Repository navigation
Scoring tuning: raw-% rock fragment + site weight 0.25 + dis_max consistency (v2.4.0) - #397
Merged
Merged
Conversation
garobrik
approved these changes
Sep 25, 2026
johannesparty
force-pushed
the
feat/soilid-site-weight
branch
from
September 25, 2026 18:19
741d04f to
3d51128
Compare
johannesparty
force-pushed
the
feat/explain-html-renderer
branch
2 times, most recently
from
September 25, 2026 18:59
21dcf66 to
1d7cb78
Compare
…eight Three low-risk scoring cleanups (see SOILID_TUNING.md §7 for the analysis): - global dis_max: np.nanmax(dis_mat_list) -> 1.0, matching the US path. Gower distances are ~[0,1] so a fixed, interpretable, query-independent penalty is cleaner than the empirical max. Accuracy effect negligible (Δtop1 ≈ -0.04). - depth weight: surface 0-20cm weight 0.2 -> 1.0 (uniform) in both US and global. Tuning showed the surface de-weighting slightly hurt accuracy (~+0.3pt top1); kept as a two-part concat so the surface band stays an easy knob. - cross-reference comments on the duplicated horizon-property bounds tables (us_soil.global_prop_bounds <-> global_soil.GLOBAL_HORIZON_PROP_BOUNDS). Unit-test snapshots regenerated against the pinned soil-id-db image and verified deterministic. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…s midpoint Both the US and global horizon comparison fed rock fragment into ranking as a 5-bucket class midpoint (0/8/25/48/80 via getCF), for both the pit and the candidate. Quantizing the candidate created a cliff at every class boundary — 15% vs 16% scored as 8 vs 25 (near-total mismatch) while 16% vs 35% scored identical — turning small real differences into large scored ones and erasing within-class differences. Use the candidate's RAW rock-fragment % (c_cfpct_intpl, already computed and already what the displayed profile uses via rf_lyrs) instead of the binned c_cfpct_intpl_grp, in all three candidate-assembly paths (global, US main, US OSD merge). The pit stays a class midpoint (getCF_fromClass) since the user only reports a class; that midpoint is on the same raw-% scale, so the feature bounds and the #377 range floor are unchanged. Sand/clay are left binned (they derive from the 2-D texture class, not a 1-D threshold set — separate question). Result-affecting: horizon/US/global snapshots must be regenerated and the change validated on the bulk-test suite. Version bumped 2.2.0 -> 2.3.0 so clients flush cached matches. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
c9b168c switched candidate rock fragment from the binned class midpoint to the raw %, which is result-affecting, but landed without regenerating the unit snapshots. Regenerate + verify (deterministic) against the pinned soil-id-db. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The site (slope/elevation) score was over-weighted relative to the horizon score. Tuning against the US bulk test and a confirmation real run give +0.34 pt top-1 (61.56 -> 61.90 on the raw-% base); top-3/top-5 flat, recall unchanged (pure reorder — pulls correct soils from rank 3/4 up to rank 1). See SOILID_TUNING.md. Result-affecting: version 2.3.0 -> 2.4.0; snapshots regenerated + verified. Note: the rfv/LAB-b horizon down-weights an earlier optimization found are NOT applied — they were compensating for the class-binning noise that the raw-% rock fragment change (2.3.0) already fixed; re-optimizing on the raw-% base leaves them neutral, so site_wt is the only surviving lever. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The uniform-depth change (in the consistency commit) was net ~0 on the raw-% base: it helps surface-heavy pits (+~1 pt) but slightly hurts deep pits (-0.07), which cancel. De-weighting the disturbed topsoil (0.2x) is also the more defensible model for the common deep-data case, so restore it. dis_max=1.0 (the genuinely-more- correct part of that commit) is kept. The principled improvement would be data-presence-aware weighting (de-weight the surface only when there's deeper data), not a constant -- filed under "Potential future changes" in JOHANNES_NOTES_2026, low priority since the effect is ~0. Snapshots regenerated + verified. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…CM 50 -> 35 Leptosols/lithosols/rendzinas/rankers are defined at ~25-30 cm to rock; 50 cm was a provisional placeholder (#375). 35 cm is closer to the taxonomic definition with a small margin. Result-affecting for pits whose effective bedrock is 35-50 cm and have a shallow-soil candidate (one global snapshot moved: a shallow candidate's horizon score is now demoted). Version 2.4.0 -> 2.5.0; snapshots regenerated + verified. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…cally (#398) process_distance_scores sorted rows by cond_prob desc but then grouped with sort=True, which re-ordered the groups alphabetically by compname before the top-12 truncation — so high-probability soils were silently dropped in favor of lower-probability ones whose names sort earlier (whenever a buffer had >12 groups). Affects both US and global candidate selection. Fix: groupby(sort=False) so the top-12 follows the cond_prob-desc order. Also fixes the following no-op loop that rebound a local `group` (its distance sort + min_dist were discarded): write the transformed group back into the list. This was harmless in practice (both are recomputed downstream) but is now correct. Adds regression tests. Result-affecting (8 snapshots moved: 4 US + 4 global, all locations with >12 groups); version 2.5.0 -> 2.6.0; snapshots regenerated + verified. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
johannesparty
force-pushed
the
feat/soilid-site-weight
branch
from
September 25, 2026 19:05
3d51128 to
6fd3ddb
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stacked on #396 (
feat/explain-html-renderer). A batch of tuning-driven scoringchanges, each validated against the US bulk test on the NAS (see
SOILID_TUNING.md/JOHANNES_NOTES_2026.mdfor methodology + numbers).Changes (net top-1, real US bulk runs)
c9b168c)c_cfpct_intplinstead of the 5-bucket class midpoint — removes the class-boundary cliff + within-class flattening. Pit still a class midpoint (user only reports a class).8ac63bf)dis_maxnanmax → 1.0 (6b3c9c8)missing/recall unchanged throughout — these reorder candidates, they don'tchange which soils are retrieved. Version bumped 2.2.0 → 2.4.0; unit snapshots
regenerated + verified deterministic against the pinned soil-id-db.
Not included / notes
3fa7b0c): net ~0 onthe bulk test (helps surface-heavy pits, slightly hurts deep pits) and
de-weighting the disturbed topsoil is the more defensible model — so the surface
weight stays at its original 0.2 (this cancels in the PR's net diff).
adopted — the raw-% change superseded them (re-optimizing on the raw-% base
leaves them neutral).
🤖 Generated with Claude Code