Route model: correct for reporting bias

- core/atlas/bias.py: for each rare roll (shadow, meadow Blue) and each source group,
  the factor by which the group over-reports the rare outcome relative to complete
  logging (bot, bulk logs), with a 90% range; rare outcomes from biased groups count
  1/rho; a roll the complete logs have (almost) never shown is learned from them only
- planner/policy.yaml: groups, rolls and thresholds; estimates published in
  policy.json under model.reportingBias
- fit_alpha: a roll seen only one way is pooled fully
- docs: goal accounts follow the best route, research accounts use Thompson sampling
  (decisionMode atlas_thompson)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UCzTd31GphpMXDhQqWrCMQ
This commit is contained in:
ekcSchiano
2026-11-10 02:29:09 +00:00
co-authored by Claude Opus 5.5
parent c0355616b6
commit 2271fe0976
8 changed files with 244 additions and 18 deletions
+17 -4
View File
@@ -148,21 +148,30 @@ def ctx_key(v, names):
# ---------------------------------------------------------------- 2. counts
def count_tables(prepared):
"""Counts for every chance event at every back-off level, plus each encounter's outcomes."""
def count_tables(prepared, weights=None):
"""Counts for every chance event at every back-off level, plus each encounter's outcomes. weights: per run,
{chance event: weight of the outcome it observed} (reporting-bias corrections, core/atlas/bias.py)."""
T = {}
for node, sp in CHANCE.items():
t = T[node] = {}
for v in prepared:
for i, v in enumerate(prepared):
y = OBSERVE[node](v)
if not y:
continue
w = weights[i].get(node, 1) if weights else 1
if w == 0:
continue
for L in range(len(sp["context"]) + 1):
names = sp["context"][:L]
if any(v.get(k) is None for k in names):
break
c = t.setdefault(f"{L}:{ctx_key(v, names)}", {})
c[y] = c.get(y, 0) + 1
c[y] = c.get(y, 0) + w
for t in T.values(): # weighted counts are fractions: keep them short and stable
for c in t.values():
for k, x in c.items():
if isinstance(x, float):
c[k] = round(x, 4)
enc = T["encounter"] = {}
for v in prepared:
if not v.get("encCtx"):
@@ -261,6 +270,10 @@ def fit_alpha(T, node):
pairs.append(([t[key].get(k, 0) for k in cats], pooled(parent)))
if not pairs:
return {"alpha": ALPHA, "logLik": None}
# a roll that has only ever come out one way says nothing about whether routes differ: pool fully (the fit
# would otherwise drive every route's chance of the unseen outcome to zero)
if sum(1 for k in cats if t.get("0:", {}).get(k, 0) > 0) < 2:
return {"alpha": max(ALPHA_GRID), "logLik": None}
best_a, best_ll, lls = ALPHA, -math.inf, {}
for a in ALPHA_GRID:
ll = 0