MONDAY.md

Monday handoff (May 4, 2026)

State of the repo at end of Sunday May 3 / overnight into May 4. Demo is Sunday May 10.

Overnight pass (Sunday evening → Monday)

Eight priorities closed against audit/2026-05-03-evening-audit.md:

1. pitch/cold_open.md restored (was accidentally deleted in 1cb5ee6).
2. Granite Guardian / refusal-classification leftovers removed. Mellea is the sole grounding mechanism, period.
3. Trace UI is now clickable. Click any specialist row to reveal its raw structured output (formatted JSON, copy button, max-height + scroll). This is the auditability contract: every claim in the briefing is traceable to the specialist that produced it directly inside the UI, not just the citation appendix.
4. Buffered-footprint overlap for the three Point-geometry register specialists. NYU Langone / Stuyvesant HS / P.S. 89 now correctly register inside_sandy_2012=true. Each output records its footprint_buffer_m.
5. Map renders register-asset pins (subway / school / hospital / NYCHA-centroid) coloured by Sandy exposure with click popups showing name + [doc_id]. NYCHA polygon-fill is queued for when geometry_geojson lands in the dataclass.
6. floodnet_forecast specialist. TTM r2 forecast on the nearest FloodNet sensor's flood-event recurrence. Reuses the (512, 96) singleton already loaded for ttm_311_forecast. no new model class loaded into memory. The strongest single TTM win for the NYU CUSP audience.
7. Trace UI groups TTM specialists under one parent node forecasting.granite-timeseries-ttm-r2 [N instances] so the "one foundation model, multiple data streams" architectural story is legible without reading per-row metadata.
8. experiments/ cleanup: dropped two empty dirs (05_sam2_promptable, 06_chronos_bolt_forecast), renamed 05_terramind_finetune → 05a_terramind_finetune_micro to dedupe with the active NYC fine-tune dir, removed Riprap.zip from repo root.

Commit chain: a2143fc … through ed6ae9d. Morning handoff doc at audit/2026-05-04-morning-handoff.md summarises what to verify and what's queued next.

Where Sunday ended

All four keep-list items resolved + 4 register specialists shipped + AMD fine-tune prep green.

Item	Status	Path
Pitch cold-open locked	✓	pitch/cold_open.md
TerraMind-NYC fine-tune eval spec	✓	experiments/05_terramind_nyc_finetune/eval/eval_spec.md
200-query adversarial set + refusal eval	✓ (planner pivot)	experiments/06_granite_guardian/
Subway-entrance specialist (Sheepshead Bay)	✓	experiments/07_mta_entrances/
NYCHA-developments specialist (Red Hook)	✓	experiments/08_nycha_developments/
DOE-schools specialist (Coney Island)	✓	experiments/09_doe_schools/
DOH-hospitals specialist (Coney Island)	✓	experiments/10_doh_hospitals/
FSM integration of all 4 register specialists	✓	app/registers/, app/fsm.py, app/reconcile.py, web/static/agent.js
AMD droplet TerraMind smoke + STAC manifest	✓	129.212.182.52:/root/terramind_nyc/

End-to-end smoke on "Coney Island Brooklyn" produced citations [mta_entrance_56], [nycha_dev_239], [nycha_dev_166] alongside [rag_mta] and [nyc311]. Family-prefix chip routing works.

Last commit: 86861be (FSM integration of 4 register specialists).

Decisions locked

• Refusal classification dropped entirely. Planner-level classifier hit FN=0% but FP=7% (gate was <5%). Granite Guardian itself was already abandoned (laptop-infeasible). After the audit surfaced that the planner shim was documented-but-never-wired, the decision is now Option C: drop refusal handling. Cold-start framing scopes the audience; Mellea rejection sampling enforces grounding integrity; the four-tier glyph margin carries the epistemic-honesty signal. The GuardianRefusal.svelte component is deleted (was only ever rendered on a documentation page). Demo's integrity beat is the Mellea grounding-failure reroll on the curated Hollis 0.19% → 19% case. experiments/06_granite_guardian/ is preserved as a "considered and rejected" artifact for the methodology paper.
• AMD path: 129.212.182.52 is production, not 165.245.134.44. CLAUDE.md says the latter; fix CLAUDE.md to match reality. Production vLLM is on .52. The TerraMind container shares the GPU with vLLM; both fit on one MI300X.
• TerraMind manifest is 1028 paired chips, 2021-05 → 2026-04, NYC 5-borough hull +5 km, S2-cloud <30%, ≤3-day pair window. One year (2022-05 → 2023-04) returned 0 due to PC API intermittency. acceptable for the micro-fine-tune.

First thing Monday morning

1. Refresh Microsoft Planetary Computer signed URLs. They have ~1 hr TTL; the manifest from Sunday evening is stale by morning. On the droplet:

   ssh root@129.212.182.52
   docker exec -it terramind bash
   cd /root/terramind_nyc
   python build_manifest.py --refresh-only manifest_train.jsonl
   python build_manifest.py --refresh-only manifest_holdout.jsonl
   

   (Recipe is in /root/terramind_nyc/NOTES.md on the droplet.)

2. Kick off TerraMind-NYC fine-tune. Spec at experiments/05_terramind_nyc_finetune/eval/eval_spec.md. Budget is 30 GPU-hours; alarm at 25 (set on the droplet). Predicted actual: ~0.16 GPU-hours at bs=8 / 3 epochs. Don't run anything experimental until eval-spec gates pass on the held-out set.

3. Decide bucket (A ship-in-demo / B publish-only / C revert):

   • A: ship the fine-tuned checkpoint as a Riprap specialist.
   • B: publish to HF as msradam/TerraMind-1.0-NYC with model card, don't ship in demo. Bucket B is fully acceptable per the spec. Civic-tech publication discipline is the durable goal.
   • C: discard checkpoint, no public artefact.

Working on Monday

• TerraMind-NYC fine-tune (above).
• Mellea grounding-failure demo prep. The pitch demo is the Hollis 0.19% → 19% case where Granite emits a number with the wrong order of magnitude and Mellea catches it. Demo script needs to:
  • Show the failed first attempt (banner: "Mellea reroll: numerics grounding failed").
  • Show the second attempt with the corrected number.
  • Show the audit panel with the pass/fail per-requirement.
  • Show wall-clock for the reroll (target: under 30 s end-to-end).
  • Currently reproducible via scripts/probe_mellea.py --query "Hollis" --runs 5. The demo script is the visual version.
• MTA Sandy-recovery citation layer. Parse the MTA "Hurricane Sandy: Three Years Later" report into per-station-id facts so the subway-entrance specialist can emit [mta_recovery_<station_id>] doc messages alongside the exposure ones.
• NYCHA polygon-fill on the map. Overnight session shipped NYCHA developments as centroid pins on the map (graded by pct_inside_sandy ≥ 50%). The next tightening is to add a geometry_geojson field to app/registers/nycha.py's DevelopmentFinding dataclass and route through SSE so register-polygons actually renders graded fills (the layer + source are already present in RipMap.svelte).
• PLUTO/Building-Footprints join for Stuyvesant Town etc. Overnight pass shipped buffered-point overlap (NYU Langone, Stuyvesant HS, P.S. 89 now correctly flip to inside_sandy_2012=true). The 100m hospital buffer / 50m school buffer is honest but coarse; PLUTO + actual building footprints is the next step for the very-large-campus assets.

Outstanding through Friday

In rough priority order:

1. More specialists:
   • FEMA OpenFEMA NFIP claims tract-aggregated (pending).
   • NWS NWPS reach-level forecast + USGS NWIS Bronx / Saw Mill / Hutchinson rivers.
   • NYC DEP CSO outfalls + Bluebelt + Green Infrastructure specialist (CSS-vs-MS4 distinction for ASCE).
   • Three more TTM r2 specialists (USGS streamgage stage, NWS rainfall accumulation, NYC 311 sewer-backup citywide rate). FloodNet forecast already shipped in the overnight pass.
2. Visual identity refresh: Carto Positron, IBM Plex, four-tier epistemic palette, WeasyPrint PDF export, trace UI as <details> tree.
3. WCAG 2.2 AA pass.
4. Methodology paper draft (6-8 page PDF). Goal: Saturday May 9.
5. Historical-event mode. Vintage-cutoff queries. Saturday.
6. Five Build-in-Public posts through the week.
7. 5-minute hackathon pitch + 3 demo queries. Friday rehearsal.
8. ASCE talk materials. May 13 (post-hackathon).

Sharp edges to remember

• Static assets cache hard. When iterating on Svelte or agent.js, hard-reload (⌘⇧R). No cache-busting in place.
• HF Space sleeps after idle. Free tier; first request after sleep is a 30-90 s cold start. Ping the space before any demo.
• vLLM cold compile. First few requests against a fresh vllm serve log surprisingly low throughput while ROCm kernels JIT. Run benchmarks 3+ times before believing them.
• Sandy GeoJSON has self-intersection issues that blow up unary_union. Use buffer(0) (caught and fixed for NYCHA; may surface again for any new polygon-overlap specialist).
• DEP column is Flooding_Category (int16), not depth_class. Documented in NYCHA RESULTS.md.
• Centroid-edge join false-negatives on NYU Langone / Stuyvesant / P.S. 89 because their centroid points lie just outside the OEM Sandy polygon despite real 2012 basement flooding. PLUTO footprint join is the queued fix.
• Don't restart uvicorn while a model is mid-generation. Ollama keeps the request alive but the FastAPI handler dies, leaving the user staring at a dead stream.

Files to read in order on Monday morning

1. This file.
2. experiments/05_terramind_nyc_finetune/eval/eval_spec.md. The contract for what training output triggers ship/publish/revert.
3. experiments/06_granite_guardian/RESULTS.md. The Guardian → planner pivot decision record (so you know why Guardian is in the repo but not on the demo path).
4. experiments/07_mta_entrances/RESULTS.md. The canonical register-specialist pattern (the other three follow it).
5. CLAUDE.md. Fix the AMD droplet IP (165.245.134.44 → 129.212.182.52) at the same time as the first edit of the day.

Status as of 2026-05-03 ~12:50 ET

• Both git remotes (origin + huggingface) up-to-date through 86861be.
• HF Space rebuild was not triggered on the FSM-integration commit; do git push huggingface main when you want to deploy. (You may want to wait until Monday afternoon so a broken HF rebuild doesn't eat morning time.)
• Local Ollama has both granite4.1:3b and granite4.1:8b warm.
• AMD droplet 129.212.182.52 has the terramind container running with TerraTorch 1.2.7 + pystac-client + planetary- computer installed in system Python; HF cache populated.
• 200-query adversarial set + planner-pivot eval results reproducible from experiments/06_granite_guardian/ in ~3 min.
• Mellea probe still works: scripts/probe_mellea.py --query "Hollis" --runs 5.