Farming Route Calibration

This page records the reproducible regression corpus behind farming-route reconstruction. It asserts route, evidence-strength, phase and topology counts per fixture replay, and never labels a route as good, bad, safe, or forced.

gem 0.12's comparative segment context and its tags were removed in 0.13, and with them this page's tag observations.

Reproduce the corpus

The committed summaries live in tests/fixtures/opendota/farming_routes_corpus.json. Full .dem files remain gitignored and are synchronized from tests/fixtures/opendota/manifest.json.

uv run python scripts/sync_opendota_fixtures.py --all-active
uv run python scripts/audit_farming_routes.py \
  tests/fixtures/opendota/8822520406.dem \
  tests/fixtures/opendota/8855188139.dem \
  tests/fixtures/opendota/8855242704.dem \
  tests/fixtures/opendota/8856501050.dem \
  tests/fixtures/opendota/8860187335.dem \
  tests/fixtures/opendota/8868259993.dem \
  tests/fixtures/opendota/8974053011.dem \
  --check tests/fixtures/opendota/farming_routes_corpus.json

The command parses gem's replay stream and rebuilds farming routes with the public defaults. OpenDota JSON snapshots are not substituted for positional or resource evidence.

Real-match coverage

The September 2026 audit parsed seven active fixtures: one short performance baseline, three feature-regression matches, and short, medium, and long TI2026 matches. With the version 4 camp catalog (zones centred on the camp spawners, and every camp's type and owner checked against the replays) it covers 70 player routes, 5,814 camp-local segments, and 206,950 sampled route points.

Match	Segments	Phases	Calibration role
`8822520406`	352	early, mid	Short performance baseline
`8855188139`	658	early, mid, late	Roshan, purchases, and inventory regression
`8855242704`	1,266	early, mid, late	Long final-interval and XP regression
`8856501050`	1,654	early, mid, late	93-minute stress fixture
`8860187335`	910	early, mid, late	Medium extended fixture with multiple Aegis windows
`8868259993`	312	early, mid	Canonical short integration fixture
`8974053011`	662	early, mid, late	Aegis-denial lifecycle regression

Every player route is included rather than selecting only high-farm heroes. That captures the different movement patterns of professional carries, mids, offlaners, and supports without pretending the replay's coarse lane_role number is an authoritative strategic-role label.

All four topology areas were exercised: 3,013 jungle, 1,207 triangle, 767 river, and 827 flooded-area segments. The phase split was 912 early, 2,413 mid, and 2,489 late segments. Lane-to-jungle movement and brief crossings are therefore present in the factual segment totals rather than hand-picked as subjective examples.

Evidence-strength observations

Evidence strength	Observed segments
`strong_farm_evidence`	2,007
`weak_farm_evidence`	1,319
`transit_like`	2,488

These counts demonstrate that the reconstruction does not promote every camp touch to farming. They are sensitivity observations, not precision or recall: the corpus has no defensible ground-truth field for player intent or complete camp clears.

Threshold and topology policy

A change to FarmingRouteConfig or the camp catalog should update the boundary tests, the corpus summaries and this page in the same pull request. Counts alone are not a reason to tune a threshold toward a preferred story.

Camp ownership, lane affinity, and area are explicit 7.41 topology annotations. The zone centres are the 7.41 replays' camp spawner positions (catalog version 4); the per-type radii are unchanged from the hand-drawn catalog. Geometry and topology versions stay separate fields.

Future calibration should add real fixtures when they expose a new factual boundary or missing-data condition. Human-reviewed strategy labels may be studied separately, but must not replace the evidence-first regression corpus.