# Felisyn decisions 020: at this timing the cat circuit relays its input

Results of [protocol 017](https://www.felisyn.com/research/protocol-017/README.md). The protocol was frozen at 2026-10-23T20:46:02Z:
- README SHA-256 `9cb808a3afc6c3d87be274b35e0831f573eaded39f8cfec6ac79a14945a6b2bd`;
- contract SHA-256 `72a58b2fd3d01277d88bc1642e04aed9af55fa79a47ac9b3bda9d7a901268b10`.

Simulation only: no token, wallet, exchange, key or real order was involved.

## The runs

All three ran one after another on the research server, on the declared 120-step sequence:

| Run | Network | Start to end (UTC) | Exit | Time | Peak memory | Export |
| --- | --- | --- | --- | --- | --- | --- |
| intact-A | published connectivity | 23 Oct 20:47:16 to 24 Oct 00:25:16 | 0 | 3 h 38 min | 5.51 GiB | 43.0 MB, `aff1d687…` |
| intact-B | another random instance (`pynn_seed 6`) | 00:25:16 to 04:37:45 | 0 | 4 h 12 min | 5.49 GiB | 40.8 MB, `0b277de9…` |
| feedforward-only | all 12 cortical projections at zero weight | 04:37:45 to 09:45:23 | 0 | 5 h 8 min | 5.88 GiB | 76.2 MB, `f084e686…` |

- **Exports.** They are too large for this site. Their full SHA-256 are in `receipts/exports-sha256.txt`, and they are kept on the research server and in the project's private workspace.
- **Input checks.** Before the analysis, every export passed protocol 017's frozen input checks (`receipts/*-precheck.json`): the declared command line, timing, lead-in, presented sequence, presentation split, and resolved weights and seeds.
- **Analysis.** Protocol 017's `analyze_decisions.py` then ran unchanged from the frozen package, from 09:47:09 to 09:47:21 UTC (`results/analysis.log`). It wrote the four-stage decision log of every run and replayed each one exactly from an independent recount of the spikes.

## Declared result

**Overall reading: "no carry-over".** In the protocol's words: "At this timing the circuit shows no material linear carry-over of the previous input: any carry-over this test detects is below the declared margin of 0.10. Following the project plan, this controller design stops here."

The primary reading is "no material carry-over": neither intact network instance shows material carry-over.

| Run | Decodes | Matches the rule | c, carry-over (95 % interval) | Category | Gain b | Error SD |
| --- | --- | --- | --- | --- | --- | --- |
| intact-A | yes | 93.3 % | +0.008 (−0.007 to +0.024) | no material carry-over | 1.00 | 2.2° |
| intact-B | yes | 91.7 % | +0.009 (−0.009 to +0.024) | no material carry-over | 0.94 | 2.8° |
| feedforward-only | yes | 95.0 % | −0.035 (−0.051 to −0.021) | small carry-over | 0.79 | 5.7° |

- **What c measures.** c comes from the fit ŝ(t) = a + b·s(t) + c·s(t−1): it is how much the previous input moves the current decision signal.
- **Both intact intervals** include 0 and lie within ±0.025, a quarter of the ±0.10 margin. Each is about 0.03 wide.
- **The decoded signal** follows the input closely in both intact runs: gain 1.00 and 0.94, orientation errors of 2.2° and 2.8° (SD).
- **"Matches the rule"** is the share of buy, sell and hold requests equal to those of the noiseless rule applied to the input.

## The lesion (secondary, declared)

Both runs pass the decoding gate, so the lesion readings apply. Each difference is intact-A minus feedforward-only, on the same bootstrap steps:

| Measure | Difference (95 % interval) | Reading |
| --- | --- | --- |
| c | +0.043 (+0.023 to +0.064) | more positive with intracortical connections |
| abs(c) | −0.028 (−0.046 to −0.004) | smaller with intracortical connections |
| gain b | +0.208 (+0.185 to +0.229) | higher with intracortical connections |
| error SD | −3.5° (−4.0° to −2.8°) | lower with intracortical connections |

In plain words:
- **History weight.** With its intracortical connections the history weight is 0.043 more positive, and the net carry-over is near zero. Without them, decisions lean slightly away from the previous input: a small repulsion, c = −0.035, within the margin. These data cannot say which mechanism produces the lesion's repulsion. The protocol's fixed reading attributes carry-over that survives the lesion most likely to the input stage.
- **Gain and errors.** Without them, the decoded signal is compressed toward the middle (gain 0.79). That compression is why its orientation errors are larger (SD 5.7° against 2.2°): its mean error is +9.1° for strong negative signals and −5.9° for strong positive ones. Around its own fitted line, its scatter is about the same as intact (2.3°, against 2.2° and 2.4°). Its requests still match the rule on 95 % of steps.

Two cautions, both fixed in advance or measured here:
- **Operating point.** The lesion also changes layer 4's operating point. Without its cortical connections, layer 4 fired about seven times as many spikes per decoding window: 14.5, against 2.0 and 1.9 in the intact runs (`results/analysis.log`). The comparison therefore mixes the loss of the cortical connections with a different firing regime. The next protocol is designed to control for this with a gain-matched lesion.
- **Sign.** By the protocol's reading of signs, negative c means repulsion. These data cannot say which mechanism produces it.

## Replication (secondary, declared)

- **Agreement.** The two intact instances made the same request on 93.3 % of steps. Independent errors would agree on 90.0 % to 97.5 %. The reading is "within": they disagree about as independent errors would.
- **Shared residual.** Their regression residuals correlate at 0.44. Their LGN input differed completely: no LGN neuron had identical spike trains in the two runs, because the instances use different seeds. The shared part of their errors therefore comes from what the instances have in common: the stimulus sequence, the decoder and the model's fixed structure. One example is a small decoding bias at particular orientations: both intact runs read the middle signal bin about 2° low.
- **The lesion's LGN.** As designed, intact-A and the feedforward-only run received identical LGN spike trains (100 % of LGN neurons), so the lesion comparison is paired down to the input noise.

## Other secondary measures (reported only)

- **Accuracy against the error null:** within, for all three runs. Their errors cost about as many decisions as errors drawn independently within each signal bin.
- **The extended regression,** with lag 2 and the previous input's size:
  - the lag-2 weight is within ±0.03 in every run;
  - intact-B shows a borderline weight of the previous input's size, g = −0.035 (−0.079 to −0.000);
  - the lesion shows a positive interaction of the current input with the previous input's size, h = +0.057 (+0.007 to +0.105).
  - None of these changes the reading.
- **Residual lag-1 autocorrelation:** within its null, for all three runs.
- **The simulated executor,** never a criterion:
  - final values 99.66 (intact-A), 99.82 (intact-B) and 99.47 (feedforward-only), against 99.54 for the noiseless rule, from a starting cash of 100;
  - rejections 3, 1 and 6, against 2.
  - The market is synthetic, so these numbers have no financial meaning.

## What it means, as declared

At 210 ms gratings separated by 147 ms blanks, the cat circuit passes the current input on almost perfectly. It does not carry the previous input into the next decision by any material amount. Following the project plan, this controller design stops here.

The cortical circuit is not idle, though. Without its connections, the decoded signal is compressed toward the middle and decisions lean slightly away from the previous input. With them, the signal keeps its full scale and the net carry-over is near zero.

The next design was pre-registered before this result existed: protocol 021, with back-to-back 70 ms steps and a gain-matched lesion.
- Its spec's hashes were recorded in the project handoff between 23:22 UTC on 23 October and 00:01 UTC on 24 October.
- Its branching rule sends every reading except "decoding failed" to that design, so no part of it was chosen from these numbers. Any later change must be listed in that protocol.
- Its preparatory runs (an identity check, probes and calibration) began on the research server at 09:46 UTC on 24 October.

## What this cannot show

- **Not the full model, and no biological claim.** These are decisions of a 2 mm reduction of the published model, with a declared encoder and decoder.
- **One market, one timing, one encoder and one decoder.** "No carry-over" means no material linear carry-over at this timing, not that the circuit has no memory.
- **Open loop.** Decisions do not feed back into the market or the stimuli.
- **No financial claim.** The market is synthetic and the outcome is simulated.

## Files

- `README.md`: this summary.
- `results/results.json`: everything the frozen analysis computed. It covers each run's measures and nulls, the lesion and replication pairs, the LGN identity check, the readings, and the provenance of every input.
- `results/decisions-intact-A.json`, `results/decisions-intact-B.json` and `results/decisions-feedforward-only.json`: the four-stage decision logs (model output, decoded request, executor decision and simulated outcome) for every step.
- `results/analysis.log`, `results/analysis-start-utc.txt` and `results/analysis-end-utc.txt`: the analysis receipt.
- `receipts/`, per run: the gzipped run log, its start, end and exit receipts, and the input pre-check. Also `exports-sha256.txt`.
- `figures/decisions.png` and `.svg`: carry-over per run against the margin and the error null.
- `figures/decisions-signals.png` and `.svg`: the decoded signal against the input.
- `figures/draw_decisions.py`: the plotting code, which reads only the files in `results/`.
- `manifest.json` and `decisions-evidence.zip`.

## Attribution

**Model.** Antolík J, Cagnol R, Rózsa T, Monier C, Frégnac Y, Davison AP (2024), *A comprehensive data-driven model of cat primary visual cortex*, [PLOS Computational Biology](https://doi.org/10.1371/journal.pcbi.1012342), published under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). LSV1M v1.0 is under its MIT license, and the Mozaik simulation framework under CeCILL.

**Runs, analysis and figures.** Our own, from protocol 017's frozen code.
