NEON Cross-Product Response Atlas

Explore: Launch the hosted atlas · Project home · Repository

The capstone of the NEONize family — an exploratory cross-product atlas of direct weather–response associations. The suite has nine product apps. This current baseline consumes seven of them (small mammals, breeding birds, plant diversity, vegetation structure, plant phenology, mosquitoes, and ground beetles); water chemistry and aquatic invertebrates are released companion products but are not integrated. The atlas lines those seven older pinned inputs up at shared terrestrial field sites on a common annual calendar.

2026-08-05 suite synthesis: all nine companion passes are complete. The measured Driver v2 synthesis rejects a blind pin refresh, authorizes no new ecological vote or canonical artifact change, and defers every complementary-app build pending exact feasibility support. The current atlas remains the frozen, exploratory v1 snapshot described below.

Phenology v2: the effect-blind adapter and primary-model registry and synthetic Seal 1 are published. Seal 2 stopped at its frozen legacy schema gate: one of 46 bundles has a typed required-column trend table with zero rows, so formal parity was not attempted. Current response support and every effect path remain sealed; the v1 artifact is unchanged.

Discharge F0 passed / F1 implementation published / acquisition pending token: the pre-payload specification, pure reducer, synthetic fixtures, and values-free response ledger freeze exact RELEASE-2026, response-key, two-table/QC, UTC-clock, zero/missing, and special-site rules before any discharge payload is accessed. The ledger freezes the prior pass-10/manual projection at Git blob c2aefd1aa7db8b1d7de4bf0551b1c95cba73f7a8 / SHA-256 79bb45911ab734ffc64444f248ac17ca42a78005707657fbe16effaef25e5296; F0 CI checks only its bytes and exact Inverts commit/tree metadata, never an RDS blob. PR #59 exact head 0a8b71ccb3c5f47c13a1f7f59d73dd2a297e6d5e passed run 31024370208 and merged as 28f00ecef8091a41af58db3f82ef9519ce940ceb; merged-master run 31024947729, Pages run 31024946671, and the public HTTP 200 check passed on that exact merge. Receipt PR #60 exact head 8954c1f8934356c1c1c41706c35cc42680a8b027 passed run 31026452964 and merged as b75996a85809ed0cd8ba89121e0de18e22063cc7; merged-master run 31027151110 and Pages run 31027144433 passed, and the public root returned HTTP 200 with bytes matching that exact merge. F1 implementation PR #61 exact head bacb1ea4def8dc70b4a1304fc75f4c9dc406bbd2 passed run 31034934400 and merged as c8ebe39f2ff6a45a523a64666caf5e15ce93b278. A GitHub-only runner-context validation defect was repaired by PR #62 exact head ce85c7c8bdaae490f99a38621dcafeb25ff05c1a, which passed run 31035951191 and merged as f16ab3ec06659c82a50c32276d0538365e20ef02. Final merged-master run 31036991024 and Pages run 31036988796 passed all seven jobs and deploy stages; the public root bytes match that exact merge. The protected discharge-f1 environment now requires reviewer tgilbert14, disables administrator bypass, and permits only literal master; its Driver-owned NEON_TOKEN is not yet configured, so no authenticated F1 run has occurred. F0 is passed. F1 is authorized only for the values-free availability/file inventory and reviewed non-observation schema metadata; F2 and observation-table bytes remain unauthorized, and discharge support remains UNMEASURED. The eventual gate can measure only response-anchored discharge-site-year and QC-pass record presence and may reopen review—not a build, metric, effect, prior, vote, Driver rebuild, or Driver byte change.

Construct warning: the current measurements do not supply a defensible annual production/seed-resource rung or a mediated-path test. The layered display is a co-display of candidate bottom-up pathways, not a tested trophic cascade.

Design history and future requirements: docs/CASCADE-ROADMAP.md is a historical expansion roadmap, not current product documentation. Its older chain proposals are retained to show what would have to be remeasured and prospectively registered before any future sequential or mediation analysis could be defended.

Why it's careful (and what it refuses to do)

Each NEON site currently has only 8–13 annual rows, while link-specific usable overlap can fall to 3–11 years. At that length a lag-aware correlation scan finds spurious signals everywhere. So the design is deliberately conservative — grounded in a literature review of bottom-up driver ecology and short-time-series statistics:

• Literature-motivated, build-locked, and explicitly exploratory. Every scorecard link has a literature-motivated sign + lag, and the current interface locks those settings while it runs. Repository history shows, however, that the family evolved while these same data were being inspected; it was not preregistered. The estimates and p-values are therefore exploratory association screens, not confirmatory tests. The app shows literature-motivated pairings only where their measurement limits can be stated honestly; for example, it posts no green-up→bird-abundance direction because trophic-mismatch studies do not supply one for this detection index. The Lag Explorer's max-statistic adjustment limits the additional lag/season search in the current interface, but cannot undo that historical selection.
• n-gated. Below 6 overlapping years no verdict is given — just the aligned series. At n≥6 a circular moving-block bootstrap CI sets the per-site direction verdict; an exact calendar-gap-aware circular-shift p is reported for transparency but does not set the tier. Its link-specific finite-randomization floor is 1 / (n_null + 1); the app exports n_null, p_floor, and full calendar span so missing years cannot masquerade as adjacent observations.
• Direction over magnitude. The cross-site association summary gives one raw-level vote per testable site for each of the two vote-eligible temperature–green-up rows, under an explicit (and imperfect) site-independence assumption. Its exact-binomial p is one-sided against 0.5 sign symmetry; the table also reports Holm familywise and Benjamini-Hochberg FDR adjustments. A within-site multi-link tally is descriptive only because links at one site reuse drivers and responses. A spatial sensitivity collapses the identical raw vote population to one majority per NEON domain (50/50 domains abstain), without adding another p-value.
• Machine-readable selection disclosure. Every bundle records a versioned prior-family status: literature-motivated and locked for that build, but historically co-developed with the analyzed data. A future confirmatory analysis would need an immutable registry and genuinely held-out observations.
• Snapshot-aware phenology. Both green-up constructions are retrospective full-panel standardizations: recurrence/connectivity gates, species centers, and additive year effects are refit when the eligible panel changes. A refresh can therefore revise historical index values rather than only append a year. Source and bundle hashes make each published snapshot exactly auditable; neither index is a prospectively fixed annual reading.
• Never "drives," "causes," or "the cascade holds." A handful of annual points cannot establish causation or mediation. No SEM/path model, no naive year-pooling, and no inference from visual arrow order.

Conditional grouping rule (an explicit sensitivity, not a measured biome class)

Sites retain a deliberately visible descriptive grouping based on a keyword heuristic over the one-line site description: descriptions containing desert, sagebrush, or semi-desert enter the dryland-keyword group; every other site enters the default other-site group. The legacy internal keys remain water-limited and temperature-limited, but they must not be read as measured aridity, resource limitation, or a validated biome classifier. Dry forests, semiarid grasslands, Mediterranean sites, and mixed sagebrush–conifer sites can be misgrouped. The bundle stores both the assignment basis and the exact rule. The current vote-eligible temperature–green-up associations include all sites rather than conditioning their p-values on this heuristic; non-green-up links are contextual only.

Seasonal climate windows (winter Oct–Mar, monsoon Jul–Sep, spring temperature Mar–May) are reconstructed from complete monthly overlays because annual aggregation can blend distinct periods. A named window is still only a proxy: unless a response is calculated over the same dates, the app does not call it a same-season or mechanistically direct test.

Tabs

• Overview — what is measured, which direct associations are actually vote-eligible, and the construct warning.
• Layered Timeline — standardised (z-scored) annual signals co-displayed by measurement layer; direct-link cards keep visual patterns separate from tested pairings. At dryland-keyword sites, a Seasonal Climate panel compares annual, winter, and monsoon associations that an annual total can blur.
• Driver Lab — pick a response; every current build-locked literature-motivated pairing is screened, with the aligned-pairs scatter (now carrying its r / n / p / 95% CI on the figure) and a tier-honest fit line.
• Across NEON — the scoreboard: each vote-eligible link pooled across sites (one vote per testable site) + a site × link grid encoded by glyph and colour, with context-only and under-supported rows shown rather than hidden.
• About — the full current pairing table + the honesty manifesto + selected references.

What it surfaces

Exact counts and p-values are generated from the committed bundle and shown in the app rather than copied into hand-maintained prose. This matters because coverage gates and monthly source refreshes change the denominators. The pooled table separates vote-eligible rows from pairings that are context-only by design; the latter are not merely under-supported and cannot become inferential through more years alone. It reports raw, Holm-adjusted, and FDR-adjusted values plus detrended, consecutive-year-change, and alternate composition-standardization direction sensitivities. The annual-temperature and March–May-temperature green-up screens are presented together; neither climate proxy is guaranteed to precede the onset observations.

Seasonal aggregation comparison (single-site, non-pooled diagnostic)

The seasonal desert comparison is an illustration, not a pooled result. Its current r/n/p values are read from the same cached link rows as the plots; they are never hard-coded in prose. Raw plant richness is now explicitly effort-confounded context (expected_class='none'), not a vote: the bundle persists sampled plot and plot/subplot/scale counts until a defensible coverage-standardized richness series is available.

Run it

R 4.5.x: shiny::runApp(".", port = 8194). The default site is SRER (Santa Rita, a Sonoran-desert case used to compare annual and seasonal driver summaries).

The required deployable data family is:

• data/cascade.rds — annual signals, pairwise screens, pooled direction summaries, site context, and provenance;
• data/search_index.rds — a fingerprinted search derivative of that bundle;
• data/cascade_meta.rds — the fingerprinted companion meta-analysis; and
• data/neon-cascade-codebook.csv — the generated field contract.

App boot verifies the manifest's exact 12-file checksum map before sourcing local code or deserializing any RDS, then cross-checks the three RDS lineage stamps; inconsistent generations refuse to start.

Data and measurement limits

data/cascade.rds is list(annual, signals, priors, codebook, suite_links, pooled, site_meta, meta). The annual per-site table—including precip_winter, precip_monsoon, and temp_spring—is assembled from seven sibling apps’ committed bundles plus the small-mammal app’s NEON-tower climate overlays. The builder does not refetch NEON data or execute sibling application code.

Annual temperature and precipitation require all 12 distinct months; winter requires 6/6 months, monsoon 3/3, and spring temperature 3/3. Every month count is retained. Biological support fields (individuals, sampled plant units, trap-nights/catches, and observed bird point-count occasions) travel beside the signals. Bird flyovers are excluded, but annual zero-detection visits can be absent. Beetle effort contains catch-bearing events only, so that denominator is outcome-conditioned rather than a valid all-event CPUE denominator. Mosquito effort-only years remain real zero-catch observations.

Phenology excludes left-censored first visits and uses a species-centered, equal-species, DOY-anchored timing index with explicit contributor, species, exclusion, and onset-interval-width support. The additive species+year construction is a full-precision sensitivity estimate. Both are retrospective full-panel standardizations: a refresh can revise earlier values when the eligible species×year panel changes.

fruiting_pct is an opportunistic maximum over observed eligible months, not a fixed-season annual estimate; fruiting_n_eligible_months and fruiting_peak_n_individuals expose unequal opportunity. The vegetation context is a mean live basal area per hectare conditional on plots with qualifying stem records. veg_n_plots, veg_area_eligible_plots, and veg_record_plots expose the denominator. The source cannot distinguish sampled-zero from unsampled plots, so zero-stem and unobserved plots are not imputed and this value must not be read as universal site-wide standing stock or productivity.

For the two five-site-eligible temperature–green-up rows, the companion uses REML random effects with Knapp–Hartung inference, reports a 95% prediction interval and heterogeneity before directional p-values, and applies Holm adjustment over the two-row family. It remains an exploratory sensitivity analysis.

Reproducible rebuilds

Run only Rscript scripts/rebuild_all.R. Set CASCADE_ROOT to the directory containing all seven current input repositories.

Ubuntu 24.04 with R 4.5.2 and the dated Posit snapshot is the canonical release-byte environment, and CI requires its scientific artifacts to reproduce exactly. Windows keeps strict schema, class, attribute, key, text, support, and scientific-decision checks plus only explicitly named bounded full-precision diagnostics; artifacts are never rounded to manufacture cross-platform identity. Each repository must be at its canonical origin; the archived data/sites scope (plus mammal data/env) must be clean and all consumed inputs tracked. Unrelated editor metadata outside those inert data roots cannot affect the build. The builder records the exact commits and input hashes, creates an immutable git archive snapshot of each recorded commit, verifies the extracted inputs against that inventory, and reads only those snapshots. A final live-source check detects ordinary source changes during the run.

Unless overridden, the analysis cutoff is deterministic: UTC year of the newest source commit minus one. CASCADE_LAST_COMPLETE_YEAR is accepted only as an exact four-digit historical year no later than the source-derived ceiling.

The single-writer rebuild lock prevents concurrent publishers. An empty isolated generation is populated from an in-memory snapshot of the build/deploy code, then runs cascade → search → companion meta-analysis → raw-source/artifact contracts → manifest write and verification → malformed/mixed-generation rejection tests → UI/server serialization and initial reactive smoke → final manifest verification. Only then are the five generated files checksum-verified and promoted; ordinary failures restore the previous byte-exact family, and the live manifest is verified again before the lock is released. The manifest is promoted last, and app boot checks its exact 12-file checksum map before sourcing repository code or deserializing RDS, so an interrupted promotion refuses to serve a mixed generation.

The manifest's exact allowlist is the verified runtime contract. A git-backed Connect deployment may still archive support files from the repository target directory, so the allowlist is not an assertion that scripts, reviews, README.md, or docs/ are physically absent. The app never sources or reads those support files; docs/index.html remains the separate GitHub Pages landing page. Built by Desert Data Labs · desertdatalabs@gmail.com. Not affiliated with NEON/Battelle/NSF.