Spaces:
Sleeping
Sleeping
File size: 7,077 Bytes
7952f32 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 | """Reward engine — per-turn (dense, small) and terminal (sparse, large).
Implementation follows PROPOSAL.md §5 verbatim. The two halves are pure
functions over lightweight envelopes so the server can call them without
threading state through the reward module.
Decisions worth flagging:
* ``All-behavioral-passing`` bonus is awarded only when there is at least
one behavioral test. The gate for the token-efficiency bonus, however,
treats zero behavioral tests as vacuously satisfied (so a tier-0 task
with no behavioral tests can still earn token-efficiency reward).
* ``type_checks_ok`` is tri-state: ``True`` / ``False`` / ``None``. ``None``
means the type-check gate didn't run (e.g. mypy isn't wired yet); the
+3 bonus is suppressed in that case.
"""
from __future__ import annotations
from dataclasses import dataclass, field
from enum import Enum
# Coefficients (PROPOSAL.md §5.1). Override at call time if you want.
ALPHA_TOKEN_COST: float = 0.0008
PER_TURN_COST: float = -0.1
MUTATION_FAIL: float = -2.0
SCHEMA_REJECTION: float = -2.0
DUPLICATE_ACTION: float = -1.0
# Terminal magnitudes (§5.2)
STRUCTURAL_PER_SAT: float = 1.0
BEHAVIORAL_PER_PASS: float = 3.0
ALL_STRUCTURAL_BONUS: float = 5.0
ALL_BEHAVIORAL_BONUS: float = 5.0
TYPE_CHECK_BONUS: float = 3.0
MATERIALIZE_FAIL_PENALTY: float = -8.0
TOKEN_EFFICIENCY_MAX: float = 5.0
# ---- per-turn -------------------------------------------------------
class ActionOutcome(str, Enum):
"""Coarse classification used by ``score_turn``.
``SUCCESS`` — mutation or info action returned ``ok=True``.
``FAILURE`` — handler raised :class:`ActionError` (rollback path).
``MALFORMED`` — pydantic schema rejected the action at parse time.
"""
SUCCESS = "success"
FAILURE = "failure"
MALFORMED = "malformed"
@dataclass(frozen=True)
class TurnReward:
base: float # outcome-dependent component
duplicate: float # 0 or DUPLICATE_ACTION
per_turn: float # PER_TURN_COST
token_cost: float # alpha * tokens_returned, negated
@property
def total(self) -> float:
return self.base + self.duplicate + self.per_turn + self.token_cost
def to_dict(self) -> dict[str, float]:
return {
"base": self.base,
"duplicate": self.duplicate,
"per_turn": self.per_turn,
"token_cost": self.token_cost,
"total": self.total,
}
def score_turn(
*,
outcome: ActionOutcome,
is_duplicate: bool,
tokens_returned: int,
alpha: float = ALPHA_TOKEN_COST,
per_turn_cost: float = PER_TURN_COST,
) -> TurnReward:
if outcome is ActionOutcome.SUCCESS:
base = 0.0
elif outcome is ActionOutcome.FAILURE:
base = MUTATION_FAIL
else: # MALFORMED
base = SCHEMA_REJECTION
return TurnReward(
base=base,
duplicate=DUPLICATE_ACTION if is_duplicate else 0.0,
per_turn=per_turn_cost,
token_cost=-alpha * max(0, tokens_returned),
)
# ---- terminal -------------------------------------------------------
@dataclass(frozen=True)
class TerminalReward:
structural: float # +1 per structural constraint satisfied
behavioral: float # +3 per behavioral test passing
bonus_all_structural: float
bonus_all_behavioral: float
bonus_type_checks: float
penalty_materialize: float # 0 or MATERIALIZE_FAIL_PENALTY
efficiency: float # gated by all-structural AND all-behavioral
components: dict[str, object] = field(default_factory=dict)
@property
def total(self) -> float:
return (
self.structural
+ self.behavioral
+ self.bonus_all_structural
+ self.bonus_all_behavioral
+ self.bonus_type_checks
+ self.penalty_materialize
+ self.efficiency
)
def to_dict(self) -> dict[str, object]:
return {
"structural": self.structural,
"behavioral": self.behavioral,
"bonus_all_structural": self.bonus_all_structural,
"bonus_all_behavioral": self.bonus_all_behavioral,
"bonus_type_checks": self.bonus_type_checks,
"penalty_materialize": self.penalty_materialize,
"efficiency": self.efficiency,
"total": self.total,
"components": self.components,
}
def score_terminal(
*,
n_structural_satisfied: int,
n_structural_total: int,
n_behavioral_passing: int,
n_behavioral_total: int,
materialization_ok: bool,
type_checks_ok: bool | None,
tokens_used: int,
budget: int,
) -> TerminalReward:
if n_structural_satisfied < 0 or n_structural_total < 0:
raise ValueError("structural counts must be non-negative")
if n_behavioral_passing < 0 or n_behavioral_total < 0:
raise ValueError("behavioral counts must be non-negative")
if budget <= 0:
raise ValueError("budget must be positive")
structural = STRUCTURAL_PER_SAT * n_structural_satisfied
behavioral = BEHAVIORAL_PER_PASS * n_behavioral_passing
all_structural = (
n_structural_total > 0 and n_structural_satisfied == n_structural_total
)
all_behavioral_present_and_passing = (
n_behavioral_total > 0 and n_behavioral_passing == n_behavioral_total
)
bonus_all_structural = ALL_STRUCTURAL_BONUS if all_structural else 0.0
bonus_all_behavioral = (
ALL_BEHAVIORAL_BONUS if all_behavioral_present_and_passing else 0.0
)
if type_checks_ok is True:
bonus_type_checks = TYPE_CHECK_BONUS
else:
bonus_type_checks = 0.0
penalty_materialize = (
0.0 if materialization_ok else MATERIALIZE_FAIL_PENALTY
)
# Efficiency bonus is gated on all-structural AND all-behavioral satisfied.
# When n_behavioral_total == 0 the behavioral half is vacuously satisfied
# for the gate's purposes (otherwise tier-0 tasks could never earn it).
behavioral_gate_ok = (
n_behavioral_total == 0
or n_behavioral_passing == n_behavioral_total
)
efficiency = 0.0
if all_structural and behavioral_gate_ok:
ratio = max(0.0, (budget - tokens_used) / budget)
efficiency = TOKEN_EFFICIENCY_MAX * ratio
return TerminalReward(
structural=structural,
behavioral=behavioral,
bonus_all_structural=bonus_all_structural,
bonus_all_behavioral=bonus_all_behavioral,
bonus_type_checks=bonus_type_checks,
penalty_materialize=penalty_materialize,
efficiency=efficiency,
components={
"n_structural_satisfied": n_structural_satisfied,
"n_structural_total": n_structural_total,
"n_behavioral_passing": n_behavioral_passing,
"n_behavioral_total": n_behavioral_total,
"materialization_ok": materialization_ok,
"type_checks_ok": type_checks_ok,
"tokens_used": tokens_used,
"budget": budget,
},
)
|