# Conventional growth baseline pilot — 14 October 2026

The conventional builder, scoring oracle, ordinary expansion and analytic learner
work within this calibration scope. Expansion preserves old outputs and carried
state exactly at zero adaptation. This fixed exploratory pilot supplies no useful
added-capacity result on query-sign accuracy: ablating the added branch improves
new-task accuracy in both seeds. No developmental advantage was tested.

Original assessment: https://chronomata.ai/experiments/developmental-computation--assessment-2026-10-14/.
Its original two comparative protocols remain unexecuted. The unchanged original
witness reran with 17 checks passed; that is an engineering rerun, not independent
reproduction of cited papers. Original files and the original result record remain
preserved separately.

Execution status: completed. Publication status is recorded separately by the
publisher receipt and HTTPS verification. Review: exact-byte editorial and automated
screening of selected artifacts. Reproduction status: not independently reproduced.
Freeze was local and pre-measurement, not public preregistration.

## Methods and limitations of the implemented baseline

Read amendment.md and frozen.json before interpreting measurements. Two seeds
(101, 202), three arms (parent continuation, ordinary 8-to-12 growth, same-topology
scratch child), 128-event episodes, old K=C=3 and new K=C=4, 32 historical episodes
and 32 adaptation episodes. Adaptation alternates old/new episodes 50:50. Each task
gets four scored reset episodes at adaptation checkpoints 0, 8, 16 and 32. Scored
streams never supplied updates or code repair; they remain exploratory permanently.
Held-out widths 16/24 and task demands 6/8 were not consumed.

The pure builder emits affine/tanh routing features, gated memory, bilinear readout,
parameters, ports, state/delay fields, schedule, provenance and resource costs.
Public WRITE address gates and tanh(value) memory are substantial supplied prior
information. Routing permutations and symbolic bank answers are absent from the
model and builder. A fixed eight-position schema and active masks are supplied
openly. No parser, general language, developmental rule, biological simulator,
architecture search, 3D simulation or active-trial integration was implemented.

NumPy was already available in an isolated bundled CPU runtime. The model uses
small explicit analytic derivatives, checked by finite differences; no autodiff
engine or shared environment was installed. Plain SGD, learning rate 0.08,
episode-mean half squared query loss and float64 were fixed before pilot scores.
Memory updates depend on public events, not feedback. Optimizer state resets
consistently; plain SGD carries no moments. Scratch repeats the actual old stream,
then receives the same adaptation stream with exactly the child's branch topology.

Requested delay-bin weights are equal over 1–8, 9–32 and 33–64 events. Unavailable
bins fall back to available bins. Actual realized counts are in pilot_results.json;
they are not represented as equal realized exposure. Initialization and eligible
query rules were frozen before measurement.

## Gates, counterexamples and failed attempts

The new engineering gates passed 22 checks. They include hand-calculated overwrite,
unwritten-address, delayed-query, permutation and reset/carry answers; all nine
K,C calibration combinations; inactive positions; independent width arguments;
reset and live-state transfer; delay-buffer rejection; gradient access and learner
loss reduction on calibration-only data. Reset and carried-state maximum output
and old-state errors were 0 at a float64 tolerance of 1e-12. Trained parents also
passed reset/carry retention at 0 error for both pilot seeds. Exact-arithmetic
preservation follows from unchanged old computation plus an isolated zero readout;
these finite-precision observations do not establish a universal growth theorem.

Erasing carried state produces a maximum error of 0.17035025549758948. Mismatched
delay transfer is explicitly rejected. The new readout receives gradients after
a public WRITE fills its memory features; feature weights gain gradients after
the zero readout changes. Maximum finite-difference disagreement was about
4.3e-11. Initially zero memory plus zero readout cannot receive a bilinear readout
gradient until a WRITE. The first gate attempt exposed that test setup mistake;
its failed attempt and pre-pilot amendment remain recorded. No scored feedback
was used to fix it. Calibration loss fell from 0.53959 to 0.40768.

Publisher test/startup failures are retained: the sandbox blocked the first RSS
monitor, the first fixture used a macOS symlink alias, fragment rendering lacked
legacy links, and one assertion expected the server error rather than the earlier
immutable-revision rejection. A subsequent local publication validation caught the missing own-assessment link in the fragment renderer; its correction and targeted regression passed before deployment. These were implementation/test failures, not discarded
scientific results. The fixed pilot itself had no failed or omitted cell and no
second optimizer, seed extension or pilot version was run.

## Per-seed results

Accuracy is the mean of four episode query-sign accuracies. Full checkpoint rows
are in results.csv; per-episode counts, loss, compute and ablations are in
pilot_results.json. Tiny seed counts are not a powered scientific comparison.

| Seed | Arm | Old accuracy before/after adaptation | New accuracy before/after adaptation |
| --- | --- | --- | --- |
| 101 | Continue parent | 72.06% / 72.06% | 66.28% / 66.55% |
| 101 | Ordinary growth | 72.06% / 72.06% | 66.28% / 66.00% |
| 101 | Scratch child | 72.06% / 75.00% | 65.99% / 70.67% |
| 202 | Continue parent | 76.37% / 76.37% | 74.39% / 70.23% |
| 202 | Ordinary growth | 76.37% / 76.37% | 74.39% / 71.45% |
| 202 | Scratch child | 76.37% / 76.37% | 75.62% / 71.22% |

At the predeclared final ablation, removing the added branch left old sign accuracy
unchanged and increased new sign accuracy to 66.83% (seed 101) and 73.63% (seed 202).
Full/ablated new loss was 0.39126/0.40336 and 0.34368/0.35565. Thus the branch
contributes numerically and reduces squared loss, but does not help the selected
query-sign metric. Parameter count or nonzero weights alone would have given the
wrong capacity conclusion. Old-task acquisition plateaus and partial new-task
accuracy do not establish learned exact routing or mastery. The symbolic oracle
checks task generation, not a discovered network algorithm.

## Historical charges and resources

All three arms receive equivalent historical/adaptation supervision. The first two
share actual trained parent checkpoints; parent cost is charged explicitly, not
called free. Scratch repeats old history under the same data envelope. Parent
history cost is 2,967,104 and 2,944,256 declared operation equivalents; scratch
history cost is 4,287,872 and 4,255,232. Final adaptation cost is 2,946,048 and
2,954,112 for continuation, versus 4,257,792 and 4,269,312 for either larger arm.
Consequently data are matched and compute is not; elapsed history and arm times
are reported separately. No lifecycle saving or matched-compute growth advantage
is inferred.

Actual retained pilot/calibration totals: 41,472 event exposures, including 512
engineering calibration events; 37,838,912 conservative learning multiply-add
equivalents. Pilot training supplied 30,060 query labels (the calibration gate's
labels are additional and counted within its 512-event ceiling). Evaluation used
27,648 event exposures and 14,247,936 operation equivalents. The convention charges
forward work conservatively on every event, including WRITEs that bypass routing,
and explicit backward contractions/elementwise work on queries. This exceeds
optimizer-update overhead and is a declared upper ledger, not hardware FLOPs.
The pilot body used about 1.013 seconds wall/1.011 seconds CPU and 32,538,624 bytes
process peak RSS. Attempt/resource totals, failed attempts and uncertainty allowances
are in resources.json and attempts.json. Remote publication checks are separate.

One numerical worker and one thread per numerical library; no GPU, API/model
benchmark, paid service or additional agent. CPU/job limits and wall watchdogs
were enforced; a 100 ms process-tree RSS watchdog killed sampled excess. This
was not an instantaneous hard RSS or address-space limit. Structural model limits,
actual sizes and the retained-artifact ceiling are disclosed. No private active
trial state was inspected to establish a resource reservation; no conflicting
static reservation was supplied.

## Claim-to-evidence map

| Claim | Evidence | Limit |
| --- | --- | --- |
| Original witness reruns unchanged | witness_rerun.json and original source digest | Engineering witness, no cited-paper reproduction |
| Task and oracle obey the declared cases | gate_results.json; task.py; hand examples in gates.py | Calibration domain only |
| Ordinary growth retains outputs/state | gate_results.json; per-seed trained-parent zero-step records | Float64 operator-specific result |
| Added branch has gradient access | gates.py analytic/finite-difference checks | After a public WRITE; no universal optimization guarantee |
| Fixed conventional learner executes | frozen.json; source-freeze.json; pilot_results.json; results.csv | Two exploratory seeds; no developmental advantage |
| Added branch fails sign-accuracy capacity criterion | Final predeclared ablation in pilot_results.json | It helps squared loss; this does not imply universal uselessness |
| Shared publishing survives finite concurrency | publisher_verification.json | Eight producers, bounded retries; no arbitrary-N throughput benchmark |

## Executable reproduction

In a separate disposable directory containing all source/config downloads, use
Python 3 with NumPy (the executed runtime's versions are in pilot_results.json).
Set OMP_NUM_THREADS, OPENBLAS_NUM_THREADS, MKL_NUM_THREADS,
VECLIB_MAXIMUM_THREADS and NUMEXPR_NUM_THREADS to 1. Run `python3 gates.py`, then
`python3 pilot.py`. The pilot refuses an existing pilot_results.json, verifies
its gate/config hashes and writes results in its own directory. Never execute
this in the preserved completed continuation or overwrite its outcome files.
The local run_job.py provides the documented CPU/wall/RSS monitoring where process
sampling is permitted. Downloaded attempt ledgers have normalized command paths;
private local commands and raw logs remain excluded.

## Unexecuted research hypothesis

The short next-step.md note specifies an equal-information direct spatial builder
versus contextual growth rule, with topology, geometry delays, intrinsic cell time
constants and temporal responses distinguished. It proposes one delay permutation
and a nonperiodic task, while preserving an identical-final-system output null check.
It neither executes a new campaign nor declares 3D necessary or frequencies an
explanation of intelligence. The broader hypothesis remains open.