Abstract
CI-008 explained its null result by sight: a boat sees a gun only at the gun's own range, too late to choose around it. It proposed pricing the guns the team had seen (backlog 32). This step measured that explanation before building anything. Over 16 seeds, the boat had seen its killer in the 10 s before in 128 of 164 landfall deaths (78%); a teammate alone had seen it once. The explanation is withdrawn, and backlog 32 closes with no part to build. Boats land under guns they can see. Backlog 33 proposes standing off a covered beach, by a weight whose default Paul sets.
Index terms: partial observability, measurement before design, correction.
I. Introduction
A. Intent
To read backlog 32's premise, as CI-008 planned, before building its part.
B. The prompt
yes 0. commit. push. continue backlog
(The same message as CI-008; this is its continuation.)
II. Decision Log
| # | Cause | Decision | Not taken, and why |
|---|---|---|---|
| D1 | CI-008 D6: read first | Sample each trial seat's sight at 1 Hz; class each landfall death's killer | Building the remembered-gun price first |
| D2 | Cost | 1 Hz sampling under CENSUS_ACTS only | Every tick: 20 times the cost for a share bounded either way |
| D3 | F: the boat saw its killer in 78% | CI-008's explanation withdrawn, marked in LR 31; backlog 32 closed | Building 32 anyway: a teammate added a gun once in 164 |
| D4 | F: boats land under guns they see | Backlog 33, standing off, as a weight | Choosing its default: Paul's to give |
III. Methods
In brain_census (fleet.rs), at each 1 Hz sample and for each living trial seat, w.observe(p).hostile_tanks() records (seat, owner) → tick. At a TankDestroyed of a trial seat whose last code was landfall and whose by is set, the killer counts as seen by the boat if (owner, by) is within 10 s, and by a teammate if any other trial seat's is. Null run, CENSUS_SEED0 0, 6, 11 with CENSUS_SEEDS 6, 5, 5; naval.map, 900 s, respawns 3, the fair census.
IV. Results
| Seen by | Deaths | Share |
|---|---|---|
| both | 128 | 78.0% |
| neither | 35 | 21.3% |
| a teammate only | 1 | 0.6% |
| the boat only | 0 | 0 |
Null: 0.00 on every seed. Tests: brain 71 passed and 8 ignored; client 31; net 14; proto 3; sim 50.
V. Discussion
Two explanations in two steps were wrong: LR 29's equal shares (CI-007) and LR 31's sight (this report). Both were reasoned from the code rather than measured, and both were caught by the next step's read. They cost no code, because each read came before the part it would have justified. The working rule is to measure the premise of a backlog item before building it.
Next: Paul's default for landfall_wait; then backlog 33.