Roman Belcher
Machine estimate. These numbers come from the strike detector at its balanced operating point — not a human. Expect a ~9–15% undercount, and counts are standup-scoped where segments exist. No conditional tells or defensive reads are shown here — those need a verified sweep. How we measure →
Machine estimate: 64% of the strikes the detector saw from Roman Belcher were punches; the other 36% were kicks.
Watch the Film Study▾
greyed: no playable video registered · times index the registered upload
Machine estimate — NOT human-verified. Counts are the strike detector accepted at its balanced operating point (punch pcal ≥ 0.437, kick ≥ 0.540); a high-recall surfacer, not a log.
Undercount ~9–15%: the hand log itself misses an estimated ~8.8% of strikes and machine recall is separate and unverified here — every count is a floor.
Balanced operating point was derived on the pre-CatBoost pcal scale; the crowned model shifted that scale, so these counts are provisional pending recalibration.
standup-scoped: in-round events inside standup spans only
engine provenance ▸
Machine estimate: 43% of Roman Belcher's detected output comes in bursts — strikes less than 1 second apart.
Watch the Film Study▾
greyed: no playable video registered · times index the registered upload
Machine estimate — NOT human-verified. Counts are the strike detector accepted at its balanced operating point (punch pcal ≥ 0.437, kick ≥ 0.540); a high-recall surfacer, not a log.
Undercount ~9–15%: the hand log itself misses an estimated ~8.8% of strikes and machine recall is separate and unverified here — every count is a floor.
Balanced operating point was derived on the pre-CatBoost pcal scale; the crowned model shifted that scale, so these counts are provisional pending recalibration.
standup-scoped: in-round events inside standup spans only
burst = inter-strike interval < 1.0s
engine provenance ▸
Machine estimate. These numbers come from the strike detector at its balanced operating point — not a human. Expect a ~9–15% undercount, and counts are standup-scoped where segments exist. No conditional tells or defensive reads are shown here — those need a verified sweep. How we measure →
Machine estimate — vs James Barney
FIRST READMachine estimate: 43% of Roman Belcher's detected output comes in bursts — strikes less than 1 second apart.
Watch the Film Study▾
greyed: no playable video registered · times index the registered upload
Machine estimate — NOT human-verified. Counts are the strike detector accepted at its balanced operating point (punch pcal ≥ 0.437, kick ≥ 0.540); a high-recall surfacer, not a log.
Undercount ~9–15%: the hand log itself misses an estimated ~8.8% of strikes and machine recall is separate and unverified here — every count is a floor.
Balanced operating point was derived on the pre-CatBoost pcal scale; the crowned model shifted that scale, so these counts are provisional pending recalibration.
standup-scoped: in-round events inside standup spans only
burst = inter-strike interval < 1.0s
engine provenance ▸
Machine estimate: Roman Belcher throws 36% of his detected strikes from the left side. Sides are physical left/right — no stance asset exists for this fight.
Watch the Film Study▾
Machine estimate — NOT human-verified. Counts are the strike detector accepted at its balanced operating point (punch pcal ≥ 0.437, kick ≥ 0.540); a high-recall surfacer, not a log.
Undercount ~9–15%: the hand log itself misses an estimated ~8.8% of strikes and machine recall is separate and unverified here — every count is a floor.
Balanced operating point was derived on the pre-CatBoost pcal scale; the crowned model shifted that scale, so these counts are provisional pending recalibration.
standup-scoped: in-round events inside standup spans only
no stance asset for this fight — hands are physical left/right, not lead/rear (a per-fight stance asset would unlock lead/rear)
engine provenance ▸
Machine estimate: 64% of the strikes the detector saw from Roman Belcher were punches; the other 36% were kicks.
Watch the Film Study▾
greyed: no playable video registered · times index the registered upload
Machine estimate — NOT human-verified. Counts are the strike detector accepted at its balanced operating point (punch pcal ≥ 0.437, kick ≥ 0.540); a high-recall surfacer, not a log.
Undercount ~9–15%: the hand log itself misses an estimated ~8.8% of strikes and machine recall is separate and unverified here — every count is a floor.
Balanced operating point was derived on the pre-CatBoost pcal scale; the crowned model shifted that scale, so these counts are provisional pending recalibration.
standup-scoped: in-round events inside standup spans only
engine provenance ▸
Machine estimate: the detector shows Roman Belcher's pace fading — 3.4 strikes a minute in round 1, 2.5 in round 5.
Watch the Film Study▾
greyed: no playable video registered · times index the registered upload
Machine estimate — NOT human-verified. Counts are the strike detector accepted at its balanced operating point (punch pcal ≥ 0.437, kick ≥ 0.540); a high-recall surfacer, not a log.
Undercount ~9–15%: the hand log itself misses an estimated ~8.8% of strikes and machine recall is separate and unverified here — every count is a floor.
Balanced operating point was derived on the pre-CatBoost pcal scale; the crowned model shifted that scale, so these counts are provisional pending recalibration.
standup-scoped: in-round events inside standup spans only
per-round counts sit on a Poisson floor of ±18–25% — read trends, not precision
engine provenance ▸
Machine estimate: Roman Belcher averaged 3.6 strikes per minute of standup time vs James Barney.
Watch the Film Study▾
greyed: no playable video registered · times index the registered upload
Machine estimate — NOT human-verified. Counts are the strike detector accepted at its balanced operating point (punch pcal ≥ 0.437, kick ≥ 0.540); a high-recall surfacer, not a log.
Undercount ~9–15%: the hand log itself misses an estimated ~8.8% of strikes and machine recall is separate and unverified here — every count is a floor.
Balanced operating point was derived on the pre-CatBoost pcal scale; the crowned model shifted that scale, so these counts are provisional pending recalibration.
standup-scoped: in-round events inside standup spans only
rate is per standup-minute (exposure-normalized)
engine provenance ▸
The detector counted 28 strikes vs James Barney — 18 punches, 10 kicks. Machine estimate, not human-verified.
Watch the Film Study▾
greyed: no playable video registered · times index the registered upload
Machine estimate — NOT human-verified. Counts are the strike detector accepted at its balanced operating point (punch pcal ≥ 0.437, kick ≥ 0.540); a high-recall surfacer, not a log.
Undercount ~9–15%: the hand log itself misses an estimated ~8.8% of strikes and machine recall is separate and unverified here — every count is a floor.
Balanced operating point was derived on the pre-CatBoost pcal scale; the crowned model shifted that scale, so these counts are provisional pending recalibration.
standup-scoped: in-round events inside standup spans only
engine provenance ▸
Machine estimate — vs Griffen Leombruno
FIRST READonly 9 machine-detected strikes — 25 needed
unlock: a longer processed fight unlocks this
only 9 machine-detected strikes (< 25)
unlock: a longer processed fight or verified sweep unlocks this
only 9 machine-detected strikes — 25 needed
unlock: a longer processed fight or verified sweep unlocks this
Machine estimate: the detector shows Roman Belcher throwing more as the fight goes on — 0.7 strikes a minute in round 1, 2.6 in round 3.
Watch the Film Study▾
greyed: no playable video registered · times index the registered upload
Machine estimate — NOT human-verified. Counts are the strike detector accepted at its balanced operating point (punch pcal ≥ 0.437, kick ≥ 0.540); a high-recall surfacer, not a log.
Undercount ~9–15%: the hand log itself misses an estimated ~8.8% of strikes and machine recall is separate and unverified here — every count is a floor.
Balanced operating point was derived on the pre-CatBoost pcal scale; the crowned model shifted that scale, so these counts are provisional pending recalibration.
standup-scoped: in-round events inside standup spans only
per-round counts sit on a Poisson floor of ±18–25% — read trends, not precision
engine provenance ▸
only 9 machine-detected strikes over 0.8 standup-minutes — 25 needed
unlock: a longer processed fight unlocks a stable pace
The detector counted 9 strikes vs Griffen Leombruno — 6 punches, 3 kicks. Machine estimate, not human-verified.
Watch the Film Study▾
greyed: no playable video registered · times index the registered upload
Machine estimate — NOT human-verified. Counts are the strike detector accepted at its balanced operating point (punch pcal ≥ 0.437, kick ≥ 0.540); a high-recall surfacer, not a log.
Undercount ~9–15%: the hand log itself misses an estimated ~8.8% of strikes and machine recall is separate and unverified here — every count is a floor.
Balanced operating point was derived on the pre-CatBoost pcal scale; the crowned model shifted that scale, so these counts are provisional pending recalibration.
standup-scoped: in-round events inside standup spans only
engine provenance ▸
Machine estimate — vs Chris Perciavalle
FIRST READMachine estimate: 77% of Roman Belcher's detected output comes in bursts — strikes less than 1 second apart.
Watch the Film Study▾
greyed: no playable video registered · times index the registered upload
Machine estimate — NOT human-verified. Counts are the strike detector accepted at its balanced operating point (punch pcal ≥ 0.437, kick ≥ 0.540); a high-recall surfacer, not a log.
Undercount ~9–15%: the hand log itself misses an estimated ~8.8% of strikes and machine recall is separate and unverified here — every count is a floor.
Balanced operating point was derived on the pre-CatBoost pcal scale; the crowned model shifted that scale, so these counts are provisional pending recalibration.
standup-scoped: in-round events inside standup spans only
burst = inter-strike interval < 1.0s
engine provenance ▸
Machine estimate: Roman Belcher throws 47% of his detected strikes from the left side. Sides are physical left/right — no stance asset exists for this fight.
Watch the Film Study▾
Machine estimate — NOT human-verified. Counts are the strike detector accepted at its balanced operating point (punch pcal ≥ 0.437, kick ≥ 0.540); a high-recall surfacer, not a log.
Undercount ~9–15%: the hand log itself misses an estimated ~8.8% of strikes and machine recall is separate and unverified here — every count is a floor.
Balanced operating point was derived on the pre-CatBoost pcal scale; the crowned model shifted that scale, so these counts are provisional pending recalibration.
standup-scoped: in-round events inside standup spans only
no stance asset for this fight — hands are physical left/right, not lead/rear (a per-fight stance asset would unlock lead/rear)
engine provenance ▸
Machine estimate: 100% of the strikes the detector saw from Roman Belcher were punches; the other 0% were kicks.
Watch the Film Study▾
greyed: no playable video registered · times index the registered upload
Machine estimate — NOT human-verified. Counts are the strike detector accepted at its balanced operating point (punch pcal ≥ 0.437, kick ≥ 0.540); a high-recall surfacer, not a log.
Undercount ~9–15%: the hand log itself misses an estimated ~8.8% of strikes and machine recall is separate and unverified here — every count is a floor.
Balanced operating point was derived on the pre-CatBoost pcal scale; the crowned model shifted that scale, so these counts are provisional pending recalibration.
standup-scoped: in-round events inside standup spans only
engine provenance ▸
Machine estimate: the detector shows Roman Belcher's pace holding steady — 14.6 strikes a minute in round 1, 13.7 in round 3.
Watch the Film Study▾
greyed: no playable video registered · times index the registered upload
Machine estimate — NOT human-verified. Counts are the strike detector accepted at its balanced operating point (punch pcal ≥ 0.437, kick ≥ 0.540); a high-recall surfacer, not a log.
Undercount ~9–15%: the hand log itself misses an estimated ~8.8% of strikes and machine recall is separate and unverified here — every count is a floor.
Balanced operating point was derived on the pre-CatBoost pcal scale; the crowned model shifted that scale, so these counts are provisional pending recalibration.
standup-scoped: in-round events inside standup spans only
per-round counts sit on a Poisson floor of ±18–25% — read trends, not precision
engine provenance ▸
Machine estimate: Roman Belcher averaged 15.9 strikes per minute of standup time vs Chris Perciavalle.
Watch the Film Study▾
greyed: no playable video registered · times index the registered upload
Machine estimate — NOT human-verified. Counts are the strike detector accepted at its balanced operating point (punch pcal ≥ 0.437, kick ≥ 0.540); a high-recall surfacer, not a log.
Undercount ~9–15%: the hand log itself misses an estimated ~8.8% of strikes and machine recall is separate and unverified here — every count is a floor.
Balanced operating point was derived on the pre-CatBoost pcal scale; the crowned model shifted that scale, so these counts are provisional pending recalibration.
standup-scoped: in-round events inside standup spans only
rate is per standup-minute (exposure-normalized)
engine provenance ▸
The detector counted 129 strikes vs Chris Perciavalle — 129 punches, 0 kicks. Machine estimate, not human-verified.
Watch the Film Study▾
greyed: no playable video registered · times index the registered upload
Machine estimate — NOT human-verified. Counts are the strike detector accepted at its balanced operating point (punch pcal ≥ 0.437, kick ≥ 0.540); a high-recall surfacer, not a log.
Undercount ~9–15%: the hand log itself misses an estimated ~8.8% of strikes and machine recall is separate and unverified here — every count is a floor.
Balanced operating point was derived on the pre-CatBoost pcal scale; the crowned model shifted that scale, so these counts are provisional pending recalibration.
standup-scoped: in-round events inside standup spans only
engine provenance ▸
Fight IQ is the fragile, high-value layer — conditional tells, response windows, combinations, defensive reads. It needs a near-perfect event stream, so it stays locked on a machine sheet. Nothing below is computed from the auto-log.
What he throws after X
conditional next-strike tables need a verified event sequence
Signature combinations
combination frequencies need a near-perfect, verified strike order
Counter-windows
who-threw-then-who timing needs a verified two-sided event stream
Defensive tells (guard-drop)
guard telemetry needs verified trigger windows
A scout pack includes TAPE's verified sweep within ~5 days — a sweep is a human pass over the machine log, confirming every strike against the tape. Can't wait? Sweep it yourself, or with your coach, and these reads turn on with receipts.
Verified verdict profile
Operator verdict-ledger evidence, kept separate from temporal tendencies. Complete reviews may support totals and gated rates; partial reviews remain labeled samples; reviews below the coverage floor expose only their withholding reason.
The three frozen verdict gates
- Per-fight mix shares require n ≥ 25 reviewed real attempts; below it, counts only.
- Complete-fight pooled mix shares and own rates require n ≥ 25 real attempts.
- Defense clean-absorb rates require a complete review and n ≥ 25 opponent real attempts.
- Partial reviews are non-random samples: counts and eligible sample mixes are not fight totals or accuracy.
- Physical side means left/right, not lead/rear stance. Jab/cross for swept strikes is stance-INFERRED from pose: a windowed ankle-vote at the strike instant when decisive (94.3% vs hand-log truth), else the fight-level lead; strikes where both abstain stay "straight" rather than guess.
- A single reviewer produced these verdicts; unresolved attribution/validity verdicts are a floor on missingness.
Coverage & provenance
machine estimate — no human labels yet; descriptive stats only
machine estimate — no human labels yet; descriptive stats only
machine estimate — no human labels yet; descriptive stats only
Not measurable from broadcast
the single-feed footage can’t support these — they are not a locked upgrade, they simply aren’t recoverable. Saying so is the product.
Landed / grazed / missed
outcome carries ~no machine signal at this capture and is unreliable even by eye from a single broadcast feed
Fine punch subtype (jab / cross / hook)
capped by 72% annotator self-agreement and the single-feed capture wall
Clinch & ground interior
the corpus is standup-scoped — a single broadcast feed can't resolve grappling exchanges
Feint intent
intent-without-contact is not recoverable from broadcast video
Known limits
- single-annotator ground truth; measured undercount ~8.8% on the deepest-swept fight — global estimate 11–15%; every count is a floor, not an exaggeration
- tendency numbers are standup-scoped and in-round; clinch/ground exchanges are not measured
- stance (lead/rear) is a per-fight constant from a pose asset; a mid-fight stance switch would mislabel lead/rear symmetrically
- per-round counts sit on a Poisson floor of roughly ±18–25%; product language is trends and conditionals, never single-round precision
Every fight on this fighter's file — the event, the result, and how deep the verified data goes.
Want more of their fights processed? Start a pack or request a fight.