Second-generation analysis

Fire VASE changes the response being explained: ordered wildfire development, not only final area, duration, or peak growth. The manuscript narrative develops that project-level argument. This page documents the corrected empirical implementation: developmental morphology is constructed before weather is considered, and weather remains an external, non-causal explanatory layer. The v2 technical manuscript supersedes the climate-centered v1 generation for quantitative claims.

The second-pass validation freezes a bounded scientific story: broad shape gradients and nonrandom temporal ordering are supported, but event-weather predictability is weak and heterogeneous. The new main-16.pdf and supplementary-2.pdf define the narrative direction and intended robustness questions; validated repository outputs, rather than unresolved draft values, control the numbers reported here.

What changed

True observed daily peak replaces catalog area divided by duration. Entropy uses normalized reconstructed growth; it is undefined for incomplete or zero-total histories. Current catalog and reconstructed totals happen to agree.
The primary morphospace uses only normalized growth-allocation bins for 10,246 fires with at least three consecutive observations. Area, duration, observation count and weather are external attributes. PC1 explains 34.1%; five axes explain 89.4%. Neither compression nor a wedge proves biological constraint: duration/count-preserving nulls also compress strongly.
Nested event-weather models share 9,212 complete primary fires and use region, year-block and strict crossed spatiotemporal holdouts. The old 0.349 cross-response median is withdrawn; each response is reported separately.
Only exact calendar-day transitions are used for subsequent growth. Weather is sampled at day-t newly burned-area centroids, not final-event centroids. The shared autoregressive cohort has 87,944 transitions from 31,700 fires. Additive weather gives a small regional increment (~0.005 R²); interactions give ~0.018 above state. These are reconstructed-data associations/predictions, not proof of causation or information availability in an operational system.
Matching uses unique partners, declared calipers and nuisance strata. Representative median-mismatch examples replace adversarial maxima. Observed mismatch rates are compatible with conditional permutation ranges; mechanistic interpretations remain hypothesis-generating.

What survived scientific validation

At least seven consecutive observations leaves 1,171 fires. Relative to the primary fit, five-axis distance ranks on shared anchors correlate 0.969, but 15-neighbor overlap is 71.8% and five-axis coverage falls to 74.2%. Broad gradients are more stable than exact local neighbors or extreme exemplars.

In a recorded 4,000-fire sample, first-half growth allocation averages 0.541 versus 0.500 after shuffling the same increments, and detected pulses average 1.249 versus 1.413. This supports nonrandom observed ordering, not a uniquely biological low-dimensional wedge. Excluded endpoint attributes remain moderately associated with shape. No formal test establishes the absence of latent classes.

Adding year to duration/count/area/region/month controls leaves small weather associations: mean VPD versus late allocation has partial rank correlation -0.069; precipitation versus pulses is only 0.025, with a regional interval including zero. The weather-complete population is CONUS-only in these inputs.

VPD-by-current/prior growth products add about 0.006-0.012 R² above the other weather interactions across holdouts. This is supported with caveats: coefficients vary by region, fire size and partial edge years; all exposures are reconstructed end-of-day observations. Matching caliper sensitivity confirms design-dependent coverage, not ecological mismatch prevalence.

The repository's analysis/scientific_validation/ contains the 26-point final report, correction audit, complete A-M claim matrix, scientific story, and PRISM_HANDOFF.md. Those files control manuscript interpretation; the original analysis/v2/ remains the separately identified candidate baseline.

Reproduce

Use the existing real data lake; absent required inputs raise an error. No new external data download or synthetic fallback is part of this workflow.

PYTHONPATH=src:scripts OPENBLAS_NUM_THREADS=1 MPLCONFIGDIR=/tmp/fire-vase-v2-mpl .venv/bin/python manuscript_figures/00_run_all.py --generation v2 --data-lake data_lake/fire-vase-data-lake-v0.1
.venv/bin/python -m pytest tests/test_analysis_v2.py -q

The configuration is config/analysis_v2.json, seed 20260828. Use --render-only with the figure command after successful statistics generation. The full pipeline also runs the bounded scientific-validation stage. To rerun only that stage or verify its artifacts:

PYTHONPATH=src:scripts MPLBACKEND=Agg MPLCONFIGDIR=/tmp/fire-vase-v2-mpl OPENBLAS_NUM_THREADS=1 .venv/bin/python scripts/validate_fire_vase_science.py
PYTHONPATH=src:scripts MPLBACKEND=Agg MPLCONFIGDIR=/tmp/fire-vase-v2-mpl .venv/bin/python scripts/verify_fire_vase_science.py

Full test collection currently fails in two pre-existing modules importing the missing tests.helpers.contracts; the remaining suite passes (132 tests, two skipped). Dedicated v2/validation tests are included. This is a disclosed testing gap, not a fully green repository test claim. The v1 comparison is deliberately separate:

PYTHONPATH=scripts/figures:scripts:src .venv/bin/python scripts/reproduce_v1_comparison.py

Outputs and provenance

Output	Repository location
Audit and old/new conclusions	`analysis/v2/audit_report.md`, `issue_audit.csv`
Per-event growth/entropy/date audit	`analysis/v2/event_semantic_audit.csv.gz`
PCA, nulls, stability and every model comparison	`analysis/v2/*.csv`
Source and exposure provenance	`input_audit.json`, `input_hashes.json`, `day_t_weather_manifest.json` in `analysis/v2/`
Code/configuration/output hashes	`analysis/v2/run_manifest.json`, `publication_manifest.json`
Five main figures and three supplements, PDF/PNG/SVG	`figures/v2/`
Second-pass audit, claims, stress tests, handoff and frozen hashes	`analysis/scientific_validation/`
Current manuscript source and PDF	`docs/manuscripts/fire_vase_developmental_morphology/manuscript_v2.md`, `output/pdf/fire_vase_v2_manuscript.pdf`
Preserved references and reproduced v1 outputs	`archive/comparison_v1/`

Large regenerable Parquet tables remain local and ignored by Git. The source package retains its FIRED/gridMET citation and reuse terms. No data are relicensed by this analysis. The full manuscript describes the strict cohort's selection limits, conditional rather than refitted-model uncertainty, the small number of region clusters, satellite/date uncertainty and matching assumptions.

Figures

Figure 1: dated real growth observations and VASE encoding.

Figure 2: shape-only morphospace, observation support, leading loadings and real glyphs.

Figure 3: external weather mapping and weak response-specific blocked prediction.

Figure 4: exact subsequent-day growth and incremental weather skill above state.

Figure 5: caliper-constrained unique pairs and conditional mismatch references.

Second-pass supplements

Figure S2: observation-threshold, null, endpoint, adjustment and matching-caliper tests.

Figure S3: VPD-specific joint support, coefficient heterogeneity and held-out ablation.