possible form of context dependence or morphology change. Timing reassignment exchanges the first unequal control/memory edge-delay pair per address in edge-ID order after construction. It preserves topology, operators, parameters/holds, initial state and the delay multiset. Identity/path-invariant permutations are defects. It is a counterfactual transport law, not an alternative physical embedding. Comparable interventions apply to B and D. Exact serialized D replay in every ordinary episode is a semantic control; construction history should supply no execution benefit to an identical final system. ## Outcomes, resources and stopping Primary: intact unconditional D minus B query fraction per setting, equally averaged over seven held-out settings, then over twelve seed groups. Invalid construction or model resource exhaustion scores zero on scheduled queries, without removing them. Aggregate interruption yields an incomplete matrix: unexecuted cells are not zeros. Report all seed/setting/condition results, in-range scores, challenge, timing, correctness, failure classes, actual construction/execution costs and replay. The prespecified screening margin is two percentage points, not a power claim. Describe uncertainty only over the twelve paired seed aggregates (SD, SE and a descriptive t11 interval), not independent queries/episodes/architectures. One local numerical CPU worker; numerical library threads 1. Inherited finite limits are tightened to 64 cells, 512 edges, 64 developmental ticks, 512 developmental offers of at most 32768 logical bytes, 64 messages/8192 payload bytes per event, 64 queued messages globally/eight per cell, one MiB serialized model, two seconds construction, 4096 logical episode ticks and 300 seconds/job. Aggregate ceilings: 3600 CPU seconds, 5400 executed-job wall seconds, one GiB RSS, 128 MiB retained new study/release bytes. Previous recorded pilot CPU/wall plus its 60-second allowance remain charged, with another conservative 60 seconds reserved for setup/analysis bookkeeping. Process CPU and file-size rlimits plus wall, sampled RSS and retained-file watchdogs apply. 100 ms sampling is not instantaneous hard RSS enforcement. Python interpreter/object overheads appear in sampled RSS; payload bytes and serialized state are distinct. Remote publication is separate static-file work, with unmeasured remote CPU. Oracle, causal timing, reset or replay defects stop the attempt. All failures are preserved. No new campaign, deadline adjustment, extra arm or seed extension follows scores. A small author rerun of one representative intact pair and its interventions is required in a disposable directory; it is not independent reproduction. Public release contains only scoped source/configuration/outcomes and a result-first report. Publication uses the existing reviewed publisher, preserving prior versions. A blocked publisher never prevents local scientific closeout. Reproduce in a disposable directory containing these downloaded source/configuration files and freeze.json/fixture-results.json. `python3 run_job.py smoke` runs a small fully held-out paired case and all three conditions. `python3 run_job.py full` runs the single frozen matrix, refusing existing results or non-frozen status. For a full reproduction, copy the distributed frozen-status.json to status.json first. Keep thread variables at 1 (the runner sets them). Never run over the preserved original.