Roshan Conversion Calibration

This page records the reproducible validation work behind the provisional-v1 Roshan conversion ruleset. It separates replay facts from tag observations: the corpus asserts attribution, lifecycle, boundaries, fight association, and missing-data state, but never declares a Roshan strategically “good” or “bad.”

Reproduce the corpus

The committed factual expectations live in tests/fixtures/opendota/roshan_conversion_corpus.json. Full .dem files stay 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_roshan_conversion_corpus.py --check

The audit parses the replay stream with gem. It does not substitute the OpenDota JSON snapshot for gem's protocol and entity evidence.

Real-match coverage

The September 2026 audit parsed eight existing local fixtures plus the dedicated Aegis-denial fixture, spanning 31 Roshan conversions and one short match with no Roshan. The committed check selects representative facts from six matches:

Match	Factual role
8868259993	Short game with no Roshan
8860187335	Three bounded windows and one in-game pause; natural expiry, inferred consumption, and a stolen pickup still held when the Ancient fell
8855188139	Multiple inferred consumptions and a late window that closes the game
8856501050	Seven Roshans across three pauses, including an expired window with no associated fight and a final window cut short by the game ending
8822593932	A fight already underway at acquisition and a final window without economy series
8974053011	Observed Aegis denial followed by a later normal pickup and inferred consumption

The local replay set did not contain every rare failure mode. Missing pickup, unresolved killers, unknown structure/Tormentor attribution, allied structure denies, summon/source fallback, and threshold edges therefore have targeted deterministic unit tests. This distinction is explicit in the corpus metadata rather than representing synthetic cases as real-match facts.

Attribution and lifecycle findings

• Source 2 attacker_team was present for every audited Roshan kill and is now the preferred source. Player and unique roster-name fallbacks remain for older or incomplete data.
• The real denial fixture produces aegis_fate_source="denial_event"; its later holder death produces holder_death_inference. Natural five-minute endings produce nominal_expiry; an Aegis still held when the Ancient falls produces game_end_boundary.
• Item-entity deletion cannot reliably distinguish pickup, consumption, expiry, transfer/drop, and cleanup. It is not used to overrule the bounded death inference.
• All audited objective and Tormentor events with known protocol/team evidence remained attributed; unknown targeted cases remain counted and uncredited.
• Each conversion ends no later than one tick before the next Roshan. Claimed fight indexes are unique across consecutive windows.

Tag observations

The 31 audited conversions produced these default-tag frequencies:

Tag	Observed count
fight_advantage	8
objective_gain	21
resource_gain	20
territorial_expansion	0
vision_expansion	18
tormentor_secured	6
game_closing	5
counter_conversion	7
no tag	0

These counts were re-audited after the game-clock correction. The first audit ended every window at the last recorded tick, which can run many minutes past the Ancient falling, so game_closing could never fire on a real replay (0 hits) and final windows reported missing late evidence. It also mapped minute-level net worth/XP samples to ticks without subtracting pauses, which shifted resource windows after any pause. Correcting both moved game_closing from 0 to 5, counter_conversion from 3 to 7, resource_gain from 22 to 20, and no-tag windows from 2 to 0. The thresholds themselves are unchanged.

The halves were re-audited when region_of moved from a band along the x = y diagonal (which was really the mid lane) to the river traced on the 7.41 map. Mid-lane wards and positions now count towards a half, and the river and Roshan pits towards neither. That moved vision_expansion from 15 to 18; every other tag count, including the zero for territorial_expansion, is unchanged.

These are sensitivity observations, not precision/recall labels. There is no defensible subjective ground truth in the fixture, so the audit cannot claim a false-positive rate by treating analyst opinion as fact. The defaults are therefore unchanged and remain explicitly provisional-v1. In particular, the zero hits for territorial expansion are recorded for future review rather than “fixed” by lowering thresholds against this small sample.

RoshTagThresholds makes every boundary inspectable. Unit tests exercise exact threshold transitions and counter-conversion's two-dimension minimum. Raw values remain the preferred interface regardless of which tags fire.

Territory sensitivity

RoshTerritoryConfig makes cell size, bucket width, maximum sample gap, occupancy requirements, player-time completeness, and depth percentile reproducible inputs.

The validation suite establishes:

• 70% player-time completeness is inclusive; 69.99% remains unavailable.
• stationary occupancy is invariant when the bucket width changes from 30 to 60 seconds.
• changing cell size changes the absolute sampled area, as expected, which is why coverage remains a raw geometric measure rather than “true control.”
• weighted p90 depth ignores a small deep-position outlier that a maximum would promote.
• gaps longer than ten seconds are not interpolated.

The default remains 600-unit cells, 30-second buckets, a ten-second maximum sample gap, 70% completeness, and weighted p90 depth. Lane or objective topology was not added because this corpus does not demonstrate that the added interpretation would be more reliable than the current transparent geometry.

Review policy

A future threshold change should update the ruleset name, the boundary tests, this page, and the corpus observations in the same pull request. A larger human-reviewed study may add strategic labels separately, but must not replace the factual replay expectations used for regression testing.