SentinelPulse

Save model reasoning for alerts that need it. A tier-0 gate applies deterministic rules to suppress known noise at zero tokens and escalates unresolved alerts to Meta Muse (muse-spark-1.3), which turns each escalation into a triage note with hypotheses and next actions. On a 7-fixture replay, gate off vs on: 0 noise model calls, -13.2% tokens, -19.8% latency, 0/7 incidents missed. Every on-call queue fills with non-incidents: health checks returning 200, favicon 404s, CDN revalidations, blips that self-recovered before anyone looked. Most "AI SRE" stacks send it all to a reasoning model, paying in tokens, latency and attention for nothing. SentinelPulse puts a deterministic gate before the model: - gate-0 applies rules to the normalized event, dropping provable noise at zero tokens. - A cheap classifier categorizes the survivors. - Only events that clear gate-0 reach tier-1 triage on Meta's Muse API (muse-spark-1.3), which drafts the triage note, hypotheses and next actions. - A sandboxed remediator can propose a minimal patch; the authenticated PR path is opt-in and unused in the demo. The gate does not make the model smarter; it keeps noise away from it. Same 7 fixtures replayed, gate off vs gate on, single run: C1 incidents missed 0/7 -> 0/7 C2 noise model calls 4 -> 0 (all four blocked by rules) C3 tokens 5533 -> 4802 (-13.2%) C4 end-to-end latency 36.8s -> 29.5s (-19.8%) C5 signal escalation 3/3 -> 3/3 SHA-256 freezes the evidence (bench/evidence/FREEZE-20261027.json), tying every number to exact bytes. `python3 -m demos.l3.freeze_evidence --check` fails loudly on any drift, and `bash demos/l3/demo_take.sh` runs a 4-step preflight gate. The report is a self-contained, CDN-free HTML file rendering air-gapped. Live mode refuses with exit 4 without --confirm-live, and BudgetGuard reserves calls and tokens against a hard ceiling before spend. Built solo for the Meta AI Developer Hackathon, Indie track. In-window (Oct 26 - Nov 4): the full demos/l3 harness (freeze, bundle, report renderer, live runner, preflight, 90-second video pipeline) and the 2026-10-27 bench run. Disclosed as pre-window: app/ gate prototype (Oct 14-20) and 2026-10-20 bench JSON kept strictly as a regression baseline.