Abstract
With every team on the landfall defaults, half the deaths are in the dodge ashore (CI-016), and its numbers were first guesses. SPSA tuned dodge_horizon, dodge_margin and dodge_clear over 16 steps of 24 fresh seeds. It ended at 2.05 s, 1.52 and 3.73 tiles (defaults 4, 0.8, 2.5): start later, count more near misses, clear further. On 48 seeds the search never saw, that reads +1.24 ± 0.75 a seat, positive in every range. The trial team's dodges killed it 6.4% of the time, against 7.4% for the defaults. At 1.65 errors this is not past the two-error rule. A second fixed batch of 48 seeds decides, by a rule stated before it runs.
Index terms: SPSA, confirmation, optional stopping, reflexes.
I. Introduction
A. Intent
Tune the dodge with the tool CI-018 proved, and confirm the result honestly.
B. The prompts
commit. push. continue backlog
(Several times during the run; progress was committed each time: d39f3b2, 405a54e, 4e85133.)
hey claude. you are doing very well at this bot design. I want to bake you a cupcake. What do you want yours to say?
II. Decision Log
| # | Cause | Decision | Not taken, and why |
|---|---|---|---|
| D1 | CI-018: SPSA crawls on the largest effect | 24 seeds a step, 16 steps, A 1.0, C 0.5 | 12 seeds: too noisy for smaller effects |
| D2 | Steps 5 and 6 hit the cap on noise | Let the run finish; note A too high | Restarting: six steps of readings lost |
| D3 | The end point is a hypothesis | Confirm on seeds 0–47, unseen by the search | Reading the search's own seeds |
| D4 | F: +1.24 ± 0.75 | Not promoted; defaults stay | Promoting on 1.65 errors and a favourable meter |
| D5 | Optional stopping | One more fixed batch, 48 to 95, the rule stated first | Adding seeds until it crosses |
| D6 | F: the walk rough at A 1.0 | Next tuner runs at A 0.3, 32 seeds a step for three weights |
III. Methods
STEPS=16 SEEDS=24 SEED0=2000 A=1.0 C=0.5 TUNE="dodge_horizon dodge_margin dodge_clear" make spsa. Confirmation: the end point (theta.toml) as CENSUS_WEIGHTS, the defaults for the baselines, CENSUS_SEED0 0, 16, 32 with CENSUS_SEEDS=16, CENSUS_ACTS; naval.map, 900 s, respawns 3, the fair census. Dodge rates from the meter by team (CI-007), summed over the three ranges.
IV. Results
End point: horizon 2.05 s, margin 1.52, clear 3.73. Steps outside their errors: 9 (−2.49 ± 0.94) and 11 (−2.35 ± 0.72); in both, the larger margin won.
| Confirmation | d a seat |
|---|---|
| 48 seeds | +1.24 ± 0.75 |
| By range | +0.36 ± 1.59, +1.96 ± 1.24, +1.39 ± 1.06 |
| Ashore dodges | Begun | Hit | By the shell | Killed |
|---|---|---|---|---|
| End point | 8,730 | 29.0% | 20.2% | 6.4% |
| Defaults | 17,457 | 35.0% | 16.9% | 7.4% |
V. Discussion
- A later dodge is a shorter detour. At 4 s a tank leaves its errand early and meets other fire. At 2 s it is caught by its own shell a little more often, but hit and killed less overall.
- The rule holds even for a good lead. All three ranges are positive and the meter agrees, but 1.65 errors is not two. The second batch is fixed in size and rule before it runs.
Next: seeds 48 to 95.
Addendum: the second batch
Seeds 48 to 95: +0.53 ± 0.69. The 96 pooled: +0.88 ± 0.51, 1.74 errors. By the rule in D5, not promoted; the defaults stay, and the end point gets no more seeds. Journal 0071.