Phase 1 / run policy

Direct local inference policy

Final benchmark inference should run locally on the M3 Max, using direct server or CLI execution as needed, while Workers preserve audit, metadata, scoring, and publication flow.

Decision

local
Use direct server or CLI execution for final inference.

Public route

smoke
Keep Worker/Tunnel prediction route for demos and health checks.

Evidence

31.99 GB
Median peak RSS across telemetry-enabled SimpleFold-3B mid CLI repeats.

Route Taxonomy

RouteUseFinal Phase 1 evidence?
Public Worker/Tunnel predictionHealth checks, public smoke tests, demosNo, unless marked as route test
Direct local HTTP predictionShort/medium jobs where no timeout risk is observedYes, with route noted
Direct SimpleFold CLILonger SimpleFold-3B jobs and guarded smokesYes, with progress and RSS telemetry recorded
Async queue / pollingFuture public long-job orchestrationYes, after validation

Required Environment

DIRECT_M3MAX_URL=http://127.0.0.1:8001
DIRECT_PREDICTION_MODE=simplefold_cli
SIMPLEFOLD_CLI_BIN=../m3max-server/.venv/bin/simplefold
SIMPLEFOLD_CLI_TIMEOUT_MS=7200000
DIRECT_M3MAX_TIMEOUT_MS=3600000
RESULTS_DIR=./results-phase1

Repeat/cache-state rule

Final benchmark cards must preserve route, seed, sample count, checkpoint files, auxiliary cache files, progress samples, and peak RSS where available. When repeated runs exist, manuscript tables should report median wall-clock and median peak RSS; one-off pLDDT and timing values must be labeled as single-run estimates.

The initial SimpleFold-3B control on 13CZ:A produced median wall-clock 378.916 s, median peak RSS 31.99 GB, median sequence-aligned lDDT 0.8613, and median TM-score 0.9224. Full details live in the repeat/cache protocol.

Paper Wording

Inference was executed locally on the M3 Max; Cloudflare Workers coordinated metadata, scoring, audit, and publication.