Listening for events…
← All Docs

Sprint — 2026-04-08

Theme: "The Pulse needs to be smarter" Living document. Updated as work progresses.

Sprint Goals

Four ordered objectives. Don't skip ahead unless one is blocked.

# Issue Type Why now
1 Pulse persistence + replay (new) Infrastructure The Pulse drops on refresh — needs ring buffer + replay endpoint so new visitors see recent activity
2 #103 Forbush decrease catalog Build automated detection from NMDB neutron monitors. Pulse will stream them once detected.
3 #45 Earthquake-weather coupling paper Test the controversial barometric-pressure-triggers-quakes hypothesis. We have 1.6M USGS quakes + Open-Meteo pressure.
4 #44 Granger V4 — full causal network The big paper. CO2 + methane + tidal + pressure cross-influences. Updates the network paper.

Status Tracker

#1 — Pulse persistence + replay

  • Status: Not started
  • Notes: Need a small in-memory ring buffer (~100-200 events) on the API side, plus a GET /api/v1/pulse/recent endpoint. PulseTicker.astro fetches it on connect to populate the rail before live events arrive.
  • Acceptance: Refreshing the page shows the last 30-50 events immediately, then continues with live streaming

#103 — Forbush decrease catalog

  • Status: Not started
  • Notes: Algorithm: rolling 30-day baseline median, detect drops >1% sustained for >12 hours, cluster across stations to suppress single-station noise. Cross-reference DONKI CMEs.
  • Acceptance: A forbush_event metric in the database, populated automatically when conditions are met. Pulse picks up new ones via the existing streamer.

#45 — Earthquake-weather coupling paper

  • Status: ✅ ACCEPTED + Dana running
  • Headline: Clean null on the barometric-trigger hypothesis
    • LA subset (the only natural-tectonic test): d=+0.03, p=0.696
    • Houston cluster (17 events) correctly disclosed as Eagle Ford induced seismicity contaminated by a single storm front
    • 1/10 tests survives Bonferroni (Houston, induced)
    • 7-day-before placebo honestly documented as unrunnable on 13-day window
  • Mike R1: 5 critical caught (SEA arithmetic error, fake placebo, induced seismicity not disclosed, Granger cherry-pick, hardcoded DSN)
  • Mike R2: MINOR REVISIONS — caught the fact that the citation fix was wrong too (Beresnev 1996 + Tanaka 2002 weren't barometric-null papers). Replaced with Rydelek 1988 (the textbook reference).
  • Bonus: RNG determinism bug fixed during R1
  • Acceptance: ✅ Issue closed (commits 52dad71, 618246f)

#44 — Granger V4 — full causal network

  • Status: ✅ ACCEPTED + Dana running
  • Headline: After HAC correction for serial correlation, only water_level → wave_height is a physically defensible Granger edge (HAC p=1.2e-5, replicates V3)
  • The HAC pivot (introduced in R1 revision after Mike's catch):
    • Plain OLS gave 3 surviving edges
    • HAC kept 1 physical (water_level → wave_height) and 1 confounder-flagged (earthquakes → temperature, weekly periodicity)
    • streamflow → earthquakes dropped from F=21.19 to HAC p=0.12 — confirmed as serial correlation artifact
  • Empirical FWER from 200 permutations: 0.24 vs analytic 0.05, confirming plain-OLS Granger is anticonservative on daily environmental data
  • Honest data audit: 4 of 6 requested new metrics excluded from BOTH tested set AND Bonferroni denominator due to insufficient data (CO2 daily, lunar tidal proxy passed; surface pressure, dst, solar X-ray, methane excluded)
  • Clean nulls: CO2 not Granger-linked to any geophysical/space-weather metric; lunar tidal proxy clean for earthquake triggering after correction
  • Mike R1 caught: pre-registration of 60-day threshold, missing HAC correction, us_aqi over-differencing, F=5.62 duplicate spot-check (was coincidence, not bug)
  • Mike R2 verdict: ACCEPT (commits a32c91d, 335ac86)
  • Acceptance: ✅ Issue closed, Dana copyedit running

Detours & Side-Quests

Anything that pulls us off the main schedule gets logged here.

(none yet)

Issues Encountered

Bugs, blockers, surprises.

(none yet)

Out of Scope (this sprint)

  • #111 Realtime alert classifier — depends on this sprint's Pulse persistence
  • #107 Knowledge graph 2.0 — big effort
  • #106 Lab page redesign — big effort
  • #110 NASA FIRMS — needs API key
  • #94 Space-Track + AMS — needs API keys
  • #97-100 Michael's web touches

Side-Quest Budget

If we finish all four early, hit one of:

  • #26 Backfill solar flux F10.7 (6 months of data)
  • #85 Real-time propagation dashboard
  • #27 NWS alert type taxonomy

Sprint End State (target)

  • Pulse persistence working (refresh shows recent events)
  • 1 new detection algorithm deployed (Forbush)
  • 2 new accepted papers (#45, #44)
  • Knowledge graph regenerated
  • Sprint doc updated with receipts

Sprint End State (actual)

✅ All four objectives complete. Plus a side-quest.

  • ✅ #1: Pulse persistence at GET /api/v1/pulse/recent, ring buffer of 200, replay on connect
  • ✅ #2: Forbush decrease detector at scripts/forbush_detector.py, registered hourly, idempotent. Caught the April 3 G2 event on first run.
  • ✅ #3: Earthquake-barometric-trigger paper accepted — clean null (LA d=+0.03, Houston disclosed as Eagle Ford induced seismicity contaminated by storm front)
  • ✅ #4: Granger V4 paper accepted — HAC correction collapsed 2 of 3 surviving edges, confirming water_level → wave_height as the only physically defensible Granger link
  • ✅ Side-quest: Knowledge graph builder reads metrics: from workspace.json + 5 workspaces patched. Detached 27-node WSPR cluster + 19-node HamQSL cluster collapsed into the main cluster (487 → 540 nodes, 1455 → 1593 edges).
  • ✅ Total accepted papers: 24 (was 22 at sprint start)

Daily Log

2026-04-08

  • Sprint planned. Schedule above.
  • Pre-sprint: 22 papers accepted, The Pulse running 14h+ with 330 events broadcast (310 quakes, 20 fireballs), 959 nodes in knowledge graph.
  • Pulse currently has no replay — new visitors see an empty rail until the next event.
Live Feed