feat(learning): S9 — D7 læringssløyfe: verdict-inbox, fail-closed promoteringsgate, artefakt-sourced persona

- inbox.py (§4.2+§5): VerdictDocument med verbatim-id-regel; write_verdict
  authoring-primitiv (deterministisk JSON); load_inbox tolerant (skip, aldri
  raise; sortert på filnavn); merge_inbox_into_store first-write-wins,
  idempotent, skriver aldri (rolle-splitt §3 steg 7)
- promotion.py (§6): promote fail-closed mot {approved,
  approved_with_adjustment}; eksplisitt påkrevd timestamp; minimal frontmatter
  (rationale → description, aldri strukturerte læringsfelt); path-safe token
  med content-hash-fallback; idempotent index-lenking med fast nøytral label
- persona.py (§4.3): load_persona_example fail-fast (run-path-vokabular,
  marker ⊆ rationale); drop_persona_verdict artefakt-sourced ved kalltid mot
  delt shared/-artefakt
- experience.py (kirurgisk): seeding leser verdict_id VERBATIM fra frontmatter
  — re-minting ville kollidert distinkte promoterte kandidater
- 43 nye load-bearing tester (step7/step8/persona), 164/164 uten API-nøkkel;
  to-runs-bevis med fersk store + tom-inbox-kontroll; fire detach-bevis kjørt
  røde og revertert grønne

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QdSfQdND84oeq2mbjueLTS
This commit is contained in:
Kjell Tore Guttormsen 2026-07-03 07:36:15 +02:00
commit 22bfc80dda
7 changed files with 947 additions and 2 deletions

View file

@ -0,0 +1,132 @@
"""The verdict file and the inbox folder contract (method-spec §4.2, §5).
The long feedback loop's folder interface: an expert (or, in simulation, the
persona) drops one verdict per ``{id}.json`` file into an inbox folder; a
separate, later run picks it up fully resumable, no live-session assumption.
Role split (§3 Step 7, unwaivable): the system READS the inbox (tolerant load,
merge into the store); writing is the authoring primitive's job, used only by
the expert/persona side a run never persists its own captured verdict back.
Decision vocabulary at this layer: the file carries the expert decision as a
plain string the run-path feedback contract (§4.1) and the promotion gate's
accepted set (§6) are where the vocabulary is policed, not the raw file layer.
"""
from __future__ import annotations
import json
from pathlib import Path
from pydantic import BaseModel, Field, ValidationError
from portfolio_optimiser_claude.experience import (
CandidateFeatures,
VerdictRecord,
VerdictStore,
mint_verdict_id,
)
class ProposalFeatures(BaseModel):
"""§4.2 ``proposal_features``: the structural features of the judged candidate.
``description`` is surface text deliberately excluded from both
similarity ranking and id minting.
"""
affected_codes: list[str]
measure_type: str
claimed_saving_nok: float
description: str
class VerdictDocument(BaseModel):
"""One verdict file (§4.2). A LOADED ``id`` is kept verbatim — never re-minted."""
id: str = Field(min_length=1)
decision: str = Field(min_length=1)
rationale: str = Field(min_length=1)
proposal_features: ProposalFeatures
@classmethod
def from_candidate(
cls,
features: CandidateFeatures,
*,
decision: str,
rationale: str,
description: str,
) -> VerdictDocument:
"""Author a verdict for a candidate — the id is minted per §4.2."""
return cls(
id=mint_verdict_id(features),
decision=decision,
rationale=rationale,
proposal_features=ProposalFeatures(
affected_codes=sorted(features.affected_codes),
measure_type=features.measure_type,
claimed_saving_nok=features.claimed_saving_nok,
description=description,
),
)
def features(self) -> CandidateFeatures:
return CandidateFeatures(
affected_codes=frozenset(self.proposal_features.affected_codes),
measure_type=self.proposal_features.measure_type,
claimed_saving_nok=self.proposal_features.claimed_saving_nok,
)
def to_record(self) -> VerdictRecord:
return VerdictRecord(
verdict_id=self.id, # verbatim (§4.2) — raw number formatting could diverge
decision=self.decision,
rationale=self.rationale,
features=self.features(),
)
def write_verdict(inbox_dir: Path, verdict: VerdictDocument) -> Path:
"""The authoring primitive (§5): ``{id}.json``, written deterministically.
Creates the directory if needed; sorted keys, 2-space indent. The disk
layer is LAST-write-wins per file (§4.2).
"""
inbox_dir.mkdir(parents=True, exist_ok=True)
path = inbox_dir / f"{verdict.id}.json"
payload = json.dumps(verdict.model_dump(), sort_keys=True, indent=2)
path.write_text(payload, encoding="utf-8")
return path
def load_inbox(inbox_dir: Path) -> list[VerdictDocument]:
"""Tolerant load (§5): the raw layer is written out of band — skip, never raise.
A missing folder yields zero verdicts; files that are not ``.json``, fail
to parse, or lack a required top-level key are SKIPPED. Deterministic
order: sorted by filename.
"""
if not inbox_dir.is_dir():
return []
verdicts: list[VerdictDocument] = []
for path in sorted(inbox_dir.iterdir(), key=lambda p: p.name):
if path.suffix != ".json" or not path.is_file():
continue
try:
verdicts.append(VerdictDocument.model_validate(json.loads(path.read_text("utf-8"))))
except (OSError, ValueError, ValidationError):
continue
return verdicts
def merge_inbox_into_store(store: VerdictStore, inbox_dir: Path) -> int:
"""Merge, never replace (§5): per-verdict add, first-write-wins per id.
Runs BEFORE the Step-1 fold, so a passed-in store's existing verdicts
survive (cross-project threading) and repeated merges are idempotent.
Returns the number of inbox verdicts ingested; never writes anything.
"""
verdicts = load_inbox(inbox_dir)
for verdict in verdicts:
store.add(verdict.to_record())
return len(verdicts)