Farming Patterns

Farming Patterns is an experimental, evidence-first view of sampled hero routes through calibrated neutral-camp zones. It answers one question:

Which camp-local route segments were observed, and how much replay evidence supports treating each segment as farming rather than transit?

The route builder is available as gem.build_farming_routes(match). The same records feed the HTML report and the flat DataFrame exports.

Current scope

The analysis reconstructs camp-local routes without scoring them. It uses facts already present on ParsedMatch:

• sampled positions from ParsedPlayer.position_log;
• calibrated camp geometry from camp_zones.json;
• neutral damage and deaths from the combat log;
• sampled total XP and total-earned-gold endpoints.

It does not change replay parsing or add extractor state. Samples before ParsedMatch.game_start_tick are excluded, so pre-horn movement does not become a farming segment.

gem 0.13 removed the comparative "context" and tags that 0.12 attached to each segment. Join the facts you need yourself: gem.region_of, the wards, towers and Aegis events on ParsedMatch, and the players' positions.

Public API

import gem

match = gem.parse("match.dem")
routes = gem.build_farming_routes(match)

for route in routes:
    print(route.player_id, route.status, route.status_reasons)
    for segment in route.segments:
        print(
            segment.camp_id,
            segment.start_tick,
            segment.end_tick,
            segment.evidence_strength.value,
            segment.evidence_reasons,
            segment.evidence_gaps,
        )

The public records are:

• FarmingRoute: one result per parsed player;
• FarmingRoutePoint: one sampled position with deterministic camp membership;
• FarmingRouteSegment: one camp-local window and its support evidence;
• FarmingRouteConfig: the inspectable reconstruction thresholds;
• FarmingCampZone: normalized camp geometry;
• FarmingEvidenceStrength and FarmingBoundaryReason: stable string enums.

Deterministic route reconstruction

The default configuration is:

All time conversions use 30 replay ticks per second.

Zone selection

For each position sample:

1. If the hero was already assigned to a camp and remains inside that camp's exit-hysteresis geometry, retain the current camp.
2. Otherwise, collect every camp whose entry geometry contains the sample.
3. Resolve overlap by normalized distance to the camp geometry.
4. Resolve an exact tie by ascending camp ID.

This avoids catalog-order dependence and prevents boundary jitter from rapidly switching between adjacent camps.

Boundaries

Every segment exposes a start_reason and end_reason:

• zone_entry / zone_exit;
• camp_change;
• sample_gap;
• large_jump;
• log_end.

Sample gaps and large jumps are hard discontinuities. They never merge back together. A short exit followed by re-entry into the same camp may merge, but the merged record keeps micro_exit_merged=True and retains the intervening out-of-zone point. in_zone_sample_count therefore remains distinct from sample_count.

Evidence attribution

Each segment retains:

• exact start/end ticks and duration;
• camp ID and camp_type;
• every sampled route point;
• total and in-zone sample counts;
• sampled-window coverage and largest supported gap;
• whether a micro-exit was merged;
• neutral deaths and neutral damage;
• cumulative total-earned-XP and total-earned-gold deltas when both endpoint samples are fresh;
• evidence reasons and evidence gaps.

Neutral events must fall inside the segment window, target an npc_dota_neutral* unit, and be attributable through either attacker_name or damage_source_name. When an event has coordinates, those coordinates must also lie inside the segment's camp zone. An event without coordinates remains attributable by the bounded time window and source.

XP comes from ParsedPlayer.total_earned_xp_t, not the level-local xp_t counter that resets on level-up. XP and gold stay None when fresh endpoints are unavailable. They are never converted to zero. The route status becomes partial when a segment has these resource gaps.

Evidence strength

Evidence strength is deliberately small and composable:

The labels do not assert that the hero cleared the camp, earned every resource from that camp, or intended to farm. Fresh XP/gold deltas remain visible as window facts, but a brief touch is not promoted solely because passive or off-zone resources changed.

Camp facts on each segment

Each segment carries its camp's owner team, lane affinity and area, the catalog, geometry and topology versions, and the contiguous distance travelled.

Availability

Each FarmingRoute has one of three statuses:

• complete: route reconstruction succeeded and every segment has fresh resource endpoints;
• partial: route reconstruction succeeded, but one or more segment resource windows are unavailable;
• unavailable: camp geometry or player position samples are unavailable.

The corresponding status_reasons and per-segment evidence_gaps are public. The bundled catalog reports version 4, 7.41 map geometry, and 7.41 camp-family/topology annotations. Each zone is centred on its camp's CDOTA_NeutralSpawner entity, which sits at the same position in every 7.41 fixture replay; neutral creeps spawn within about 100 units of it. Geometry and topology provenance remain separate fields.

DataFrames

gem.results.dataframes.build_dataframes(match, include=["analysis"]) (or gem.parse_to_dataframe(path, include=["analysis"])) adds three stable flat tables:

• farming_routes: player-level availability, catalog metadata, and counts;
• farming_route_segments: boundaries, support facts, strength, camp facts, provenance, and gaps;
• farming_route_points: sampled path, selected camp, base-zone membership, boundary_before, and segment membership where applicable.

String enums are exported as their raw values. Lists of reasons/gaps use semicolon-delimited strings, matching the other flat analysis exports.

Report behavior

The report's Farming tab shows the route and timeline for each team's cores (see Match Reports). Each segment row lists the camp, its type, the visit's duration, the hero's neutral kills inside the camp zone, and the gold and XP the hero earned during the visit. Evidence strength stays available in Python and the DataFrame exports but is not rendered.

The playback trail uses the route builder's point-to-camp assignments rather than recalculating geometry in the report. This keeps Python, DataFrame, and HTML behavior aligned.

Calibration

The farming-route corpus records factual counts per fixture replay (segments, points, evidence strength, phase, camp area) so that a change to route reconstruction shows up as a difference. It makes no judgment about whether a route was strategically correct. See Farming Route Calibration.

Source map

• src/gem/analysis/farming.py
• src/gem/data/camp_zones.json
• src/gem/results/dataframes.py
• src/gem/reports/sections/vision.py