cal-disp v0.5 CalVal report (draft for sign-off)¶

Release: cal-disp v0.5 (DISP-CAL-S1), Venti plan T50
Status: ready for sign-off, 2026-10-15, for review with Talib / OPERA CalVal
Candidate: cal-disp gamma-release e027309, Venti main f1da8aa (frame table 0.6-draft, area DB 1.1-draft), confirmed on all 10 frames through the frame table on 2026-10-15 (section 1.1). The table below is from the T46 runs; section 1.1 has the candidate's own numbers.
Baseline: cal-disp gamma 0.3 (Venti 600 km plane, tropo off, unwrap off)
Validator: disp2vlm_validation main d39956d. It uses only independent GNSS (UNR daily stations + MIDAS), never the calibration grid.

1. Verdict¶

v0.5 with remove-restore passes the release gate on all 10 frames. Houston (F08882) passes once its frame-table settings from the edge pass (2026-10-14) are applied: separate island regions and a packaged stable-pixel fit mask (venti-dev #35, cal-disp #15; section 4.1). Without them it fails, and the earlier proposal was a waiver.

#	Category	Frame	Gate	DD sill vs gamma	Category condition
1	Coastal subsidence + islands	F08882 Houston	PASS (+ RR, regions, fit mask)	−60%	met: subsiding RMSE 5.93 ≤ 6.13, defo_area 1.81 ≤ 2.52 mm/yr (4.1)
2	Stable inland	F08886 Oklahoma City	PASS	−81%	met: stable median −0.11 mm/yr (n = 24)
3	High relief + islands	F16940 Los Angeles	PASS	−72%	sill drops ✓; tropo mode not run in T46 (frame table: stratified; earlier trade study −15% sill)
4	Northeast / GIA	F08622 New York	PASS	−69%	LOS RMSE 1.23 mm/yr; VLM vertical, single look only (descending data gap, 4.4)
5	Fast basin	F36542 Central Valley	PASS (+ RR)	−69%	stations outside the bowl: median +2.42, RMSE 5.04 (gamma +10.61 / 12.80); bowl kept by remove-restore (4.2)
6	Coseismic event	F16941 Ridgecrest	PASS (+ RR)	−68%	step leakage into CAL 44% → 10% with remove-restore; non-spanning pairs unaffected (4.3)
7	Sparse GNSS	F07081 Montana	PASS	−64%	7 stations, stable RMSE 1.14 mm/yr; σ realism not assessable
8	Non-NA plate	F23211 Hawaii (PA)	PASS (+ RR)	−42%	Kīlauea area + 2018 event in the DB; plate PA recorded in the frame table (not used by any calibration step)
8	Non-NA plate	F35991 Puerto Rico (CA)	PASS	−81%	plate CA recorded in the frame table (not used by any calibration step)
–	VLM pair	F38238 Houston desc	PASS	−53%

1.1 Confirmation run with the release candidate (2026-10-15)¶

The T46 runs fed each frame's settings through its parameter file. The confirmation run calibrated every pair of the 10 frames the way production will run: the candidate code, the v0.5 parameter set, the materialized frame table and Venti's area and event databases. Each frame is validated against gamma. All 2,500 pairs calibrated; all 10 frames PASS.

Frame	DD sill vs gamma	gated median / RMSE (mm/yr)	T46 (same columns)
F08882 Houston	−60% [−64, −56]	+0.21 / 4.86	−60%, +0.21 / 4.86
F08886 Oklahoma City	−81% [−88, −63]	−0.11 / 1.37	identical
F16940 Los Angeles	−77% [−79, −74]	+0.67 / 3.33	−72%, +0.55 / 3.28
F08622 New York	−69% [−72, −65]	+0.11 / 1.23	identical
F36542 Central Valley	−68% [−72, −64]	+1.84 / 4.32	identical (San Joaquin +15 km run)
F16941 Ridgecrest	−68% [−74, −62]	+0.45 / 2.16	identical
F07081 Montana	−64% [−66, −62]	+0.97 / 1.79	identical
F23211 Hawaii	−44% [−58, −27]	+0.46 / 2.40	−42%, +0.48 / 2.41
F35991 Puerto Rico	−88% [−91, −83]	+0.02 / 3.18	−85%, −0.42 / 3.43
F38238 Houston desc	−46% [−53, −35]	+2.62 / 4.84	−53%, +2.47 / 4.87

Checked on every pair or every sampled product:

Los Angeles and Puerto Rico applied stratified tropo on every pair (285/285 and 190/190).
Houston used its fit mask on every pair (204/204).
The frame-table overrides loaded on every pair.
calibration_std is NaN exactly where calibration is (30 products, 3 per frame, no mismatch).

What moved, and why:

Los Angeles: stratified tropo and island regions from the frame table. The sill improves further (−77% vs −72%); the median rises by 0.12 mm/yr, within the 0.2 rule.
Puerto Rico: stratified tropo (TS-T1). Better on every measure.
Houston descending: now runs with remove-restore (the Houston-Galveston area); sill −46% vs −53%, median +2.62. The median above 1 mm/yr was already reported as a TARGET-1 note.

Every other frame's configuration is the one its T46 run used, and its numbers are identical. Runs: cal_disp_e2e/runs/<frame>-final; validation and release summary: cal_disp_e2e/t50conf/.

The gate (PRD 2.7, decisions 0001, 0002) FAILS a frame if: - the double-difference (DD) sill is not below raw; - the sill rises significantly by more than 5% against gamma; - velocity RMSE or |median bias| against MIDAS rises by more than 0.2 mm/yr.

The velocity rules use the stations outside deforming areas and event footprints (Venti DB 1.0-draft). Per-frame numbers are in Baselines.

2. Requirements traceability¶

Requirement	Evidence	Status
GATE-1 sill below raw	all 10 frames, −42 to −81% vs gamma	✓
GATE-2 sill vs baseline	no significant rise on any frame	✓
GATE-3 velocity vs baseline	10/10; Houston gated RMSE 4.86 vs gamma 5.51 (with its edge-pass frame settings)	✓
GATE-4 golden epoch	n/a until the v0.5 golden (T49, after sign-off)	open
TARGET-1 \|median bias\| ≤ 1 mm/yr (reported)	met on 8/10 in the confirmation run; not met: Central Valley +1.84, Houston desc +2.62	partial
R-V1 independent GNSS only	`GATE.independent_gnss_only`; TS-G1 is the sole grid reader and gates nothing	✓
R-V2 metrics per class	`velocity_by_class` in every run	✓
R-G2 50 km GNSS buffer	cal-disp PR #10	✓
R-G4 frozen grid snapshot + provenance	snapshot `unr_grid_0.3_IGS20_20261006` on S3 (T48); `/metadata/gnss_provenance` written (cal-disp PR #13). The T46 runs used per-frame UNR downloads; the snapshot is wired in at T49	partial
R-G5 GNSS-side remove-restore	`exclude_defo_nodes` + re-interpolation (cal-disp #11)	✓
R-G6 plate per frame	Venti frame table records NA, PA (F23211), CA (F35991); no calibration step consumes it (UNR grid is IGS20; plate only matters for a plate-motion layer, not implemented)	partial
R-S2 / R-S4 local-linear 50 km, coherence^8 robust	v0.5 config	✓
R-S5 exclude InSAR in defo/event areas	remove-restore (cal-disp #11)	✓
R-U1 unwrap correction off	off; TS-U1 no-go; T62 Houston: precise but incomplete	✓ (off)
R-E1 σ_CAL, k per frame	k = 3.9 (TS-G1); opt-in σ₀	✓
R-E2 DISP noise documented	"add 10 mm per point and pair", checked by TS-E4: coverage 91–99% on 9/10 frames (Puerto Rico about 16 mm)	✓
R-E4 σ realism (reported)	TS-E4: with σ_CAL + 10 mm + σ_GNSS, robust std(z) 0.62–1.02 and coverage 91–99% on 9/10 frames (Puerto Rico 1.64 / 74%); σ_CAL alone covers about 5% (surface-only by design). VLM: 85% on LA after `sigma_rate`	✓ (9/10)
R-O3 small instance (budget 6 GB since 2026-10-14)	v0.5 golden pair through the frame table (remove-restore, fit mask, extrapolated σ): 5.75 GB; 4.0–4.1 GB without them (T47)	✓ (EC2 run open)
R-O4 / R-O5 deterministic, golden in image	at T49	open

3. Trade studies¶

Study	Result	Report
TS-G1 grid vs stations, k	k = 3.9 frame default; σ₀ 2.6–4.9 mm opt-in	`studies/TS-G1/REPORT.md`
TS-S1 DISP noise < 50 km	6–10 mm per pair, 4–7 mm per point	`studies/TS-S1/REPORT.md`
TS-B1 coherence power	keep p = 8	`studies/TS-B1/REPORT.md`
TS-U1 unwrap correction	no-go for v0.5 (flag off)	`studies/TS-U1/REPORT.md`
TS-E4 product σ realism	10 mm DISP term realistic on 9/10 frames; TS-S1 per-frame values too small	`studies/TS-E4/REPORT.md`
TS-T1 troposphere by relief	0.3–1.5 km: relief alone does not decide; off for New York and Montana, stratified for Puerto Rico (velocity RMSE −0.38, \|bias\| −0.41 mm/yr); `auto` stays off in the band	`studies/TS-T1/REPORT.md`
T62 (Houston) unwrap on	precise, no mainland shifts; Galveston −18 → −12 mm/yr; incomplete	baselines.md

4. Findings the reviewers should know¶

4.1 Houston (PASS with the edge-pass frame settings)¶

Gated velocity RMSE: 5.90 against gamma's 5.51 mm/yr. Two areas drive it:
Galveston Island (GAL2, GAL7, TXGA, TDAM): −8 to −18 mm/yr;
Matagorda (TXBC, TXMG): +8.5 mm/yr.
Galveston is a disconnected water-mask region. About 24% of the raw pairs carry a whole-cycle jump against the mainland, and the 50 km surface absorbs part of it.
The category condition is evidence-based (owner, 2026-10-12): the subsiding-area RMSE must drop from raw (raw with its offset removed). It does not: 2.42 against 1.74 mm/yr.
What gamma does instead: it carries a frame-wide +4 mm/yr bias, but no added scatter.
Unwrap correction on brings the gated RMSE to 5.19, better than gamma, but it is incomplete.
Earlier proposal (superseded 2026-10-14): a waiver for Houston as a known limitation of coastal disconnected regions. The edge pass below makes it pass.

Edge pass (2026-10-13). All runs are v0.5 + remove-restore, gated under 0002.

F08882	regions only	+ stable-pixel mask (stack-mean coherence ≥ 0.7)	+ per-pair coherence ≥ 0.7
DD sill	110	101	100
gated median / RMSE / NMAD (mm/yr)	+0.73 / 5.35 / 4.20	−0.44 / 5.10 / 1.78	+0.07 / 5.91 / 2.51
defo_area median / RMSE	+0.87 / 2.42	+0.10 / 1.67	+0.66 / 1.96
subsiding RMSE (limit 6.13)	6.28	6.67	7.81
Matagorda TXBC / TXMG	+8.6 / +8.3	+1.4 / +0.8	+2.0 / +1.9
Galveston GAL2 / TXGA	−9.5 / −16.4	−18.1 / −13.5	−23.5 / −17.1
verdict	FAIL	FAIL (subsiding only)	FAIL

Regions (venti-dev #32/#33, on for F08882) keep the island cycle off the mainland.
Fitting only on stable scatterers fixes Matagorda. The 50 km surface was absorbing the phase-bias signal of decorrelating crop fields around the GNSS towns.
A per-pair threshold is weaker than the stack-mean mask (it also fixes most of Matagorda, but scatter returns on the coast and Galveston). Early pairs have 20–30% of the frame above 0.7 (against 9% for the stack mean), so crop fields that are coherent for a while still enter the fit.
Galveston remains the open failure. It is the sub-region error (T62), and it gets worse when the island offset is estimated from fewer pixels.

Final (2026-10-14): island offsets from every valid island pixel. The packaged mask (venti-dev #35) only removes pixels from the fit. The island offsets still use all valid island pixels, because Galveston's offset went wrong when it came from its few stable pixels.

F08882	packaged mask, island offsets from stable pixels	packaged mask, island offsets from all pixels
verdict	FAIL	PASS
DD sill (vs gamma)	96	102 (−60%)
gated median / RMSE / NMAD	+0.15 / 5.96 / 2.01	+0.21 / 4.86 / 2.95 (gamma +4.22 / 5.51 / 1.87)
subsiding median / RMSE (limit 6.13)	−0.04 / 7.90	+0.13 / 5.93
Galveston GAL2 / GAL7	−24.0 / −24.2	−10.0 / −10.1
Matagorda TXBC / TXMG	+1.6 / +1.4	+6.5 / +6.3

Houston passes.
Matagorda's stations sit in a land region of their own. That region also gets an offset, which picks the crop bias up again. Matagorda is a documented residual.
Galveston's sub-region error (TXGA −17) is unchanged (T62).

4.2 Central Valley¶

Without remove-restore, the 50 km surface removes up to about 60 mm/yr of real subsidence inside the bowl. Station RMSE rises from 17 to 24 mm/yr.
With remove-restore, the in-bowl median is +8.9 against +32.6 mm/yr, and the stations outside are unaffected (+2.42 / 5.04).
The +8.9 mm/yr left inside the bowl, and the frame's +2.4 mm/yr median outside it, are open items.

Edge pass (2026-10-13): San Joaquin polygon widened by 15 km (venti-dev #34, area DB 1.1-draft). The 10 km polygon left the bowl rim in the fit, and the 50 km surface absorbed the rim subsidence.

F36542, v0.5 + RR	10 km	+15 km
verdict / DD sill	PASS / 103	PASS / 107
defo_area n / median / RMSE (mm/yr)	20 / +8.86 / 34.82	40 / +1.42 / 20.55
stable median / RMSE	+2.14 / 4.06	+1.83 / 4.21

Most of the bowl residual is gone. The Kern polygons are outside this frame and keep 10 km.

4.3 Ridgecrest¶

The event DB (Mw 7.1 and 6.4 footprints) excludes spanning pairs from the fit.
With remove-restore the gate passes (median +0.45 against gamma's +0.27, within the 0.2 rise limit). Plain v0.5 misses it by 0.03.

Step leakage (PRD category 6, measured 2026-10-12). The test regresses CAL on DISP inside the event footprint, on 300 m blocks. In a pair spanning the events, DISP there is dominated by the coseismic step (std 149 mm).

pair	v0.5: slope / CAL std	v0.5 + RR: slope / CAL std
2019-05-29 → 07-10 (spans both events)	+0.44 / 77 mm	+0.10 / 27 mm
2019-05-29 → 07-04 (before the events, control)	+0.47 / 7.2 mm	+0.47 / 7.2 mm (identical)

Without remove-restore, about 44% of the step leaks into CAL. With it, about 10% does.
Where the rest comes from: far-field coseismic motion outside the drafted footprint (rupture length × 0.75 buffer). Widening the footprint is a T41.4 / edge-pass item.
Pairs that don't span the event are bit-identical with and without the DB (PRD 2.8).
The control's slope (and ~0.9 outside the footprint) is the calibration removing long-wavelength DISP error, not leakage.

Edge pass (2026-10-13): wider footprints rejected. Buffering the footprints to 1.5 × rupture length removes the residual leakage inside them, but it takes GNSS constraint away from the surrounding area:

spanning pair	current footprint	1.5 L footprint
inside: slope / CAL std	+0.10 / 27.1 mm	−0.03 / 19.6 mm
outside: slope / CAL std	+0.85 / 17.7 mm	+0.74 / 18.9 mm
frame gate	PASS	FAIL (median +0.30 higher)

The current footprints stay. The ~10% leakage is accepted and documented.

4.4 VLM (product built on v0.5 DISP-CAL)¶

LA (F16940 + F18904): vertical rate against GNSS Up, median −0.42, NMAD 1.44, RMSE 2.67 mm/yr on 172 stations seen by both looks; 80% are within 2 mm/yr. It meets the owner's 1–2 mm/yr target.
New York: the descending DISP has a gap from 2020 to 2022, so most pixels come from a single look. That still gives a rate NMAD of 1.3 mm/yr.

4.5 Uncertainty¶

σ_CAL follows R-E1, calibrated by TS-G1.
σ_CAL is the surface-only term. At the pair level it is tiny (median 0.2–0.7 mm) and covers about 5% of the residuals. That is by design and must stay explicit in the product text.
The user budget is σ_CAL + 10 mm (DISP) + σ_GNSS (TS-E4). It is realistic on 9 of 10 frames and slightly conservative.
Puerto Rico (tropical island, troposphere uncorrected) needs about 16 mm.

5. Open items before or after the release¶

Item	When
Sign-off and waivers (this report)	before T49
T49: flip defaults to v0.5, image `cal-disp:0.5.0-rc`, golden regenerated in the image, CHANGELOG	after sign-off
GATE-4 golden-epoch check on the v0.5 golden	with T49
T37.3: the remaining metadata fields (schema / frame-table versions, components)	with T49 (golden update)
TS-T1 tropo by relief	done 2026-10-14: Puerto Rico stratified (Venti frame table)
Confirmation run with the release candidate through the frame table	done 2026-10-15: 10/10 PASS (section 1.1)
Edge-case refinement pass: done 2026-10-14 for Central Valley (San Joaquin +15 km), Ridgecrest (footprints kept), Houston (regions + packaged fit mask: PASS). Open: Galveston sub-region (TXGA, T62), Matagorda residual (+6.5 mm/yr)	after venti-dev #35 / cal-disp #15 merge

Sign-off¶

Role	Name	Decision	Date
Owner
OPERA CalVal

Note on how the T46 runs were configured (added 2026-10-14)¶

The T46 runs did not pass Venti's frame table to cal-disp. Per-frame settings that changed results (Houston's island regions and fit mask) were written into their parameter files. As a result:

uncertainty.k_grid was 3.9 on every frame, not the TS-G1 per-frame fit. That changes σ_CAL only.
Los Angeles ran with tropo off, not the table's stratified mode.

Through the frame table, the table would not have applied anyway: cal-disp looked frames up as "8882" while the table keys them as "08882" (fixed in cal-disp #16). The T49 golden run goes through the table, as production will.

Separately, calibration_std was NaN where the loclin fit coverage was low: 42–89% of land on Houston, New York, Hawaii and Puerto Rico; 5–18% on Central Valley and Montana. The σ-realism checks (TS-E4) use station pixels, which are coherent, so they did not see it. Decided 2026-10-14 (owner): σ_CAL now has a value wherever the calibration does: the nearest fitted value, inflated up to 3× with distance (Venti #38, on in cal-disp's v0.5 set, cal-disp #18). calibration is unchanged, so no verdict moves. The extrapolation is to be refined later (Venti todo).