# Maine forestry: an interpretable FIA sample

Version 2.0.0 | Krummholz AI | 2026-10-14 | Human/domain review pending

10,000 recorded tree observations selected from 1,066,726 rows in the USDA Forest Service Maine TREE snapshot downloaded October 12, 2026. Includes 28 inventory years, 93 species codes and all observed source status strata. These are observations from repeated inventories, not 10,000 necessarily distinct trees and not a current statewide census.

## Start here
Open trees-readable.csv for readable fields, or trees-readable.jsonl for typed records and identifiers preserved as strings. Read DATA-DICTIONARY.md, then try worked-examples.json. human-review.csv provides blank fields for your corrections and evidence. Source values are in selected-source-tree-rows.csv; source_csv_line includes the header as line 1.

## What changed
We sampled for coverage across inventory year, status and species; retained live, dead, legacy removed and not-presently-in-sample records where present; decoded species names and status/method codes; included measurement year separately from inventory year; and documented missingness without guessing its cause. Machine annotations are separate from editable human reviews. This is not a claim that any record was reviewed by a forester.

## Selection
Each stratum is inventory year × source status × normalized species code. Rank by SHA-256 of `krummholz-fia-v2|CN`, retain up to 8 per stratum, select round-robin across lexically sorted strata up to 10,000. All source rows were scanned. selection-log.csv reports source and selected counts per stratum, including any strata with zero selections. Small strata are intentionally overrepresented. This is a coverage demonstration, not a probability sample suitable for population estimates. Some plots are partial.

## Intended tasks and limitations
Useful for ingestion tests, source-grounded question answering, code interpretation, unit conversion and exploratory model prototypes. It is not a forestry prescription, commercial harvest recommendation, robot-action dataset, imagery ground truth, or validated production benchmark. No model has been trained/evaluated on this release.

No train/test split is supplied. Related trees, plot visits and remeasurements must be grouped before evaluating predictive models; PLT_CN identifies a visit and is not sufficient by itself to prevent cross-inventory leakage. PREV_TRE_CN and PREV_PLT_CN may point outside the sample. Species names/labels are deterministic decodings and can leak answers if used as predictors for their source codes.

TPA_UNADJ is preserved but must not be summed to claim a statewide total. This sample omits full conditions, strata, evaluation selection and all rows needed for proper population estimation. Original complete files remain unchanged. Nulls are not zeros. Protocols, field methods and modeled attributes vary across years; manual version and height-method code help interpret them.

No coordinates are included. Public FIA plot coordinates are perturbed and sometimes swapped; exact locations are confidential. Do not infer precise field positions or align these records to imagery as exact ground truth.

## Permissions and provenance
See LICENSE-AND-ATTRIBUTION.md and source-provenance.json. Official FIADB metadata identifies CC0 and permits use without additional permissions or fees. We retain USDA attribution and do not apply an MIT license to the collection. Documentation and third-party material are not automatically covered by the database license.

## Reproduce
Run build_forestry.py with Python 3.10+ and the four verified source files listed in source-provenance.json at the SRC directory defined near the top; edit BASE/SRC/OUT to your local paths. Python standard library only. Full original data and user guide are linked in provenance rather than bundled. The script verifies SHA-256 before processing. CHECKSUMS.sha256 covers all packaged files except itself.