HeadOn (no-lead) vs Pattern LIVE at long range: clean negative, offline ruler killed

2 arms x 15 runs x 7 rounds, one frozen binary from HEAD a82c864, real DrussGT,
server-side events sidecar. Shipped rack is onlyPattern, so control=Pattern-only
and headon=HeadOn-only (TR_RACK_PATTERN=off TR_RACK_HEADON=both).

  arm      dmg/run  dmgtk/run  round wins  shots/run
  control      279        211     48/105       785
  headon        14        228      0/105       580

Round wins and dmg/run both separate at p<0.0001 (MC permutation, se 0.0000),
~7x the damage MDE (35.8). Per range band (pooled, 15 runs):
  300-450: Pattern 12.3% (4590 shots) vs HeadOn 0.6% (3701)  p<0.0001, MDE 2.0pp
  450+   : Pattern  9.2% (6671)       vs HeadOn 0.4% (4177)  p<0.0001, MDE 1.1pp
HeadOn loses EVERY long-range band by 20-23x, so the whole-battle loss is not a
close-range artefact.

The offline ruler (prediction_quality_results.txt) predicted the opposite: HeadOn
meanAbs 14.61 vs Pattern 17.53 at 300-450 and 12.33 vs 16.19 at 450+, hitProxy
.105/.104 and .098/.077 (+27%). That is an open-loop replay of a FIXED enemy
track, so it cannot see that a different bullet makes the surfer dodge
differently; live, the static gun does not lead at all.

TR_PATTERN_RAD_SCALE arms were skipped: applyRadial scales aim DISTANCE along an
unchanged bearing, so it cannot express 'less lead' (bearing is what firing uses).
HeadOn confirmed to ignore bulletSpeed (head_on.nim:9), liveness OK 15/15.

Adds the range-band analyzer tools/ab/ab_range_bands.py (reuses the lead-capture
Run alignment) and the captured fixtures. Does not touch bitbrain_gun.nim /
bitbrain_campaign.md (job-100).
This commit is contained in:
2026-09-24 23:52:49 +02:00
parent a82c864c60
commit 140fe2519a
6 changed files with 583 additions and 0 deletions
+182
View File
@@ -0,0 +1,182 @@
#!/usr/bin/env python3
"""ab_range_bands.py — hit rate BY RANGE BAND per arm, from an ab_run.sh session.
python3 tools/ab/ab_range_bands.py <session_dir> [--reference ARM]
Why this exists: ab_analyze.py judges damage/run and round wins over the whole
battle, but the HeadOn-vs-Pattern claim is specifically about LONG RANGE. A
whole-battle win that comes only from close range would NOT support it, so the
range breakdown is the load-bearing view. ab_analyze.py does not emit it.
Method: reuse `Run` from common_libs/tests/analyze_lead_capture_by_range.py,
which aligns each of OUR fire events to the capture tick (owner side resolved
per run, round-startTick + local-tick search) and gives the fire-tick range
(shooter->target distance) plus the server-resolved outcome of that bullet.
A band's hit rate is then hits/shots over that band's shots, pooled across the
arm's runs. A two-sided permutation test on the PER-RUN band rates (only runs
with shots in the band) is printed alongside, so a pooled difference can be
told apart from run-to-run noise; the MDE for those n is printed too.
Bands match the offline ruler (prediction_quality.nim) and the lead-capture
report: 0-100, 100-200, 200-300, 300-450, 450+ px.
"""
import argparse
import collections
import math
import os
import random
import sys
_HERE = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, os.path.join(_HERE, "..", "..", "common_libs", "tests"))
from analyze_lead_capture_by_range import Run, BAND_LABELS # noqa: E402
Z_ALPHA_POWER = 1.959963984540054 + 0.8416212335729143
MC_SEED = 0x5EED5EED
MC_DRAWS = 200_000
def load_arm(session_dir, arm):
armdir = os.path.join(session_dir, arm)
runs = []
for fn in sorted(os.listdir(armdir)):
if not fn.endswith(".jsonl") or fn.endswith(".events.jsonl"):
continue
cap = os.path.join(armdir, fn)
ev, rj = cap[:-6] + ".events.jsonl", cap + ".rounds.json"
if os.path.exists(ev) and os.path.exists(rj):
runs.append((fn, cap, ev, rj))
out = {}
for fn, cap, ev, rj in runs:
try:
r = Run(cap, ev, rj)
except Exception as e: # noqa: BLE001
print(f" WARN {arm}/{fn}: {e}", file=sys.stderr)
continue
shots = list(r.shots())
out[fn] = shots
return out
def perm_test(xa, xb):
"""Two-sided permutation on the difference of means (exact if small)."""
na, nb = len(xa), len(xb)
if na == 0 or nb == 0:
return None
import itertools
obs = abs(sum(xa) / na - sum(xb) / nb)
pooled = list(xa) + list(xb)
n = na + nb
total = sum(pooled)
ncomb = math.comb(n, na)
if ncomb <= 20_000_000:
cnt = 0
for combo in itertools.combinations(range(n), na):
sa = sum(pooled[i] for i in combo)
if abs(sa / na - (total - sa) / nb) >= obs - 1e-9:
cnt += 1
return obs, cnt / ncomb, "exact", 0.0
rng = random.Random(MC_SEED)
B = MC_DRAWS
cnt = 0
for _ in range(B):
sa = 0
for i in rng.sample(range(n), na):
sa += pooled[i]
if abs(sa / na - (total - sa) / nb) >= obs - 1e-9:
cnt += 1
p = (cnt + 1) / (B + 1)
return obs, p, f"MC/B={B:,}", math.sqrt(p * (1 - p) / (B + 1))
def _sd(xs):
m = sum(xs) / len(xs)
return math.sqrt(sum((v - m) ** 2 for v in xs) / (len(xs) - 1))
def banded(shots):
"""{band: (shots, hits)}, plus overall."""
c = collections.Counter()
h = collections.Counter()
for s in shots:
c[s["band"]] += 1
h[s["band"]] += int(s["hit"])
c["ALL"] += len(shots)
h["ALL"] += int(sum(s["hit"] for s in shots))
return c, h
def per_run_band_rate(runs, band):
"""Per-run hit rate for a band, only runs that fired in that band."""
out = []
for fn, shots in runs.items():
sub = shots if band == "ALL" else [s for s in shots if s["band"] == band]
if sub:
out.append(sum(s["hit"] for s in sub) / len(sub))
return out
def main():
ap = argparse.ArgumentParser()
ap.add_argument("session_dir")
ap.add_argument("--reference", default=None)
args = ap.parse_args()
arms = [d for d in sorted(os.listdir(args.session_dir))
if os.path.isdir(os.path.join(args.session_dir, d))
and not d.startswith(".") and d != "frozen"]
arms = [a for a in arms if any(
f.endswith(".jsonl") and not f.endswith(".events.jsonl")
for f in os.listdir(os.path.join(args.session_dir, a)))]
ref = args.reference or arms[0]
data = {a: load_arm(args.session_dir, a) for a in arms}
print("=" * 100)
print("HIT RATE BY RANGE BAND (our shots; band = shooter->target distance px at the fire tick)")
print(f"session: {args.session_dir} arms: {', '.join(arms)} reference: {ref}")
print("=" * 100)
labels = BAND_LABELS + ["ALL"]
hdr = f"{'band':<10}" + "".join(f"{a:>26}" for a in arms)
print(hdr)
print("-" * len(hdr))
for band in labels:
cells = []
for a in arms:
c, h = banded([s for shots in data[a].values() for s in shots])
n = c[band]
cells.append(f"{n:>7} {h[band]:>5} {(100*h[band]/n if n else 0):>6.1f}%")
print(f"{band:<10}" + "".join(f"{c:>26}" for c in cells))
print("\nPER-RUN BAND RATES (shows the spread behind the pooled numbers)")
for band in labels:
print(f" band {band}")
for a in arms:
rates = per_run_band_rate(data[a], band)
txt = " ".join(f"{100*r:.0f}" for r in rates)
print(f" {a:<10} n={len(rates):>2} {txt}")
print("\nPER-BAND PERMUTATION TEST vs `" + ref + "` (per-run rates, two-sided)")
print(f"{'band':<10} {'arm':<10} {'d(pp)':>7} {'p':>9} {'method':<12} {'MCse':>7} {'MDE(pp)':>8}")
print("-" * 66)
for band in labels:
xa = per_run_band_rate(data[ref], band)
if len(xa) < 2:
continue
sd = _sd(xa)
mde = Z_ALPHA_POWER * sd * math.sqrt(2.0 / len(xa)) * 100
for a in arms:
if a == ref:
continue
xb = per_run_band_rate(data[a], band)
res = perm_test(xa, xb)
if res is None:
continue
obs, p, method, se = res
signed = (sum(xb) / len(xb) - sum(xa) / len(xa)) * 100 # arm - ref
print(f"{band:<10} {a:<10} {signed:>+7.2f} {p:>9.4f} {method:<12} "
f"{se:>7.4f} {mde:>8.2f}")
return 0
if __name__ == "__main__":
sys.exit(main())
+6
View File
@@ -0,0 +1,6 @@
# HeadOn-vs-Pattern long-range live gate (2 arms x 15 runs x 7 rounds).
# Shipped default is `onlyPattern` (Pattern is the only admitted rack gun), so:
# control = shipped Pattern-only rack, no env
# headon = Pattern off, HeadOn re-admitted -> HeadOn-only rack
control |
headon | TR_RACK_PATTERN=off TR_RACK_HEADON=both | static no-lead gun only