tinymind-native-8b-remote-handoff/bundle/evaluation/claims.py

"""Auditable claim-readiness gates for TinyMind."""

from __future__ import annotations

from dataclasses import dataclass
import json
from pathlib import Path
from typing import Any, Mapping


REQUIRED_CLAIM_CATEGORIES = (
    "quality",
    "size",
    "context",
    "stability",
    "speed",
    "quantization",
)

REQUIRED_ARTIFACTS = (
    "checkpoint",
    "int4_artifact",
    "dataset_manifest",
    "objective_report",
)


@dataclass(frozen=True)
class ClaimVerdict:
    world_best_claim_allowed: bool
    allowed_claims: list[str]
    blocked_claims: list[str]
    missing: list[str]

    def to_dict(self) -> dict:
        return {
            "world_best_claim_allowed": self.world_best_claim_allowed,
            "allowed_claims": self.allowed_claims,
            "blocked_claims": self.blocked_claims,
            "missing": self.missing,
        }


def _artifact_exists(value: Any) -> bool:
    if not isinstance(value, str) or not value.strip():
        return False
    return Path(value).exists()


def _rank1_comparison_ok(row: Mapping[str, Any]) -> bool:
    required = ("leaderboard", "metric", "model_score", "best_baseline_score", "rank", "source_url", "as_of")
    if not all(row.get(key) not in (None, "") for key in required):
        return False
    if not str(row["source_url"]).startswith(("https://", "http://")):
        return False
    try:
        return int(row["rank"]) == 1 and float(row["model_score"]) >= float(row["best_baseline_score"])
    except (TypeError, ValueError):
        return False


def evaluate_claim_readiness(evidence: Mapping[str, Any]) -> dict:
    missing: list[str] = []
    local_missing: list[str] = []
    measurements = evidence.get("measurements", {})
    if not isinstance(measurements, Mapping):
        measurements = {}

    local_smoke = bool(measurements)
    for category in REQUIRED_CLAIM_CATEGORIES:
        row = measurements.get(category)
        if not isinstance(row, Mapping) or row.get("passed") is not True:
            local_missing.append(f"measurement:{category}")

    artifacts = evidence.get("artifacts", {})
    if not isinstance(artifacts, Mapping):
        artifacts = {}
    for name in REQUIRED_ARTIFACTS:
        if not _artifact_exists(artifacts.get(name)):
            local_missing.append(f"artifact:{name}")

    missing.extend(local_missing)

    comparisons = evidence.get("comparisons", [])
    if not isinstance(comparisons, list):
        comparisons = []
    rank1 = [row for row in comparisons if isinstance(row, Mapping) and _rank1_comparison_ok(row)]
    if len(rank1) < 3:
        missing.append("comparisons:at_least_3_rank1_external")

    world_best_allowed = not missing
    allowed = ["local_smoke_verified"] if local_smoke else []
    if local_smoke and not local_missing:
        allowed.append("local_train_eval_complete")
    blocked: list[str] = []
    claim_scope = str(evidence.get("claim_scope", "world_best")).strip() or "world_best"
    if world_best_allowed:
        allowed.append(claim_scope)
    else:
        blocked.append(claim_scope)

    return ClaimVerdict(
        world_best_claim_allowed=world_best_allowed,
        allowed_claims=allowed,
        blocked_claims=blocked,
        missing=missing,
    ).to_dict()


def build_claim_dossier(evidence: Mapping[str, Any]) -> dict:
    verdict = evaluate_claim_readiness(evidence)
    return {
        "schema_version": "tinymind-claim-dossier-v1",
        "model_name": evidence.get("model_name", "TinyMind"),
        "claim_scope": evidence.get("claim_scope", "unspecified"),
        "as_of": evidence.get("as_of"),
        "verdict": verdict,
        "measurements": evidence.get("measurements", {}),
        "comparisons": evidence.get("comparisons", []),
        "artifacts": evidence.get("artifacts", {}),
        "rules": {
            "required_categories": list(REQUIRED_CLAIM_CATEGORIES),
            "required_artifacts": list(REQUIRED_ARTIFACTS),
            "external_rank1_comparisons_required": 3,
            "no_world_best_claim_without_saved_json_or_csv_metrics": True,
        },
    }


def _load_json(path: str | Path) -> dict:
    return json.loads(Path(path).read_text(encoding="utf-8"))


def _markdown(dossier: Mapping[str, Any]) -> str:
    verdict = dossier["verdict"]
    status = "ALLOWED" if verdict["world_best_claim_allowed"] else "BLOCKED"
    lines = [
        "# TinyMind Claim Dossier",
        "",
        f"- Model: {dossier.get('model_name')}",
        f"- Claim scope: {dossier.get('claim_scope')}",
        f"- As of: {dossier.get('as_of')}",
        f"- World-best claim: {status}",
        "",
        "## Allowed Claims",
    ]
    for claim in verdict["allowed_claims"]:
        lines.append(f"- {claim}")
    if not verdict["allowed_claims"]:
        lines.append("- None")

    lines.extend(["", "## Blocked Claims"])
    for claim in verdict["blocked_claims"]:
        lines.append(f"- {claim}")
    if not verdict["blocked_claims"]:
        lines.append("- None")

    lines.extend(["", "## Missing Evidence"])
    for item in verdict["missing"]:
        lines.append(f"- {item}")
    if not verdict["missing"]:
        lines.append("- None")

    lines.extend(["", "## Measurements"])
    measurements = dossier.get("measurements", {})
    if isinstance(measurements, Mapping):
        for category in REQUIRED_CLAIM_CATEGORIES:
            row = measurements.get(category, {})
            passed = row.get("passed") if isinstance(row, Mapping) else None
            score = row.get("score") if isinstance(row, Mapping) else None
            artifact = row.get("artifact") if isinstance(row, Mapping) else None
            lines.append(f"- {category}: passed={passed}, score={score}, artifact={artifact}")

    lines.extend(["", "## External Comparisons"])
    comparisons = dossier.get("comparisons", [])
    if isinstance(comparisons, list) and comparisons:
        for row in comparisons:
            if not isinstance(row, Mapping):
                continue
            lines.append(
                "- "
                f"{row.get('leaderboard')}: rank={row.get('rank')}, "
                f"metric={row.get('metric')}, model={row.get('model_score')}, "
                f"baseline={row.get('best_baseline_score')}, "
                f"as_of={row.get('as_of')}, source={row.get('source_url')}"
            )
    else:
        lines.append("- None")

    lines.extend(
        [
            "",
            "## Rule",
            "No best-in-world claim is permitted unless every required category passes, required artifacts exist, and at least three dated external rank-1 comparisons are saved.",
            "",
        ]
    )
    return "\n".join(lines)


def write_claim_dossier(evidence_path: str | Path, markdown_path: str | Path) -> dict:
    evidence_file = Path(evidence_path)
    out = Path(markdown_path)
    dossier = build_claim_dossier(_load_json(evidence_file))
    out.parent.mkdir(parents=True, exist_ok=True)
    out.write_text(_markdown(dossier), encoding="utf-8")
    json_path = out.with_suffix(".json")
    json_path.write_text(json.dumps(dossier, ensure_ascii=False, indent=2, sort_keys=True), encoding="utf-8")
    return dossier