# Contextual dense RL: integrated development readout

Descriptive adjunct specified during G training, before native generation. No additional gates or inferential claims. Two gradient snapshots cannot characterize the entire trajectory. Retention is historical source likelihood, not native capability. Exposure is treatment-dependent, not matched causal strata.

| Policy | Correct / 512 |
|---|---:|
| W | 253 |
| G | 252 |
| C | 241 |
| E | 242 |
| H | 243 |

Frozen development spending gate: **False**.

| Contrast | Net correct | Difference, pp | Paired 95% interval, pp |
|---|---:|---:|---|
| C-W | -12 | -2.34 | [-6.25, 1.37] |
| C-G | -11 | -2.15 | [-5.27, 0.98] |
| C-E | -1 | -0.20 | [-3.71, 3.32] |
| C-H | -2 | -0.39 | [-4.30, 3.52] |

| Arm | Optimizer updates /64 | Donor groups /256 | Receiver draws /2048 | Dense tokens | Clipped batches /64 | Source NLL change |
|---|---:|---:|---:|---:|---:|---:|
| G | 52 | 141 | 219 | 0 | 52 | +0.008256 |
| C | 56 | 140 | 205 | 67340 | 56 | +0.013106 |
| E | 51 | 136 | 177 | 56851 | 51 | +0.025380 |
| H | 63 | 133 | 173 | 0 | 63 | +0.038423 |

| Arm | Snapshot | RL norm | Auxiliary / RL norm | Combined norm |
|---|---:|---:|---:|---:|
| G | 1 | 358.6843 | n/a | 358.6843 |
| G | 33 | 192.9173 | n/a | 192.9173 |
| C | 1 | 358.6843 | 0.007174 | 358.6944 |
| C | 33 | 268.0769 | 0.000000 | 268.0769 |
| E | 1 | 358.6843 | 0.001534 | 358.7152 |
| E | 33 | 354.0817 | 0.002054 | 354.0168 |
| H | 1 | 358.6843 | 0.238319 | 360.1024 |
| H | 33 | 251.6415 | 0.468804 | 332.9741 |

| Arm | Training allocated GPU seconds | Matrix + attention proxy, FLOPs |
|---|---:|---:|
| G | 2502.8 | 2.4837e+15 |
| C | 2592.4 | 2.8516e+15 |
| E | 2583.0 | 2.4747e+15 |
| H | 2575.4 | 2.5663e+15 |

| Policy | Evaluation GPU seconds | Generated tokens |
|---|---:|---:|
| W | 1145.1 | 344907 |
| G | 1007.3 | 290484 |
| C | 1052.0 | 325207 |
| E | 1049.0 | 303018 |
| H | 950.3 | 286461 |

Total training/evaluation allocation including interrupted work: [15636.02, 15990.24] GPU seconds. The interval reflects missing interruption exit timing, not statistical uncertainty. Readiness, probes and shared ancestry remain additional.

Interrupted attempts:

- notes/contextual_interruption_20261012_2238/interruption.json: 136 retained draws, 4 uncheckpointed updates, 53850 retained generated tokens. This work was repeated and remains charged.

Completed jobs only; logical matrix/attention arithmetic proxy excludes norms, embeddings, nonlinearities, transfers, optimizer, grader and serialization. Checkpoint recomputation charged. Not measured hardware FLOPs.

Reused512 development; one seed/start; not compute matched; conditional paired item intervals exclude seed uncertainty; no benchmark or sustained-loop claim

Primary claims follow the unchanged frozen contrasts. C−E/C−H determine contextual/dense specificity; this report does not infer such specificity from the spending gate alone. Shared ancestry and source costs are additional, as recorded in the JSON ledger.
