Lock the trigger set and write rules #91

New Issue

2026-08-26T06:55:42+02:00

SirStone commented

2026-08-26 06:55:42 +02:00

part-of: #88

Question

Fix the concrete save policy using the pinned inputs: metric = GA champion fitness per generation, single-observation replace, no cold-start case (#89); reachable moments per #90 — onRoundEnded is the only reliable periodic trigger, abort paths offer at most a best-effort onDisconnected.

Decide:

Exact trigger list honoring the rule of thumb "save at the first moment the GA can no longer improve": own death, game end, round-end periodic, disconnect handler — which combination?
Which snapshot a save writes given the async evo thread: last champion drained from the channel vs current population best; how much staleness is acceptable?
Confirm load order stays per-opponent → global → init under best-slot semantics (ADR 0001).
Crash-safe write procedure: atomic replace so a kill mid-write cannot corrupt the stored best.
Write-cost sanity check at 1000+ rounds if the round-end trigger is included.

part-of: #88 ## Question Fix the concrete save policy using the pinned inputs: metric = GA champion fitness per generation, single-observation replace, no cold-start case ([#89](https://git.fossellini.top/SirStone/SirRoboGarage/issues/89)); reachable moments per [#90](https://git.fossellini.top/SirStone/SirRoboGarage/issues/90) — `onRoundEnded` is the only reliable periodic trigger, abort paths offer at most a best-effort `onDisconnected`. Decide: - Exact trigger list honoring the rule of thumb "save at the first moment the GA can no longer improve": own death, game end, round-end periodic, disconnect handler — which combination? - Which snapshot a save writes given the async evo thread: last champion drained from the channel vs current population best; how much staleness is acceptable? - Confirm load order stays per-opponent → global → init under best-slot semantics (ADR 0001). - Crash-safe write procedure: atomic replace so a kill mid-write cannot corrupt the stored best. - Write-cost sanity check at 1000+ rounds if the round-end trigger is included.

SirStone added the wayfinder:grilling label 2026-08-26 06:55:42 +02:00

SirStone commented

2026-08-26 07:11:15 +02:00

claimed by orchestrator session (ox-alpha), 2026-08-26

SirStone commented

2026-08-26 07:21:10 +02:00

Resolution

Locked save policy:

Trigger — exactly one save point: onRoundEnded, firing every round, dead or alive. Dropped: onDeath (subsumed seconds later by round end), onGameEnded flush, onDisconnected — one code path saves weights, ever. Abort/kill loss window bounded to ≤ 1 round of generations.
Snapshot: last champion drained from the evo thread's channel (bot-thread-visible state); the live thread is never touched.
Gating: candidate replaces the stored best iff its GA fitness exceeds the stored fitness (metric per #89). Fitness travels inside the weights file as a metadata header line above the 745 floats. Legacy headerless files load normally and rank as never-better (−inf), replaced by the first improved candidate.
Load order: unchanged — per-opponent → global → init (ADR 0001).
Write procedure: temp file + atomic rename in the same directory; a kill mid-write cannot corrupt the stored best.
File location (ruling): all weight files live in the directory of the running bot binary, resolved at runtime — not compile-time source paths.
Cost: disk writes occur only on fitness improvement; self-limiting even at 1000+ rounds.

## Resolution Locked save policy: - **Trigger — exactly one save point**: `onRoundEnded`, firing every round, dead or alive. Dropped: `onDeath` (subsumed seconds later by round end), `onGameEnded` flush, `onDisconnected` — one code path saves weights, ever. Abort/kill loss window bounded to ≤ 1 round of generations. - **Snapshot**: last champion drained from the evo thread's channel (bot-thread-visible state); the live thread is never touched. - **Gating**: candidate replaces the stored best iff its GA fitness exceeds the stored fitness (metric per [#89](https://git.fossellini.top/SirStone/SirRoboGarage/issues/89)). Fitness travels inside the weights file as a metadata header line above the 745 floats. Legacy headerless files load normally and rank as never-better (−inf), replaced by the first improved candidate. - **Load order**: unchanged — per-opponent → global → init (ADR 0001). - **Write procedure**: temp file + atomic rename in the same directory; a kill mid-write cannot corrupt the stored best. - **File location** (ruling): all weight files live in the directory of the running bot binary, resolved at runtime — not compile-time source paths. - **Cost**: disk writes occur only on fitness improvement; self-limiting even at 1000+ rounds.

SirStone closed this issue

2026-08-26 07:21:18 +02:00

SirStone referenced this issue

2026-08-26 07:21:42 +02:00

EvoBot weight checkpointing: never lose the best per-enemy weights #88

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: SirStone/SirRoboGarage#91