Abstract
CI-011 proposed pricing the landfall by the team board's hostile influence. Reading the code first showed two things. Sight is shared by the whole team, so CI-009's boat-or-teammate split was vacuous. And the board clears itself every second, so it cannot know a gun out of sight. Backlog 35 closes without a part. Measured instead: in 76 of 164 landfall deaths (46%), the team had seen the killer 10 to 60 s earlier, near where the boat died, and lost it. Backlog 36 remembers hostile tanks for the landfall, by a weight whose default Paul sets.
Index terms: memory, partial observability, team knowledge, measurement.
I. Introduction
A. Intent
Backlog 35's read, before its part.
B. The prompt
commit. push. continue backlog
II. Decision Log
| # | Cause | Decision | Not taken, and why |
|---|---|---|---|
| D1 | CI-011 D4: a read before the part | Read board.rs and World::visible_to before measuring | Measuring the board's threat at killers' positions: the code answers it |
| D2 | F (code): shared sight | CI-009's split marked vacuous in LR 32 | |
| D3 | F (code): the board clears each second | Backlog 35 closed | A board with memory: a larger change than the question needs |
| D4 | The question left: known earlier? | Log team sightings by position at 1 Hz; class each landfall killer | |
| D5 | F: 46% seen near there 10 to 60 s before | Backlog 36, landfall_memory, default Paul's | Reusing duck_memory: a 5 s memory for a different decision |
| D6 | The compiler warned of an unused total | A shadowed name fixed; the printed numbers were unaffected |
III. Methods
In brain_census (fleet.rs), at each 1 Hz sample, every hostile tank in a living trial seat's observation is appended to sightings[owner] with the tick and its position, once a tick. At a landfall death with killer k, the sightings of k whose age is 10 to 60 s are taken. If any lies within DIRECT_RANGE + 4 of the boat's position at death, the death is "seen then, within reach"; else if any exists, "seen then, elsewhere"; else "not seen". Null run, 16 seeds in ranges of 6, 5 and 5; naval.map, 900 s, respawns 3, the fair census.
IV. Results
| The killer | Deaths | Share |
|---|---|---|
| seen 10–60 s before, within reach of the death | 76 | 46.3% |
| seen 10–60 s before, elsewhere | 26 | 15.9% |
| not seen 10–60 s before | 62 | 37.8% |
Null 0.00 on every seed. Tests: brain 72 passed and 8 ignored; client 31; net 14; proto 3; sim 50; no warnings.
V. Discussion
The bots forget a hostile tank the second it leaves sight. The duck remembers for 5 s, the board for 1 s. Nearly half the boats that die on landfall die to a gun the team had seen there within the minute. Memory is the cheapest knowledge beyond sight the brain can have, and the first part a death count supports that the earlier landfall parts (landfall_shy, landfall_wait) can read at once.
Next: Paul's default for landfall_memory; then backlog 36.