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:
parent
9a4caeb419
commit
22bfc80dda
7 changed files with 947 additions and 2 deletions
|
|
@ -124,7 +124,10 @@ def seed_store_from_bundle(store: VerdictStore, bundle_dir: Path) -> int:
|
|||
Entries are keyed on the bundle's candidate features, read from the IR
|
||||
projection (fail-fast, required input). The rationale is built from the
|
||||
``description`` frontmatter plus, when present, the structured learning fields.
|
||||
Returns the number of verdict files seeded.
|
||||
A ``verdict_id`` in the frontmatter (promoted files, §6) is read VERBATIM
|
||||
(§4.2) — re-minting from the bundle features would collide distinct
|
||||
promoted candidates into one first-write-wins store slot. Returns the
|
||||
number of verdict files seeded.
|
||||
"""
|
||||
features = CandidateFeatures.from_proposal(load_validator_input(bundle_dir))
|
||||
seeded = 0
|
||||
|
|
@ -141,7 +144,7 @@ def seed_store_from_bundle(store: VerdictStore, bundle_dir: Path) -> int:
|
|||
rationale = f"{rationale} {learning}".strip()
|
||||
store.add(
|
||||
VerdictRecord(
|
||||
verdict_id=mint_verdict_id(features),
|
||||
verdict_id=concept.frontmatter.get("verdict_id") or mint_verdict_id(features),
|
||||
decision=concept.frontmatter.get("decision", _DEFAULT_SEED_DECISION),
|
||||
rationale=rationale,
|
||||
features=features,
|
||||
|
|
|
|||
132
src/portfolio_optimiser_claude/inbox.py
Normal file
132
src/portfolio_optimiser_claude/inbox.py
Normal 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)
|
||||
70
src/portfolio_optimiser_claude/persona.py
Normal file
70
src/portfolio_optimiser_claude/persona.py
Normal file
|
|
@ -0,0 +1,70 @@
|
|||
"""The persona example artifact loader (method-spec §4.3).
|
||||
|
||||
The expert-reviewer persona is a shared skill artifact: one canonical example
|
||||
verdict JSON (``decision``, ``marker``, ``rationale``). The persona judgement
|
||||
is sourced from that artifact AT CALL TIME — never from an inlined copy — and
|
||||
loading is fail-fast: the artifact is required input (contrast the tolerant
|
||||
inbox, §5). In simulation the persona plays the human on the WRITE side of the
|
||||
inbox role split: ``drop_persona_verdict`` authors a §4.2 verdict file through
|
||||
the same folder interface a production expert uses.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
from pathlib import Path
|
||||
from typing import Literal
|
||||
|
||||
from pydantic import BaseModel, Field, model_validator
|
||||
|
||||
from portfolio_optimiser_claude.experience import CandidateFeatures
|
||||
from portfolio_optimiser_claude.inbox import VerdictDocument, write_verdict
|
||||
|
||||
|
||||
class PersonaVerdict(BaseModel):
|
||||
"""The canonical persona example (§4.3): a run-path decision + traceable payload."""
|
||||
|
||||
# §4.3: MUST be a run-path value (§4.1) — the seed-only vocabulary is rejected.
|
||||
decision: Literal["approved", "rejected"]
|
||||
marker: str = Field(min_length=1)
|
||||
rationale: str = Field(min_length=1)
|
||||
|
||||
@model_validator(mode="after")
|
||||
def _marker_is_traceable(self) -> PersonaVerdict:
|
||||
# The marker is the payload a simulation follows from the persona's
|
||||
# judgement into a later run's prompt — it must live in the rationale.
|
||||
if self.marker not in self.rationale:
|
||||
raise ValueError(
|
||||
f"persona example: marker {self.marker!r} is not a substring of the rationale"
|
||||
)
|
||||
return self
|
||||
|
||||
|
||||
def load_persona_example(artifact_path: Path) -> PersonaVerdict:
|
||||
"""Load the shared artifact — FAIL-FAST on a missing or malformed file (§4.3)."""
|
||||
return PersonaVerdict.model_validate(json.loads(artifact_path.read_text(encoding="utf-8")))
|
||||
|
||||
|
||||
def drop_persona_verdict(
|
||||
inbox_dir: Path,
|
||||
artifact_path: Path,
|
||||
features: CandidateFeatures,
|
||||
*,
|
||||
description: str,
|
||||
) -> VerdictDocument:
|
||||
"""The persona judges a candidate and drops a verdict file into the inbox.
|
||||
|
||||
The judgement (decision + rationale) is read from the artifact at call
|
||||
time; the id is minted from the candidate features (§4.2); the file is
|
||||
written through the authoring primitive (§5) — the same interface a
|
||||
production expert uses.
|
||||
"""
|
||||
persona = load_persona_example(artifact_path)
|
||||
verdict = VerdictDocument.from_candidate(
|
||||
features,
|
||||
decision=persona.decision,
|
||||
rationale=persona.rationale,
|
||||
description=description,
|
||||
)
|
||||
write_verdict(inbox_dir, verdict)
|
||||
return verdict
|
||||
103
src/portfolio_optimiser_claude/promotion.py
Normal file
103
src/portfolio_optimiser_claude/promotion.py
Normal file
|
|
@ -0,0 +1,103 @@
|
|||
"""The promotion gate (method-spec §3 Step 8, §6) — an opt-in PUBLIC primitive.
|
||||
|
||||
``promote`` lifts one APPROVED verdict from the raw output layer into the OKF
|
||||
context layer as a navigable ``type: verdict`` concept file. It is NEVER wired
|
||||
into the run itself: the system reads context; the gate/persona promotes.
|
||||
Fail-closed — only human/persona-approved knowledge enters the wiki, never raw
|
||||
agent output (self-contamination). The index link label is FIXED and neutral:
|
||||
the index body flows verbatim into the rendered read-context, so a descriptive
|
||||
label would leak the learning signal around the gated fold.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import hashlib
|
||||
import re
|
||||
from pathlib import Path
|
||||
|
||||
from portfolio_optimiser_claude.inbox import VerdictDocument
|
||||
|
||||
# §6: the gate's accepted set — approved_with_adjustment exists ONLY here and
|
||||
# in bundle-seed frontmatter, never on the run path (§4).
|
||||
_ACCEPTED_DECISIONS = frozenset({"approved", "approved_with_adjustment"})
|
||||
_PROMOTED_PREFIX = "promoted-verdict-"
|
||||
# §6: FIXED neutral label — carries NO verdict signal.
|
||||
_NEUTRAL_LABEL = "Promotert ekspert-dom"
|
||||
_TOKEN_UNSAFE = re.compile(r"[^A-Za-z0-9._-]")
|
||||
_INDEX_FILENAME = "index.md"
|
||||
|
||||
|
||||
class PromotionError(ValueError):
|
||||
"""A verdict was refused at the gate (§6) — nothing written, nothing linked."""
|
||||
|
||||
|
||||
def _filename_token(verdict_id: str) -> str:
|
||||
# Path-safe, fail-closed against escaping names: sanitise to the safe
|
||||
# alphabet; a degenerate token falls back to a content hash.
|
||||
token = _TOKEN_UNSAFE.sub("", verdict_id)
|
||||
if not token.strip("."):
|
||||
token = hashlib.sha256(verdict_id.encode("utf-8")).hexdigest()[:16]
|
||||
return token
|
||||
|
||||
|
||||
def _frontmatter_text(value: str) -> str:
|
||||
# The frontmatter parser is line-oriented — prose must stay on one line.
|
||||
return " ".join(value.split())
|
||||
|
||||
|
||||
def promote(
|
||||
verdict: VerdictDocument,
|
||||
bundle_dir: Path,
|
||||
*,
|
||||
approved_by: str,
|
||||
experiment: str,
|
||||
timestamp: str,
|
||||
) -> Path:
|
||||
"""Promote one approved verdict into the bundle — fail-closed, idempotent link.
|
||||
|
||||
``timestamp`` is an explicit REQUIRED argument (no wall-clock default), so
|
||||
promotion is deterministic and reproducible. The promoted file is minimal:
|
||||
the rationale becomes ``description`` (the learning signal as prose) — the
|
||||
structured learning fields of hand-authored seeds are never reproduced.
|
||||
Ids key on candidate features (§4.2), so re-approving the same candidate
|
||||
overwrites the same file (last-write-wins): one curated verdict file per
|
||||
distinct candidate, not one per verdict event.
|
||||
"""
|
||||
if verdict.decision not in _ACCEPTED_DECISIONS:
|
||||
raise PromotionError(
|
||||
f"promotion refused: decision {verdict.decision!r} is not in the "
|
||||
f"gate's accepted set {sorted(_ACCEPTED_DECISIONS)} (§6 fail-closed)"
|
||||
)
|
||||
filename = f"{_PROMOTED_PREFIX}{_filename_token(verdict.id)}.md"
|
||||
path = (bundle_dir / filename).resolve()
|
||||
if path.parent != bundle_dir.resolve():
|
||||
raise PromotionError(f"promotion refused: {filename!r} escapes the bundle")
|
||||
path.write_text(
|
||||
"---\n"
|
||||
"type: verdict\n"
|
||||
f"title: {_NEUTRAL_LABEL}\n"
|
||||
f"decision: {verdict.decision}\n"
|
||||
f"description: {_frontmatter_text(verdict.rationale)}\n"
|
||||
f"verdict_id: {verdict.id}\n"
|
||||
f"provenance: approved_by={approved_by}; experiment={experiment}; "
|
||||
f"timestamp={timestamp}\n"
|
||||
"tags: [verdict, promoted]\n"
|
||||
"---\n",
|
||||
encoding="utf-8",
|
||||
)
|
||||
_link_from_index(bundle_dir, filename)
|
||||
return path
|
||||
|
||||
|
||||
def _link_from_index(bundle_dir: Path, filename: str) -> None:
|
||||
# §6: navigation follows only index cross-links — an unlinked file is
|
||||
# unreachable. Idempotent: re-promotion never double-links. (reference
|
||||
# limitation: the read-modify-write is not atomic — single-process MVP.)
|
||||
index_path = bundle_dir / _INDEX_FILENAME
|
||||
index_text = index_path.read_text(encoding="utf-8")
|
||||
if f"({filename})" in index_text:
|
||||
return
|
||||
link_line = f"- [{_NEUTRAL_LABEL}]({filename})\n"
|
||||
if not index_text.endswith("\n"):
|
||||
index_text += "\n"
|
||||
index_path.write_text(index_text + link_line, encoding="utf-8")
|
||||
Loading…
Add table
Add a link
Reference in a new issue