Spec §3 steps 2–5 + §8, TDD-ed offline (scripted, honesty-marked stand-in): - budget.py: BudgetMeter over TerminationContract — provider-reported usage only (missing usage fails closed), structured BudgetExceeded stop event. - loop.py: ModelClient protocol; blind parse-retry generation (never silent repair); round-capped debate with turn safety net and mandated VERDICT line; opt-in-reject checker gate (explicit REJECT overrides a validated outcome, validator rejection stands); most-recent-reason-verbatim informed refinement under max_attempts; validator_decision stamped BEFORE override, checker_decision as its own result field (§9, never conflated). - 45 new tests (121 total, no API key); four detach proofs run RED and reverted green: checker override, informed block, surfaced checker output, stamp-before-override. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QdSfQdND84oeq2mbjueLTS
162 lines
6.8 KiB
Python
162 lines
6.8 KiB
Python
"""Load-bearing: the checker gate (method-spec §3 Step 4, §11).
|
||
|
||
Either the checker actually gates, or the debate must not be called
|
||
maker-checker. RED-conditions proved here: the checker's surfaced output is
|
||
detached from the gate, OR its explicit REJECT no longer overrides an
|
||
otherwise-validated outcome. The gate is opt-in-reject (fail-open): APPROVE or
|
||
a missing marker never blocks; a validator rejection stands regardless of the
|
||
checker; the two falsifiers are recorded SEPARATELY (§9) — the provenance
|
||
field mirrors only the validator, the checker's decision is its own field.
|
||
"""
|
||
|
||
from __future__ import annotations
|
||
|
||
import json
|
||
|
||
from _scripted import ScriptedClient, reply
|
||
|
||
from portfolio_optimiser_claude.budget import BudgetMeter
|
||
from portfolio_optimiser_claude.contracts import TerminationContract
|
||
from portfolio_optimiser_claude.loop import (
|
||
CheckerVerdict,
|
||
apply_checker_gate,
|
||
run_project,
|
||
)
|
||
from portfolio_optimiser_claude.validator import Rejection, ValidatedProposal
|
||
|
||
# 100 × 1000 = 100_000 NOK total; nominal feasible 30_000 == p90 (no band).
|
||
VALID_PROPOSAL = {
|
||
"project_id": "p1",
|
||
"measure": "led-retrofit",
|
||
"affected_items": [{"code": "E01", "quantity": 100, "unit_cost": 1000}],
|
||
"claimed_saving_nok": 25000,
|
||
}
|
||
# IR-valid (≤ total) but above p90 → the validator rejects it.
|
||
OVERCLAIMED_PROPOSAL = dict(VALID_PROPOSAL, claimed_saving_nok=90000)
|
||
|
||
VALIDATED = ValidatedProposal(
|
||
validates=True,
|
||
claimed_saving_nok=25000,
|
||
nominal_feasible=30000,
|
||
p10=30000,
|
||
p50=30000,
|
||
p90=30000,
|
||
)
|
||
|
||
|
||
def _meter() -> BudgetMeter:
|
||
return BudgetMeter(TerminationContract(max_rounds=50, max_tokens=10_000))
|
||
|
||
|
||
class TestApplyCheckerGate:
|
||
def test_explicit_reject_overrides_a_validated_outcome(self) -> None:
|
||
# THE load-bearing seam: red if the override is detached.
|
||
verdict = CheckerVerdict(decision="reject", reason="double-counted savings")
|
||
outcome = apply_checker_gate(VALIDATED, verdict)
|
||
assert isinstance(outcome, Rejection)
|
||
assert outcome.reason.startswith("double-counted savings")
|
||
|
||
def test_approve_never_blocks(self) -> None:
|
||
verdict = CheckerVerdict(decision="approve", reason=None)
|
||
assert apply_checker_gate(VALIDATED, verdict) is VALIDATED
|
||
|
||
def test_absent_marker_never_blocks(self) -> None:
|
||
# Fail-open: the validator remains the sole gate on such runs.
|
||
verdict = CheckerVerdict(decision="absent", reason=None)
|
||
assert apply_checker_gate(VALIDATED, verdict) is VALIDATED
|
||
|
||
def test_a_validator_rejection_stands_regardless_of_the_checker(self) -> None:
|
||
validator_rejection = Rejection(reason="claimed saving exceeds the feasible bound")
|
||
verdict = CheckerVerdict(decision="approve", reason=None)
|
||
assert apply_checker_gate(validator_rejection, verdict) is validator_rejection
|
||
|
||
|
||
class TestRunProjectGating:
|
||
def test_checker_reject_flips_an_otherwise_validated_run(self) -> None:
|
||
# Speiltest (sesjonsplan S8): numbers pass the validator, the checker's
|
||
# surfaced REJECT flips the run outcome — red without the override.
|
||
client = ScriptedClient(
|
||
replies=[
|
||
reply("reasoning"),
|
||
reply("VERDICT: REJECT - assumptions unsupported"),
|
||
reply(json.dumps(VALID_PROPOSAL)),
|
||
]
|
||
)
|
||
result = run_project(client, "context", meter=_meter(), max_debate_rounds=1)
|
||
assert isinstance(result.outcome, Rejection)
|
||
assert result.outcome.reason.startswith("assumptions unsupported")
|
||
# §9: the two falsifiers are never conflated — the provenance mirror
|
||
# says the NUMBERS passed; the checker's decision is its own field.
|
||
assert result.validator_decision == "validated"
|
||
assert result.checker_decision == "reject"
|
||
|
||
def test_the_gate_consumes_the_checkers_last_surfaced_output(self) -> None:
|
||
# Two debate rounds, two distinct REJECT reasons: the gate must carry
|
||
# the LAST one — red if the surfaced-output seam is detached.
|
||
client = ScriptedClient(
|
||
replies=[
|
||
reply("reasoning v1"),
|
||
reply("VERDICT: REJECT - EARLY-REASON"),
|
||
reply("reasoning v2"),
|
||
reply("VERDICT: REJECT - FINAL-REASON"),
|
||
reply(json.dumps(VALID_PROPOSAL)),
|
||
]
|
||
)
|
||
result = run_project(client, "context", meter=_meter(), max_debate_rounds=2)
|
||
assert isinstance(result.outcome, Rejection)
|
||
assert result.outcome.reason.startswith("FINAL-REASON")
|
||
assert "EARLY-REASON" not in result.outcome.reason
|
||
|
||
def test_an_approved_run_passes_through_ungated(self) -> None:
|
||
client = ScriptedClient(
|
||
replies=[
|
||
reply("reasoning"),
|
||
reply("sound. VERDICT: APPROVE"),
|
||
reply(json.dumps(VALID_PROPOSAL)),
|
||
]
|
||
)
|
||
result = run_project(client, "context", meter=_meter(), max_debate_rounds=1)
|
||
assert isinstance(result.outcome, ValidatedProposal)
|
||
assert result.validator_decision == "validated"
|
||
assert result.checker_decision == "approve"
|
||
|
||
def test_a_missing_marker_is_fail_open(self) -> None:
|
||
client = ScriptedClient(
|
||
replies=[
|
||
reply("reasoning"),
|
||
reply("no explicit verdict line here"),
|
||
reply(json.dumps(VALID_PROPOSAL)),
|
||
]
|
||
)
|
||
result = run_project(client, "context", meter=_meter(), max_debate_rounds=1)
|
||
assert isinstance(result.outcome, ValidatedProposal)
|
||
assert result.checker_decision == "absent"
|
||
|
||
def test_a_validator_rejection_stands_even_when_the_checker_approves(self) -> None:
|
||
client = ScriptedClient(
|
||
replies=[
|
||
reply("reasoning"),
|
||
reply("VERDICT: APPROVE"),
|
||
reply(json.dumps(OVERCLAIMED_PROPOSAL)),
|
||
reply(json.dumps(OVERCLAIMED_PROPOSAL)),
|
||
reply(json.dumps(OVERCLAIMED_PROPOSAL)),
|
||
]
|
||
)
|
||
result = run_project(client, "context", meter=_meter(), max_debate_rounds=1, max_attempts=3)
|
||
assert isinstance(result.outcome, Rejection)
|
||
assert result.validator_decision == "rejected"
|
||
assert result.checker_decision == "approve"
|
||
|
||
def test_the_debates_converged_output_feeds_generation(self) -> None:
|
||
# §3 Step 3: the debate's converged proposer output feeds Step 2's
|
||
# generation context.
|
||
client = ScriptedClient(
|
||
replies=[
|
||
reply("CONVERGED-REASONING-MARKER"),
|
||
reply("VERDICT: APPROVE"),
|
||
reply(json.dumps(VALID_PROPOSAL)),
|
||
]
|
||
)
|
||
run_project(client, "context", meter=_meter(), max_debate_rounds=1)
|
||
generation_prompt = client.prompts("proposer")[1]
|
||
assert "CONVERGED-REASONING-MARKER" in generation_prompt
|