# Decision and execution log

Date: 2026-10-09. Machine-readable timestamps are UTC; the workspace is America/New_York. This log records decisions, evidence and concise rationales.

- Created `sandbox/flycodes` as a separate project. Existing sibling experiments and raw data were inspected read-only.
- Interpreted the initial constraint conservatively: preserve both graph topology and recorded count ratios. Only the external linear code readout learns. A separate no-recurrence model is an ablation, never a rewrite of the source graph.
- Inspected the actual Feather schemas and status counts. Cached annotations contain 211,577 records; 165,122 have `status=Traced`. “167K” was not treated as an exact local count. The unfiltered weight file contains 151,856,684 ordered segment-pair rows.
- Imported every recorded positive-count edge with both endpoints in the traced-neuron set. Retained 25,563,197 directed pairs, including 101 self-edges. No weak-edge threshold, subsampling, or rewiring. Saved integer count arrays, source/target index arrays, body-ID mapping, sparse CSR arrays and SHA-256 manifest. Source hashes match the prior workspace's independently recorded cache hashes.
- Chose fixed reservoir computing as a first method compatible with freezing all core weights. Wrote `PROTOCOL.md` before the first scored pilot. The dynamics and training budget are explicitly provisional; no benchmark-tuned settings were used.
- Verified that Docker's daemon was unavailable and that the managed sandbox denied `sandbox-exec`. Implemented a bounded AST interpreter for a declared Python subset, with no generated-code `exec`/`eval`, instead of running arbitrary generated programs on the host.
- The initial judge probe discovered that setting `RLIMIT_AS` on this macOS host raises `ValueError`. Interrupted the first feature-collection attempt (`pilot-20261009T155528-s7`) before training/evaluation. Its files and interrupted status remain. Enabled the address-space bound only on Linux; retained CPU, wall-time, instruction and collection bounds on macOS.
- The initial 28 scientific/judge tests passed after the compatibility fix. The scored pilot began at 15:57:22 UTC and completed at 15:59:34 UTC, run `pilot-20261009T155722-s7`.
- Used seed 7, 200 full-corpus updates and 449 answer-byte/EOS targets per update for each arm. All three best checkpoints were selected at step 50 using validation cross entropy, not held-out test scores. Saved all 18 generated candidates, test diagnostics, learning curves, source snapshots, checkpoints and artifact checksums.
- Measured 0/6 tasks solved for fly, 0/6 for the independently trained no-recurrence control, and 0/6 for the small Transformer. Fly/control outputs are byte-identical on all six tasks. The fly output loops through invalid `def solve` / `return solve` fragments; no manual repair or solution substitution was applied.
- Post-pilot diagnostics on a **training prompt** measured mean absolute recurrent input 0.0001036 against mean absolute byte input 0.2747754. This documents a very weak recurrent contribution under the conservative global gain; no parameters were changed after looking at this diagnostic or the test results.
- Added the optional pinned pretrained LLM adapter using official documentation. No brain weights, local files or test expected outputs are included in its requests. Python HTTPS failed DNS; bounded authenticated curl transport checks also failed host resolution. An unauthenticated curl endpoint check was reachable. No valid generation response was returned, no model accuracy was inferred, and the provider score remains null. Request JSON and transport error are retained. No key is saved in the project.
- Local HTTP serving was denied by the managed environment (`PermissionError: Operation not permitted` while binding 127.0.0.1:5077). Installed Chrome also failed to launch under the sandbox. Browser screenshots and live HTTP verification are therefore unverified; no server remains running. Added a self-contained measured-results page that can be opened directly, with execution controls disabled in saved-view mode.
- Model and graph integrity checks passed before and after training. Additional verification, source-header checks, direct inference checks and final test counts are recorded in `verification.json` and `source-verification.json` alongside this log.

## Next stage, not claimed complete

The next experiment should be specified separately: broaden the training curriculum, establish syntax generation first, examine memory on training/validation probes, consider an inhibition-aware model without changing topology, and compare multiple seeds. Repeatedly adjusting the current model against these six exposed test tasks would convert them into development data. Use new tasks for confirmation. A real pretrained LLM measurement and rendered UI review still require an environment that permits those operations.

## Local hosting follow-up

The user requested a separate local address to avoid conflicts with other applications. Inspected listening TCP ports with `lsof`: port 8765 was unused. Changed the CLI and server defaults to `127.0.0.1:8765`, and updated the launch documentation. Starting `./flycodes.sh serve --port 8765` was still denied by the managed sandbox at socket binding (`PermissionError: Operation not permitted`). This was a permission failure, not a port conflict. No other service was stopped or changed, and Flycodes is not running. The user can start that command from their own terminal.
