The document a consultant can attach to internal review. We explain how the data is sourced, how the screening memo is drafted and validated, where the model can fail, and where the consultant of record remains responsible. We don't republish the regulatory commentary itself.
Geographic clips of U.S. federal environmental datasets, packaged as a single ZIP that extracts to a folder-organized layout matched to how GIS practitioners actually work. Every pack includes the AOI boundary, the layer data in multiple formats, embedded and sidecar styling for QGIS and ArcGIS Pro, raster overlays where applicable, and the metadata + citation trail required for a regulatory submission.
pack-<job_id>/
├── README.md per-pack overview + per-layer status
├── METADATA.json machine-readable manifest (v1.2)
├── CITATIONS.md / CITATIONS.json cited source per layer
├── data/ primary delivery
│ ├── AOI_Boundary.gpkg / .geojson
│ ├── <Layer>.gpkg GeoPackage + embedded QML style
│ └── <Layer>.geojson JSON-native vector format
├── shapefiles/ legacy-tool compatibility
│ └── <Layer>.shp / .dbf / .prj / .shx / .cpg
├── kml/ Google Earth + Earth Pro
│ └── <Layer>.kml inline <Style> per feature class
├── rasters/ land cover, hazard, elevation
│ └── <Layer>_cog.tif Cloud-Optimized GeoTIFF + colormap
└── styles/ one symbology definition per layer; six rendering forms
├── HOW_TO_IMPORT.md GIS-specific quick start
├── qgis/<Layer>.qml also embedded in the .gpkg
├── arcgis-pro/<Layer>.lyrx drop-in CIM JSON, one-click symbology
└── universal/<Layer>.sld OGC SLD 1.1.0 fallbackEvery vector layer ships in four formats (GeoPackage, Shapefile, GeoJSON, KML); every raster in two (raw GeoTIFF + Cloud-Optimized GeoTIFF with an embedded colormap). The styling stack covers all four major desktop GIS environments without forcing the consultant to author symbology themselves. Data packs are deterministic: same AOI + same cached source data vintage produces the same files. No LLM is invoked for a pack. The screening memo is a separate add-on.
A 7–22 page PDF (typically ~8 / ~12 / ~18 pages for Standard / Pro / Expert respectively) structured around a fixed section template (Standard ships seven sections, Pro nine, Expert ten), so the customer always knows what they're getting before they buy. Standard and Pro memos are generated deterministically. Given the same AOI, the same federal layer data, and the same corpus version, the engine produces byte-identical output every time. The narrative is composed from a versioned library of corpus-validated sections rather than from a language model. No LLM is invoked for Standard or Pro. Expert memos add a project-context AI synthesis layer (Anthropic Sonnet 4.5) on top of the same deterministic base, for executive-summary tailoring, comparative-project framing, and cross-statute narrative. The AI layer's contribution is clearly disclosed in the memo header so reviewers can distinguish deterministic content from AI-synthesized prose. Every memo (Standard, Pro, or Expert) is validated against the active regulatory corpus before delivery, and the corpus version + a content hash of the active citation whitelist are stamped into the footer. Both values resolve to a specific section of the public Currency Ledger so any reviewer can verify which authorities were in scope when the memo was issued.
A neutral, machine-extracted feed of recent environmental-compliance documents from the U.S. Federal Register, refreshed daily. Strictly an awareness layer. Avow surfaces what was published and links straight to the authoritative source. We do not tell the user what an item means for their project; that boundary is a functional requirement, not a disclaimer. See the Monitor directly or the spec discipline in Section 9 below.
The active source set (28 federal datasets) spans BLM, CDC, Census, EPA, FEMA, HUD, MRLC, NMFS, NOAA, NPS, USACE, USDA, USFWS, USGS. Each pack's METADATA.json is the source of truth for what was actually used in that order: the per-layer agency, endpoint URL, retrieval timestamp, native CRS, AOI clip extent, and a per-layer success flag.
Avow fetches every layer live from the original agency endpoint at pack-generation time, so each pack reflects the current authoritative federal source. For a curated subset of high-volume federal datasets, Avow also maintains a versioned local mirror sourced directly from those same agency endpoints. The mirror is a backstop, consulted only when the live source is unavailable, so a transient agency-side outage doesn't block a customer's order. It is never a warehouse-of-record that pre-empts fresh data.
The remaining sources, typically smaller or naturally per-AOI in scope, have no mirror and are always fetched live, with the reliability discipline described in Section 3.
Every layer in every pack records the agency source URL (so the citation chain to the authoritative source is preserved) and a retrieval timestamp (so the consultant can verify when the underlying data was retrieved). The delivery path actually used for each layer, the live primary, a documented fallback, or the cache backstop, is recorded in METADATA.json.
Three commitments underpin every pack build:
Per-layer outcomes are recorded in admin-only telemetry so endpoint instability is detected before customers see degraded packs. Telemetry is never linked to customer email in any internal or external view.
Each layer has a single canonical symbology definition (palette, classification field, renderer type) versioned as symbology 1.0. Every output format derives from that one definition (there is no manual per-format styling), so a palette change propagates atomically across all six artifacts:
.qml files in styles/qgis/ cover tooling that reads the sidecar form..lyrx layer files in styles/arcgis-pro/ resolve to the pack's data files via relative paths, so customers don't manually re-link sources.styles/universal/ for any tool that consumes the standard (QGIS, GeoServer, MapServer, MapProxy).<Style> blocks so the layer renders correctly in Earth / Earth Pro with no external styling.The discipline: a palette change updates one place; all six outputs regenerate from it. We refuse to author symbology manually in any output format.
Not every federal endpoint returns data for every AOI. A polygon in Wyoming has no tribal-lands overlap; a small AOI may not intersect a critical-habitat layer; an upstream endpoint may be down at the moment the customer ordered. To make these outcomes legible, and to keep "the layer is empty" distinct from "the layer failed to fetch," every layer carries an explicit status. The status taxonomy distinguishes:
Each pack SKU carries a contract declaring which layers are required, which are optional, and what to do if a required layer is unavailable. The contract is evaluated at pack-finalization time. A pack ships as complete only if every required layer is accounted for; otherwise it ships as degraded, with an explicit degradation notice in the pack metadata and the per-pack README explaining what's missing and what action the consultant should take.
| Tier | Sections | Engine | Turnaround | AI involvement |
|---|---|---|---|---|
| Standard +$249 (pack add-on) | Project summary · resources present · applicable statutes · permit pathway · short risk flags · scope disclosure · citations | Deterministic | ~5 seconds | None: no LLM is invoked |
| Pro +$499 (pack add-on) | All Standard sections, plus Section 7 ESA assessment, expanded risk register, and next-steps recommendations tied to the pack contract | Deterministic | ~5 seconds | None: no LLM is invoked |
| Expert +$849 (pack add-on) | All Pro sections, plus timeline estimate and per-gap consequence analysis on every unanalyzed layer in the scope disclosure | Deterministic base + AI synthesis layer | ≤ 15 minutes | Anthropic Sonnet 4.5 for narrative tailoring + comparative framing on top of the deterministic base |
Scope disclosure appears in every memo (Standard, Pro, and Expert). It lists exactly which federal layers the screening analyzed, which were not included in the pack, and what the customer's next step is if a missing layer is material to their project (typically: upgrade to a pack that includes it, or request a custom add-on). Standard renders the disclosure as a plain list; Pro and Expert render it as structured tables tying each unanalyzed layer to its regulatory topic and a recommended uplift pack.
The engine name, tier, corpus version, and (for Expert) the AI model used are stamped into every memo footer. Tier choice is the customer's; Avow does not silently upgrade or downgrade. If the Expert AI synthesis layer is temporarily unavailable at the moment of generation, an Expert memo is delivered with the deterministic base alone and the absence of the AI layer is explicitly noted in the header, never silently omitted.
Why the deterministic approach for Standard and Pro? For licensed environmental consultants, reproducibility and auditability are the differentiating product attributes. Same input → same output means every memo is a verifiable artifact: an internal reviewer, an agency, or opposing counsel can re-run the same AOI against the same corpus version and produce the same memo. There is no run-to-run variance, no hallucination surface, and no API-latency variance. The Expert tier exists for customers who want narrative polish layered on top of that auditable base.
The corpus is the bibliographic ground truth the model is required to cite from. It is versioned (currently v2026.05.2) and entries are retired when an authority is rescinded or superseded; retirements are logged, never silent. The corpus covers the federal compliance frameworks Avow's customers operate under:
The corpus is split between a structured authority list (which is reviewable via the Currency Ledger) and proprietary explanatory commentary (not republished). Every regulatory shift that drove a corpus revision is captured in the Outcome Ledger (Section 8) with a public-record source link.
The Outcome Ledger is the append-only published log of regulatory observations that informed the corpus. The Currency Ledger (Section 7) tells you what the corpus says today; the Outcome Ledger tells you how we got there: every court ruling, Federal Register notice, or agency rule change that drove a corpus revision, with a verifiable public-record source URL.
Intake discipline:
Every published entry carries a tracked timestamp, so the reasoning behind any corpus revision is auditable as of the date it was made. Agency reviewers, internal counsel, or third-party auditors can reconstruct the regulatory-shift evidence that informed a specific corpus version.
The Monitor is a free, no-login feed of recent environmental-compliance documents from the U.S. Federal Register, refreshed daily. It is strictly an awareness layer, not an interpretation engine. The boundary is functional, not cosmetic:
federalregister.gov. The Avow page is a layer over the public record, never a replacement for it.auto-extracted tag indicating the text is Federal Register material that Avow surfaced, not authored.The Monitor is also where notable regulatory shifts surface for promotion into the Outcome Ledger. The operator (founder, at launch) sees an item in the feed that affects how Avow screens; if it's material, they write an Outcome Ledger entry against it. The Monitor is the firehose; the Ledger is the curated, durable record.
Every CFR, U.S.C., Executive Order, and case citation in any Avow memo (Standard, Pro, or Expert) is bound to the active citation whitelist tied to the memo's corpus version.
For the deterministic Standard and Pro engines, every citation is emitted from the whitelist by construction. The engine cannot produce a citation that isn't on the list. For the Expert AI synthesis layer, citations produced by the model are validated against the whitelist post-generation; any citation that is not on the whitelist, or that matches an authority retired since the memo's corpus version (for example, 40 C.F.R. §§ 1500–1508 after the January 2026 rescission), is replaced in the memo body with [CITATION UNVERIFIED — REVIEW REQUIRED] and the memo is flagged for human review before delivery.
The corpus version and a content hash of the whitelist active at the moment of generation are stamped into the memo footer. Either value resolves to a specific section of the public Currency Ledger, so any reviewing party can reconstruct which authorities were in scope when a memo was issued.
Standard and Pro tiers do not invoke an LLM, so prompt-injection is not a vector for them. The safeguards below apply to the Expert tier's AI synthesis layer.
Avow does not accept uploaded AOI files. The project area is drawn directly in the browser using a constrained polygon tool; the API receives a pure GeoJSON geometry: coordinate arrays only, no attribute payloads. This sidesteps the most common file-upload-driven prompt-injection vector entirely.
Three guards apply to any user-supplied input that could reach the Expert tier's LLM:
Standard and Pro have zero per-memo marginal cost (no LLM is invoked). The Expert tier carries LLM costs, governed by three layers of spend discipline that keep customer-facing reliability decoupled from upstream cost variance:
Applies to every tier (Standard, Pro, and Expert). A screening memo never:
Every memo carries a cover-page disclaimer banner and a per-page footer. Avow's aggregate liability for any deliverable is capped at the purchase price of that deliverable. The consultant of record (or the customer's retained licensed professional) retains full professional responsibility for any document derived from an Avow output and submitted to a federal, state, or tribal agency.
By using Avow, the customer represents that they are, or are acting under the direction of, a licensed professional and will independently verify all outputs before relying on them. Full terms are at /legal/terms.
The corpus is reviewed on a scheduled cadence and minor-versioned each month. Hotfix versions are tagged promptly on a material regulatory event: a SCOTUS decision affecting NEPA / CWA / ESA, a rescinded or finalized federal NEPA-implementation regulation, or a comparable trigger. The Currency Ledger shows the active version, last review date, and the source list under monitoring. Memos generated after the next scheduled review window opens carry a "review window approaching" banner; memos significantly older than that window are blocked from agency submission and must be re-run against the current corpus.
For methodology questions, an agency-reviewer's clarification request, or to verify a specific corpus version: [email protected].