Experiment record / developmental-computation
Conventional growth baseline: retained state, limited capacity
Two exploratory seeds and three conventional arms. Ordinary growth retained outputs/state exactly; the added branch lowered squared loss but hurt query-sign accuracy.
Question
Can a conventional builder, reliable task oracle, ordinary state-preserving expansion and functioning learner support a bounded calibration pilot?
Hypothesis
An isolated zero-readout branch preserves old behavior and later receives gradients; useful added capacity requires an actual task-performance contribution. This pilot does not test developmental advantage.
Registration
Exploratory local freeze before pilot outcomes; source/config hashes in source-freeze.json. Not public prior registration. Scored streams remain exploratory permanently.
Methods
Read report.md, amendment.md and frozen.json. Two seed groups, conventional 8-to-12 joint growth, K=C=3 old tasks and K=C=4 new tasks, fixed maximum eight-position schema, 128-event episodes, public-address gated memory and learned affine/tanh routing. Explicit analytic NumPy gradients; SGD learning rate 0.08; float64. No oracle/permutation enters the runtime.
Baselines
Continue trained parent; ordinary isolated-branch growth of the same parent; same-child-topology scratch model trained on the actual parent history then identical adaptation. Historical data matched; computation/elapsed time reported separately.
Evaluation
22 pre-training engineering checks; original unchanged witness rerun 17 passed. Trained-parent reset and carried-state retention tested before adaptation in both seeds, maximum errors 0 at tolerance 1e-12. Old/new scored episodes evaluated at 0,8,16,32 adaptation episodes; added branch ablated once after final adaptation, preserving old path/timing.
Resource ceilings
One numerical CPU worker, library threads one; total CPU ceiling 3600 seconds, executed-job wall 5400 seconds, aggregate RSS 1 GiB, retained new artifacts 128 MiB. At most 3 arms/3 seed groups; this pilot used 3/2. Joint training/calibration event ceiling 262144 and per-arm/seed 32768, learning operation ceiling one billion. Watchdogs and structural limits documented.
Stopping conditions
All task/transfer/gradient gates required before one fixed pilot. No extra optimizer variant, seed addition, task search or second pilot after outcomes. All six scheduled arm/seed cells completed; failed startup/engineering attempts retained.
Executing code
Frozen code SHA-256 values in source-freeze.json, gate_results.json and pilot_results.json. Original witness source SHA-256 0982ffb13fd18bbd5125bf4d0969e0e6aa7244730a57208462f141632ffb722a.
Configuration
NumPy 2.3.5, Python 3.12.14. 32 parent history episodes, 32 scratch history episodes, 32 adaptation episodes per arm; alternating old/new training; four independent scored episodes/task/checkpoint. Full fixed configuration is downloadable.
Complete outcomes
Seed 101 final new accuracy: continuation 66.55%, growth 66.00%, scratch 70.67%; growth ablated 66.83%. Seed 202: continuation 70.23%, growth 71.45%, scratch 71.22%; growth ablated 73.63%. Old sign accuracy unchanged under branch ablation. The branch reduced squared loss but did not improve sign accuracy. Zero-step reset/carry retention passed for both trained parents. Both original assessment comparative protocols remain unexecuted.
Realized resources
Training/calibration 41472 event exposures; conservative learning ledger 37838912 multiply-add equivalents. Evaluation 27648 events and 14247936 equivalents. Pilot body wall 1.013 seconds, CPU 1.011 seconds, peak RSS 32538624 bytes. Entire local job ledger recorded about 80.158 CPU seconds and 51.413 wall seconds, plus a conservative 60-second overhead allowance. Resource details and failed attempts are downloadable; remote publication operations are separate.
Limitations
Two exploratory seeds are not a powered comparison. No held-out widths 16/24, task strata 6/8, contextual developmental rule, 3D necessity, independent cited-paper reproduction or lifecycle saving tested. Supplied memory gates are a strong prior. Labels are matched, compute and elapsed time are not. Sampled RSS watchdog is not an instantaneous hard memory limit.
Reproducibility
In a separate disposable directory download all source/config files, use Python 3 with NumPy and one numerical thread, run python3 gates.py then python3 pilot.py. Do not rerun in the preserved continuation or overwrite results. Original assessment is linked below; full methods, failures, resource convention, claim-to-evidence table and unexecuted next step are in report.md.
Revision record
First public baseline continuation release on 14 October 2026, distinct from the original assessment; no earlier public availability claimed.
Selected outcome table
| Seed group | Arm | Adaptation episodes | Adaptation events | Adaptation mac equivalents | Old accuracy | New accuracy | Old loss | New loss |
|---|---|---|---|---|---|---|---|---|
| 101 | continue-parent | 0 | 0 | 0 | 0.7206305442339966 | 0.6627513244745528 | 0.3212597973409425 | 0.48288219751603334 |
| 101 | continue-parent | 8 | 1024 | 735616 | 0.7641752577319587 | 0.6629076558971687 | 0.32190091502419343 | 0.4115715910554431 |
| 101 | continue-parent | 16 | 2048 | 1474368 | 0.7206305442339966 | 0.6629432104394649 | 0.32434820896244887 | 0.3943224844336577 |
| 101 | continue-parent | 32 | 4096 | 2946048 | 0.7206305442339966 | 0.6655473771061317 | 0.3190680231790499 | 0.39306018237789586 |
| 101 | ordinary-growth | 0 | 0 | 0 | 0.7206305442339966 | 0.6627513244745528 | 0.3212597973409425 | 0.48288219751603334 |
| 101 | ordinary-growth | 8 | 1024 | 1063168 | 0.7641752577319587 | 0.6395990750390829 | 0.32256646340986245 | 0.4005708050358191 |
| 101 | ordinary-growth | 16 | 2048 | 2130816 | 0.7206305442339966 | 0.6656844385096404 | 0.32514502304425996 | 0.39027011809709156 |
| 101 | ordinary-growth | 32 | 4096 | 4257792 | 0.7206305442339966 | 0.6600182386659719 | 0.31761590436302156 | 0.3912628036514033 |
| 101 | scratch-child | 0 | 0 | 0 | 0.7206305442339966 | 0.6599085895431648 | 0.3154942560323846 | 0.414121349084228 |
| 101 | scratch-child | 8 | 1024 | 1063168 | 0.7499700311675858 | 0.6654103157026229 | 0.3146536362984247 | 0.38429497785400035 |
| 101 | scratch-child | 16 | 2048 | 2130816 | 0.7499700311675858 | 0.6807500651380928 | 0.3182230625324002 | 0.3729375944461267 |
| 101 | scratch-child | 32 | 4096 | 4257792 | 0.7499700311675858 | 0.7067176372242487 | 0.31091285326926166 | 0.36906256245860697 |
| 202 | continue-parent | 0 | 0 | 0 | 0.7636937804463577 | 0.7439274322169059 | 0.32347392313149975 | 0.3800200516108866 |
| 202 | continue-parent | 8 | 1024 | 749056 | 0.7636937804463577 | 0.7439274322169059 | 0.31987088244574885 | 0.36630432441522576 |
| 202 | continue-parent | 16 | 2048 | 1483776 | 0.7636937804463577 | 0.7414274322169059 | 0.31916418457137796 | 0.36113451988740364 |
| 202 | continue-parent | 32 | 4096 | 2954112 | 0.7636937804463577 | 0.7022607655502393 | 0.3209481262954229 | 0.35261292115825665 |
| 202 | ordinary-growth | 0 | 0 | 0 | 0.7636937804463577 | 0.7439274322169059 | 0.32347392313149975 | 0.3800200516108866 |
| 202 | ordinary-growth | 8 | 1024 | 1082368 | 0.7636937804463577 | 0.7440590111642743 | 0.3190804249560834 | 0.3581063784236759 |
| 202 | ordinary-growth | 16 | 2048 | 2144256 | 0.7636937804463577 | 0.7414274322169059 | 0.31693094191156657 | 0.35035243651780246 |
| 202 | ordinary-growth | 32 | 4096 | 4269312 | 0.7636937804463577 | 0.7145029239766082 | 0.31456541378742453 | 0.3436797283103683 |
| 202 | scratch-child | 0 | 0 | 0 | 0.7636937804463577 | 0.7561695906432748 | 0.2945742264104415 | 0.36441247289019296 |
| 202 | scratch-child | 8 | 1024 | 1082368 | 0.7636937804463577 | 0.7434117490696438 | 0.2938118639575641 | 0.3546936435847044 |
| 202 | scratch-child | 16 | 2048 | 2144256 | 0.7448554758937675 | 0.7539779372674109 | 0.2923553606934229 | 0.34857319970465095 |
| 202 | scratch-child | 32 | 4096 | 4269312 | 0.7636937804463577 | 0.7121637426900584 | 0.2962433379370772 | 0.34624542787501755 |
Recorded method references
Reviewed evidence & downloads
- report.md — derived. Selected project-authored source, configuration or measurements authorized for this report; no private experiment state or external dataset included.
- amendment.md — original. Selected project-authored source, configuration or measurements authorized for this report; no private experiment state or external dataset included.
- frozen.json — original. Selected project-authored source, configuration or measurements authorized for this report; no private experiment state or external dataset included.
- source-freeze.json — original. Selected project-authored source, configuration or measurements authorized for this report; no private experiment state or external dataset included.
- builder.py — original. Selected project-authored source, configuration or measurements authorized for this report; no private experiment state or external dataset included.
- task.py — original. Selected project-authored source, configuration or measurements authorized for this report; no private experiment state or external dataset included.
- learner.py — original. Selected project-authored source, configuration or measurements authorized for this report; no private experiment state or external dataset included.
- gates.py — original. Selected project-authored source, configuration or measurements authorized for this report; no private experiment state or external dataset included.
- pilot.py — original. Selected project-authored source, configuration or measurements authorized for this report; no private experiment state or external dataset included.
- run_job.py — original. Selected project-authored source, configuration or measurements authorized for this report; no private experiment state or external dataset included.
- gate_results.json — original. Selected project-authored source, configuration or measurements authorized for this report; no private experiment state or external dataset included.
- pilot_results.json — original. Selected project-authored source, configuration or measurements authorized for this report; no private experiment state or external dataset included.
- next-step.md — original. Selected project-authored source, configuration or measurements authorized for this report; no private experiment state or external dataset included.
- resources.json — derived. Selected project-authored source, configuration or measurements authorized for this report; no private experiment state or external dataset included.
- publisher_verification.json — derived. Selected project-authored source, configuration or measurements authorized for this report; no private experiment state or external dataset included.
- attempts.json — derived. Selected project-authored source, configuration or measurements authorized for this report; no private experiment state or external dataset included.
- results.csv — derived. Selected project-authored source, configuration or measurements authorized for this report; no private experiment state or external dataset included.
- witness_rerun.json — original. Selected project-authored source, configuration or measurements authorized for this report; no private experiment state or external dataset included.
These are selected public derivatives; raw transcripts, internal configuration and private source are withheld. Hashes describe the actual downloads, not unmodified original evidence.
Artifact manifest and source provenance · Permanent revision URL