# Independent developmental-computation handoff

**14 October 2026 · Research/specification complete · Experiments unexecuted**

## Recommendation

Implement the conventional modular baseline first. Keep a tiny Python IR only as a shared manifest for operators, ports, state, schedules and costs. A new language ecosystem is not justified. If the question becomes lifelong activity/reward-dependent rewiring, reproduce or extend [LNDP](https://arxiv.org/html/2406.09787v1) before creating another developmental system.

The strongest reason not to proceed is that a size-parameterized builder and ordinary function-preserving growth may already provide every required capability. A shared genome, heterogeneous roles, compact construction and exact copying do not establish a developmental learning advantage. The executed example proves engineering behavior only.

## What is delivered

- `research.md` and the self-contained `research.html`: the same critical assessment, primary-source inspection record, smallest semantic design, and exactly two draft protocols.
- `checks.py`: one bounded, two-cell engineering witness, using only Python's standard library.
- `check_results.json`: 17 passed checks, numeric outputs, source hash, resource measurements and limitations. Two startup failures were corrected before the checks ran and are recorded.

The successful witness used about 14.7 MiB peak resident memory and 0.0402 seconds process CPU. Its memory usage was bounded structurally and checked against the 512 MiB ceiling; the host rejected the requested address-space limit. This is disclosed in the results. No training, search, measured-model inference, GPU work, additional agent, deployment or Chronomata access occurred. No cited result was reproduced.

The Markdown/HTML assessment identity, section anchors, companion links, standalone resource structure and check-source hash were verified. The browser URL policy blocked local-file visual preview, so HTML layout has not been visually verified.

## Smallest next implementation, after separate authorization

1. Implement a pure, ordinary `build(memory_width, routing_width, public_ports, seed)` using the declared affine and gated-accumulator primitives. Return the same manifest fields specified in the report. Count primitive definitions as supplied prior information.
2. Implement the deterministic last-write-and-route event generator and symbolic scoring oracle. Separate calibration feedback from scored episodes and freeze seed namespaces. Verify overwrites, delays, reset behavior and context permutations before learning.
3. Add one conventional growth operation: copy old parameters/state at an external-step boundary and attach a new branch with zero readout. Preserve old normalization and update timing; declare optimizer reset and delay-buffer policy. Test zero-step output/state retention and ensure the new readout can receive gradients.
4. Only if these gates pass, add one contextual rewrite policy targeting the same manifest and one matched typed graph control. Use the same parent and, where possible, the same child topology to isolate the claimed mechanism.

Do not implement a parser, broad grammar library, arbitrary neural-layer catalog, NAS loop, biological simulator, repair system, or training campaign as part of this step. Full training remains a future pilot, not an implicit continuation of the current assignment.

## Resources and unresolved fields

The next engineering implementation needs ordinary Python, a separate local directory, one CPU worker, explicit short timeouts and bounded structures. No cloud, GPU, service, new account, pretrained model, private/live agent state or active-trial dependency is required. A future learning pilot needs separately authorized compute/time/data budgets and an independently isolated learner. If suitable dependencies are unavailable, retain specifications; do not install into or borrow from Chronomata.

The report provides provisional ceilings and matching ledgers. Before a pilot, fill in actual message/byte/compilation-wall limits, learner/optimizer settings and evaluation episode counts. After a pilot, resolve between-parent/generator variance, sample size, power, outer replicate count and finalized meaningful/noninferiority margins before confirmatory freeze. These fields are intentionally unresolved, not implied permission for an open-ended campaign.

## Decision gate

Proceed toward a scientific developmental claim only after a paired, resource-matched benefit survives the ordinary builder/ordinary growth and same-topology controls, with held-out component sizes and genuinely contributing added capacity. Interface validity alone supports an engineering IR. Immediate copying alone supports a transfer operator. If the baseline matches or wins, use it and stop development-specific claims. Reproducible improvements in the report's two discriminators would change this recommendation.

## Future integration boundary with Chronomata

Any later integration is a human-reviewed handoff of static specifications, code or result files, with explicit provenance and a new authorization. Do not insert this assistant as a subject/controller in the existing persistent-memory trial. Do not reuse or inspect its SQLite databases, caches, queues, model sessions, ports, scheduler, registries, sealed tasks or experiment budgets. Do not contact or redirect its research agent, modify its working tree/Git state, or alter prior attempt accounting.

If integration is eventually requested, consume public/static exports and run a separately budgeted comparison with separate storage and processes. Conceptual inclusion under the Chronomata research umbrella does not authorize operational coupling. This assignment stops with these files.