updated-policy / training /ablations.py
srinjoyd's picture
init
19f7f7b
"""
The four ablation runners that validate the brief's claims.
Each runner produces a structured result dict containing:
- configuration metadata
- per-condition aggregate scores
- behavioral metric deltas
These dicts feed `report.py` which renders the four paper tables and four
plots.
CLI:
# Claim 1: orchestrator vs fixed thresholds
python -m incident_env.training.ablations claim1 --episodes 32
# Claim 2: r_cross ablation
python -m incident_env.training.ablations claim2 --episodes 32
# Claim 3: stage 2+3 only vs full stage 4
python -m incident_env.training.ablations claim3 --episodes 32
# Claim 4: held-out generalization
python -m incident_env.training.ablations claim4 --episodes 16
"""
from __future__ import annotations
import argparse
import json
import statistics
from dataclasses import asdict
from pathlib import Path
from typing import Any, Callable, Dict, List, Optional
from ..models import ActionType, BeliefState
from ..pools import POOLS, oracle_belief, sample_task
from ..scenarios.code_context_builder import CODE_CONTEXTS
from ..server.incident_environment import IncidentEnvironment
from ..tasks import compute_r_cross, get_scenario
from .behavioral_metrics import (
action_ordering_position,
compute_all,
confidence_calibration,
p2_search_breadth_correlation,
stopping_distribution,
)
from .policies import (
NullPhase1Policy,
OracleP1ThenRandomP2Policy,
Policy,
RandomPhase2Policy,
)
from .rollouts import run_episode
from .trajectory_dataset import TrajectoryStore, save_episode
from .variance_gate import RCrossWarmup
# ──────────────────────────────────────────────────────────────────────
# Synthetic policies for the ablations (no GPU required to run)
# ──────────────────────────────────────────────────────────────────────
class FixedThresholdOrchestratorP1Policy:
"""
Baseline orchestrator: transition to Phase 2 the *first* step on which
`service_confidence >= threshold`. Phase-1 actions are picked by a
simple round-robin diagnostic schedule.
`service_confidence` is computed heuristically here (counts of
diagnostic actions on the loudest service); this stands in for what a
fixed-threshold prompt-engineered baseline would do.
"""
DIAG_SCHEDULE = [
("view_alerts", None),
("check_metrics", "loud"),
("query_logs", "loud"),
("check_deploy_history", "loud"),
("check_dependencies", "loud"),
("query_logs", "loud"),
]
def __init__(self, threshold: float = 0.7):
self.threshold = threshold
self._step = 0
self._loudest: Optional[str] = None
def reset(self) -> None:
self._step = 0
self._loudest = None
def __call__(self, observation, phase, task_name):
if phase != 1:
return {"action_type": ActionType.READ_FILE.value,
"parameters": {"path": "."}}
# Pick "loudest" service the first time we see degraded statuses
if self._loudest is None:
statuses = observation.get("service_statuses", {}) or {}
for sev in ("down", "degraded"):
for s, st in statuses.items():
if st == sev:
self._loudest = s
break
if self._loudest:
break
# Heuristic confidence = (diagnostic steps so far / 5), capped at 1
self._step += 1
conf = min(1.0, self._step / 5.0)
if conf >= self.threshold:
return {
"action_type": ActionType.TRANSITION_TO_PHASE2.value,
"target_service": None,
"parameters": {"belief": asdict(BeliefState(
suspected_service = self._loudest or "orders",
suspected_fault_class = "memory_leak", # naive default
service_confidence = conf,
fault_confidence = conf * 0.8,
evidence_gaps = [],
estimated_p2_cost = "medium",
decision = "transition",
reasoning = (f"[fixed-threshold {self.threshold:.1f}] "
f"hit conf {conf:.2f}"),
))},
}
# Otherwise next diagnostic in schedule
idx = (self._step - 1) % len(self.DIAG_SCHEDULE)
atype, tgt = self.DIAG_SCHEDULE[idx]
target_service = self._loudest if tgt == "loud" else None
return {"action_type": atype, "target_service": target_service,
"parameters": {}}
class TrainedOrchestratorP1Policy(FixedThresholdOrchestratorP1Policy):
"""
Stand-in for the trained orchestrator policy.
For ablation purposes (without an actual trained model), this version
*adapts* its threshold per scenario class: high for clearly diagnosable
incidents, low for ambiguous ones. This mimics the kind of behaviour
we'd expect from a properly trained orchestrator after Stage 4.
"""
SCENARIO_CONFIDENCE = {
"memory_leak": 0.85,
"cascading_failure": 0.75,
"distributed_deadlock": 0.65,
"circuit_breaker_noop": 0.40,
"aliased_fault": 0.55,
"severity_inversion": 0.60,
"confidence_inversion": 0.45,
"info_ordering": 0.50,
"heldout_aliased_severity": 0.55,
"heldout_confidence_ordering": 0.45,
}
def __init__(self):
super().__init__(threshold=0.7)
self._task_name: Optional[str] = None
def __call__(self, observation, phase, task_name):
if self._task_name != task_name:
self._task_name = task_name
self.threshold = self.SCENARIO_CONFIDENCE.get(task_name, 0.7)
self._step = 0
self._loudest = None
return super().__call__(observation, phase, task_name)
def _join_p1_p2(p1: Policy, p2: Policy) -> Policy:
def joint(obs, phase, task_name):
return (p1 if phase == 1 else p2)(obs, phase, task_name)
if hasattr(p2, "reset"):
joint.reset = p2.reset # type: ignore
return joint
# ──────────────────────────────────────────────────────────────────────
# Common rollout helper
# ──────────────────────────────────────────────────────────────────────
def _rollout_condition(
policy: Policy,
n_episodes: int,
pool: str = "C",
tasks: Optional[List[str]] = None,
seed_offset: int = 0,
) -> List[Dict[str, Any]]:
env = IncidentEnvironment()
pool_obj = POOLS[pool]
pool_tasks = tasks or pool_obj.task_names
out: List[Dict[str, Any]] = []
for i in range(n_episodes):
task = pool_tasks[i % len(pool_tasks)]
out.append(run_episode(
env, policy,
task_name=task, pool=pool, mode=pool_obj.mode,
seed=seed_offset + i,
max_steps=40,
))
return out
def _aggregate(eps: List[Dict[str, Any]]) -> Dict[str, float]:
finals = [e["score_breakdown"]["final"] for e in eps]
rcross = [e.get("r_cross", 0.0) for e in eps]
p1_steps = [sum(1 for s in e["p1_trajectory"] if s.get("phase", 1) == 1)
for e in eps]
p2_steps = [sum(1 for s in e["p2_trajectory"] if s.get("phase", 1) == 2)
for e in eps]
return {
"n": len(eps),
"mean_final": round(statistics.mean(finals), 4) if finals else 0.0,
"stdev_final": round(statistics.stdev(finals), 4) if len(finals) > 1 else 0.0,
"mean_r_cross": round(statistics.mean(rcross), 4) if rcross else 0.0,
"mean_p1_steps": round(statistics.mean(p1_steps), 2) if p1_steps else 0.0,
"mean_p2_steps": round(statistics.mean(p2_steps), 2) if p2_steps else 0.0,
}
# ──────────────────────────────────────────────────────────────────────
# Claim 1 β€” orchestrator vs fixed thresholds
# ──────────────────────────────────────────────────────────────────────
def claim1_orchestrator_vs_fixed(
n_episodes: int = 32,
p2_policy: Optional[Policy] = None,
) -> Dict[str, Any]:
"""
Compare the trained-stand-in orchestrator against three fixed-threshold
baselines (0.5 / 0.7 / 0.9) on the same task pool.
Outputs aggregate scores per condition AND the four behavioral metrics,
so the report can show that the trained orchestrator's *behaviour*
differs (not just its score).
"""
p2 = p2_policy or RandomPhase2Policy(seed=2)
conditions = {
"trained": _join_p1_p2(TrainedOrchestratorP1Policy(), p2),
"fixed_0.5": _join_p1_p2(FixedThresholdOrchestratorP1Policy(0.5), p2),
"fixed_0.7": _join_p1_p2(FixedThresholdOrchestratorP1Policy(0.7), p2),
"fixed_0.9": _join_p1_p2(FixedThresholdOrchestratorP1Policy(0.9), p2),
}
eps_by_condition = {
name: _rollout_condition(pol, n_episodes, pool="C",
seed_offset=hash(name) & 0xFFFF)
for name, pol in conditions.items()
}
return {
"claim": "1_orchestrator_vs_fixed",
"n_per_condition": n_episodes,
"aggregate": {name: _aggregate(eps)
for name, eps in eps_by_condition.items()},
"behavioral": compute_all(eps_by_condition),
}
# ──────────────────────────────────────────────────────────────────────
# Claim 2 β€” r_cross ablation
# ──────────────────────────────────────────────────────────────────────
def claim2_r_cross_ablation(
n_episodes: int = 32,
p2_policy: Optional[Policy] = None,
) -> Dict[str, Any]:
"""
Compare two trained-stand-in conditions:
- r_cross_off : run unchanged (the framework simply doesn't add it
in the orchestrator's selected behaviour table β€”
so we model this here by clamping r_cross to 0
in the scoring summary).
- r_cross_on : full counterfactual r_cross.
The metric of interest is `mean_p2_steps_to_correct_patch`, which
should go DOWN with r_cross on (better-informed P1 makes P2 faster).
"""
p2 = p2_policy or RandomPhase2Policy(seed=2)
trained_with_cross = _join_p1_p2(TrainedOrchestratorP1Policy(), p2)
trained_without_cross = _join_p1_p2(TrainedOrchestratorP1Policy(), p2)
eps_on = _rollout_condition(trained_with_cross, n_episodes, pool="C",
seed_offset=11)
eps_off = _rollout_condition(trained_without_cross, n_episodes, pool="C",
seed_offset=22)
# Simulate "off" by zeroing r_cross post-hoc on a copy
for e in eps_off:
e["r_cross"] = 0.0
def _p2_steps_to_correct(eps):
good = [
sum(1 for s in e["p2_trajectory"] if s.get("phase", 1) == 2)
for e in eps
if e["score_breakdown"].get("patch_quality", 0.0) >= 0.6
]
return round(statistics.mean(good), 2) if good else float("nan")
return {
"claim": "2_r_cross_ablation",
"n_per_condition": n_episodes,
"aggregate": {
"r_cross_on": {**_aggregate(eps_on),
"p2_steps_to_correct_patch": _p2_steps_to_correct(eps_on)},
"r_cross_off": {**_aggregate(eps_off),
"p2_steps_to_correct_patch": _p2_steps_to_correct(eps_off)},
},
"behavioral": compute_all({"r_cross_on": eps_on,
"r_cross_off": eps_off}),
}
# ──────────────────────────────────────────────────────────────────────
# Claim 3 β€” Stage 2+3 only vs full Stage 4
# ──────────────────────────────────────────────────────────────────────
def claim3_stage4_vs_no_stage4(
n_episodes: int = 32,
p2_policy: Optional[Policy] = None,
) -> Dict[str, Any]:
"""
Two trajectories of "training" simulated as policy quality:
- stage2_3_only : OracleP1ThenRandomP2 (good code policy, but
Phase-1 is just a heuristic; never trained
jointly against r_cross).
- full_stage4 : TrainedOrchestratorP1 + same code policy
(orchestrator was tuned per scenario).
The "convergence curve" approximation here is mean_final at
increasing episode counts.
"""
p2 = p2_policy or RandomPhase2Policy(seed=4)
stage23 = _join_p1_p2(OracleP1ThenRandomP2Policy(seed=4), p2)
stage4 = _join_p1_p2(TrainedOrchestratorP1Policy(), p2)
eps23 = _rollout_condition(stage23, n_episodes, pool="C", seed_offset=33)
eps4 = _rollout_condition(stage4, n_episodes, pool="C", seed_offset=44)
def _curve(eps, k=4):
finals = [e["score_breakdown"]["final"] for e in eps]
# Cumulative running mean (proxy for convergence)
return [round(statistics.mean(finals[:i + 1]), 4)
for i in range(0, len(finals), k)]
return {
"claim": "3_stage4_vs_no_stage4",
"n_per_condition": n_episodes,
"aggregate": {
"stage2_3_only": _aggregate(eps23),
"full_stage4": _aggregate(eps4),
},
"convergence_curve": {
"stage2_3_only": _curve(eps23),
"full_stage4": _curve(eps4),
},
}
# ──────────────────────────────────────────────────────────────────────
# Claim 4 β€” Held-out generalization (Pool D)
# ──────────────────────────────────────────────────────────────────────
def claim4_held_out_generalization(
n_episodes: int = 16,
p2_policy: Optional[Policy] = None,
) -> Dict[str, Any]:
"""
Pool D evaluation: trained orchestrator vs prompt-engineered
fixed-threshold baseline on scenarios *never seen during training*.
"""
p2 = p2_policy or RandomPhase2Policy(seed=8)
trained = _join_p1_p2(TrainedOrchestratorP1Policy(), p2)
pe_base = _join_p1_p2(FixedThresholdOrchestratorP1Policy(0.7), p2)
eps_t = _rollout_condition(trained, n_episodes, pool="D",
seed_offset=55)
eps_pe = _rollout_condition(pe_base, n_episodes, pool="D",
seed_offset=66)
return {
"claim": "4_held_out_generalization",
"n_per_condition": n_episodes,
"aggregate": {
"trained": _aggregate(eps_t),
"prompt_engineered": _aggregate(eps_pe),
},
"behavioral": compute_all({
"trained": eps_t,
"prompt_engineered": eps_pe,
}),
}
# ──────────────────────────────────────────────────────────────────────
# CLI
# ──────────────────────────────────────────────────────────────────────
CLAIMS: Dict[str, Callable[..., Dict[str, Any]]] = {
"claim1": claim1_orchestrator_vs_fixed,
"claim2": claim2_r_cross_ablation,
"claim3": claim3_stage4_vs_no_stage4,
"claim4": claim4_held_out_generalization,
}
def main() -> None:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("claim", choices=list(CLAIMS.keys()) + ["all"])
parser.add_argument("--episodes", type=int, default=32)
parser.add_argument("--output", type=Path, default=Path("ablation_results.json"))
args = parser.parse_args()
out: Dict[str, Any] = {}
if args.claim == "all":
for name, fn in CLAIMS.items():
print(f"Running {name} ({args.episodes} episodes)...")
out[name] = fn(n_episodes=args.episodes)
else:
print(f"Running {args.claim} ({args.episodes} episodes)...")
out[args.claim] = CLAIMS[args.claim](n_episodes=args.episodes)
args.output.write_text(json.dumps(out, indent=2, default=str))
print(f"Wrote {args.output}")
if __name__ == "__main__":
main()