# Learned route-latch placement and transfer V1 — frozen exploratory protocol

The learned scope is one scalar latch site per address on a fixed locally grown rooted tree. It does not learn the task algorithm, primitive library, topology, full growth program or within-episode memory. Search occurs across environments; selected coefficients freeze before test construction. No task targets, optimal labels or future events enter construction. All numerical work uses one local standard-library worker. Positive signed last-write/select semantics remain settled and bounded.

## Physics and additive objective

External links each take one tick; internal bonds take ceil(Euclidean distance), either one or two ticks. Routers/latches process atomically on delivery with no additional operator delay. Let D_a be write-attachment depth, d_a latch depth and R response-attachment depth on the chosen rooted tree. The eligible latch lies on the ancestor path, so write distance is D_a-d_a. A write commits after 1+D_a-d_a ticks; the query request/value round trip and response routing require 2+2d_a+R ticks. Evaluator commit logging is an observation, never a client acknowledgment or execution resource. Thus J = sum_a [w_a(1+D_a-d_a)+q_a(2+2d_a+R)]. Idle safety gaps never enter J. Exact J differs from realized sampled-event mean latency. Initialization and fixed diagnostics are excluded from stationary J and separately costed.

The common T_cap=3*215*2+2=1292 bounds any legal query's two latch paths and response path, including terminal links. Legal write completion is at most 431. Gap is 1294 plus independent integer jitter 0..100. No achieved outcome tunes the bound. Episode cap 110000 accommodates initialization plus 64 subsequent events at K<=8. IDs stay 0..63 and recycle only after queue/pending settlement and output drain. Output and queue capacities remain finite.

For each fresh host, independent public per-address write/read importance values are drawn uniformly from {1,2,4}; profiles share those draws. Each bank normalizes separately; rho=.25/.75 multiplies the read bank, 1-rho the write bank. Rational integer weights sum exactly to a positive common denominator. Writes carry independently drawn signs. No forced query minimum exists. Ordinary streams are independent episodes shared by all conditions on a given host/profile. No task events are supplied to builders.

## Local program and constraints

The unchanged growth program receives only a cell state, local boundary cues, adjacent handles/delays and a delivered message. Charged birth attempts and registrations discover the rooted tree. The scheduler holds host adjacency only for delivery, occupancy and rejection. Global quiescence and final serialization are inherited conveniences. They do not choose a placement, solve assignment or repair choices.

The placement shell retains that tree and its write/response path union. Query boundary exposes the offered q list; each write boundary exposes only its own w. For address 0..K-1 in fixed order, a query-to-write demand token accumulates path delay. At the write attachment a bid begins; moving upstream accumulates write distance and visits feasible vacant sites. Each local scorer sees x/32, d/32, 8w and 8q. Four tanh units each have four input weights and a bias; four output weights and one own-write preference give 25 binary64 parameters bounded [-4,4]. No precomputed objective product, oracle score, label or test normalization is a feature. The running reduction retains just (minimum bid, opaque winner handle). Ties use lexicographic opaque handles. The root sees one reduction, not the host or a list of bids. A downstream reservation marks the winner and installs Q and W routes; a charged acknowledgment returns before the next address. No retries or fallback allocation occur. Roles and address-bearing storage therefore come from executable local actions. Distinct sites are enforced by local reservations. Root, response and other write attachments never bid; own attachment remains available.

Each placement hop charges a padded 96-byte logical message (double fields, handles, tags, integer delays and running bid); the inherited growth hop charges 16 bytes. These exclude Python object overhead, covered by sampled RSS. Temporary state/counters, proposal counts and route tables are reported. Growth is discovered once per split host and cached as actual local states. Every policy placement reruns the shell on copied local states. Per-deployment simulated costs include growth; actual computation charges growth once. No learned placement or candidate reward caching is used. Validation padding alone may reuse a previously evaluated theta0 slot, marked cached.

D_initial uses zero readout and an explicit -.25 own-write prior; it retains the competent prior allocation. D_analytic uses w*x+2q*d, the variable immediate cost, in the same shell and order. B_original preserves exact previous source and its independent weighted Dijkstra tree/root-near greedy allocation. B_cost uses the same shortest-path construction/ties but exact minimum-cost distinct-site allocation for the new objective. B receives workload globally. No B tree is forced to match D.

The D-tree bound is an evaluator-only exact rectangular Hungarian assignment over integer costs. It includes the actual feasible D assignments and every terminal/timing constant. It is cross-checked against exhaustive tiny allocations and matrices. It is never supplied during training/validation; aggregate feedback only measures the learner's own circuit.

Other deployment limits are nonbinding for every assignment in this diagnostic space: the fixed write/response path union has <=216 cells and <=215 edges; each of K<=8 W/Q routes has at most 215 entries, plus V/A each <=215, so <=3870 entries even under a loose bound. Each route key/address/neighbor is finite; the 1 MiB JSON/field allowance dominates their representation. In fact one rooted path per key per cell yields <=18 entries/cell. Roles require exactly K distinct latches. All root/write/response attachments stay active and at least one router exists in this lattice family (more boundary/path cells than K+1); actual validator checks this. Transactions do not overlap; request, value and response are sequential, at most 647 messages and one queue record at a time, below 1024/512/16 caps. Four-byte interface and bounded output record drain after every event. Latency and episode limits follow from the independent path/gap bound. Construction caps govern the local shell, not the deployed assignment optimum: no optimal constructor claim is made. Bound is only over these feasible assignments on this fixed tree and runtime, never arbitrary graphs/trees.

## Splits, optimizer and seals

config.json specifies every namespace, seed purpose, mutation scale, bound, order, objective and selection rule. Four independent seeded searches each use eight training hosts and four validation hosts, all L4/K4 balanced across p=0/.15; both rho profiles. Runs/splits are disjoint namespaces. Host settings within each test group are newly generated independently through setting-bearing group keys; profiles share geometry and importance. Search randomness and training environments vary; algorithms/task family remain shared.

Exactly 128 evaluated vectors/run include theta0. Parent is the lowest training mean J_star, earlier-index ties; all 25 coefficients receive Gaussian mutation with scale .25 for indexes1..32, .15 for33..64, .08 for65..96 and .04 for97..127, clipped to [-4,4]. Only strictly smaller training score changes the incumbent. Repeats still consume evaluations and resources. Every vector, parent, selection/rejection, failure and context score is saved. Short all-address write/read/overwrite diagnostics gate each training circuit. The exact path expression scores only the locally constructed circuit and is checked against each runtime transaction. Three best distinct training vectors plus theta0 undergo validation; duplicate theta0 is omitted, and unused slots pad with cached theta0. Lowest validation score then earlier candidate index selects one policy per search. Validation feeds no mutation or manual edits.

First seal binds source, history, fixtures, config, precision, splits/optimizer/resources before optimization. Second seal binds all selected numerical policies, initial controls, lineage, training/validation files and unchanged source hashes before any test outcomes are generated. No test manifests/costs or D-tree oracle are inspected before that seal. Post-freeze semantic/locality/objective/oracle/replay defects stop the scientific comparison, retain artifacts, and receive no silent repair. Poor searches/invalid individual models are outcomes; no second campaign occurs.

## Test and endpoints

Twelve fresh groups cross L4/6, K4/8, p0/.15 and both rho profiles: 96 hosts, 192 contexts. For every context evaluate four intact policies, four policies built with the opposite rho cues, D_initial, D_analytic, B_original and B_cost: 2304 conditions. Importance, geometry, true events and scoring are unchanged under mismatch. Reference executions occur once per context and are shared, not four independent replications. D-tree assignment diagnostic is separate. Six settings with L6 or K8 form primary size transfer; two L4/K4 settings are fresh in-distribution references. L6/K8 were historically studied, but absent from this new optimizer.

Every final condition gates on two initialization+64-event ordinary episodes, fixed all-address/overwrite/reset checks, two opposite-answer cloned pairs, and compiled JSON replay of both ordinary episodes from generated structure/initial execution state alone. Replay sees no development/training history. Full delivery/state and construction traces are retained for group0 reference rho.25 and group0 L6/K8/p.15 both profiles. Compact event observations, boundary/trajectory/final hashes and complete host/workload/event manifests are retained elsewhere. The latter are auditable records, not independent reproduction.

Valid J_star=J; invalid/nonfunctional/per-model-overbudget J_star=1292. Penalties are never measured latency. Raw J and failures remain separate. Aggregate interruption leaves unexecuted cases unobserved. Zero initial/comparator denominator is a semantic defect and stops reaggregation. Each run's primary paired fraction is (initial J_star - learned J_star)/initial J_star, equally averaged over six transfer settings, two profiles and twelve groups; raw ticks accompany it. Repeat comparisons against analytic, original B and cost B, plus D-tree gap and mismatch-intact true-cost increase. Report four search means/full spread, per-group and per-setting/workload effects, no best-of-four selection and no significance/powered-equivalence claims. Thousands of events are not independent policy discoveries. Construction/traffic/state/work/CPU/bytes remain separate secondary costs. No superiority over untested optimizers or program compression claim follows.

## Resources and closeout

Inherited ledger 444.820223 CPU/410.215338 wall seconds includes old attempts, failed-monitor reservation and setup/audit allowances, verified from saved ledgers. Carry it against 3600 CPU/5400 executed-job wall totals. New 120-second setup/analysis/publication-bookkeeping allowance and every job/failed attempt are recorded. The sole full campaign has an 1800-second CPU/wall limit; other jobs 120 seconds. An inherited 100ms process-tree RSS/wall/artifact watchdog and CPU/file-size rlimits enforce finite jobs. Sampling cannot guarantee instantaneous 1 GiB RSS; unmeasured remote publication CPU is disclosed. Retained new study plus release must be <128 MiB. No infrastructure or monitoring redesign, extra agent, scheduler, GPU, paid service or measured-model inference.

Independent saved-data reaggregation imports no scorer/runtime and recomputes oracle targets, event costs, gates, weights and endpoints. The author rerun uses the first selected policy, group0 L6/K8/p.15 both profiles, four controls and mismatch in a disposable exact-source directory. This is an author check, not independent reproduction. Public-safe scoped files pass exact-byte editorial review before review sealing. Content, visual inspection, saved-data audit, executable check and publication verification are separate records. No earlier artifacts or failed releases are changed.