code/self_healing_finetuning/model_policy/roadmap.py

from __future__ import annotations

import json
from pathlib import Path
from typing import Any


ROADMAP_PATH = Path(__file__).with_name("roadmap.json")


def load_roadmap(path: Path = ROADMAP_PATH) -> dict[str, Any]:
    return json.loads(path.read_text(encoding="utf-8"))


def weighted_score(candidate: dict[str, Any], weights: dict[str, float]) -> float:
    context_tokens = float(candidate["context_tokens"])
    context_score = min(context_tokens / 131072.0, 1.0)
    score = (
        weights["base_capability"] * float(candidate["base_capability_score"])
        + weights["finance_bootstrap"] * float(candidate["finance_bootstrap_score"])
        + weights["context_window"] * context_score
        + weights["trainability"] * float(candidate["trainability_score"])
        + weights["license_and_operational_readiness"] * float(candidate["license_and_operational_readiness_score"])
    )
    return round(score, 4)


def roadmap_report(path: Path = ROADMAP_PATH) -> dict[str, Any]:
    roadmap = load_roadmap(path)
    weights = roadmap["quantitative_decision_weights"]
    candidates = []
    for candidate in roadmap["candidates"]:
        enriched = dict(candidate)
        enriched["weighted_score"] = weighted_score(candidate, weights)
        candidates.append(enriched)

    score_rank = sorted(candidates, key=lambda c: c["weighted_score"], reverse=True)
    context_rank = sorted(candidates, key=lambda c: int(c["context_tokens"]), reverse=True)
    current = next(c for c in candidates if c["id"] == roadmap["strategy"]["current_default_model"])

    return {
        "policy_version": roadmap["policy_version"],
        "candidate_count": len(candidates),
        "current_default_model": current["id"],
        "current_default_weighted_score": current["weighted_score"],
        "current_default_context_tokens": current["context_tokens"],
        "highest_score_candidate": score_rank[0]["id"],
        "highest_score": score_rank[0]["weighted_score"],
        "highest_context_candidate": context_rank[0]["id"],
        "highest_context_tokens": context_rank[0]["context_tokens"],
        "ranked_by_score": [
            {
                "id": c["id"],
                "track": c["track"],
                "weighted_score": c["weighted_score"],
                "context_tokens": c["context_tokens"],
            }
            for c in score_rank
        ],
        "ranked_by_context": [
            {
                "id": c["id"],
                "track": c["track"],
                "weighted_score": c["weighted_score"],
                "context_tokens": c["context_tokens"],
            }
            for c in context_rank
        ],
        "promotion_rules": roadmap["promotion_rules"],
    }


def candidate_evidence(model_id: str, path: Path = ROADMAP_PATH) -> dict[str, Any]:
    report = roadmap_report(path)
    ranked_by_score = report["ranked_by_score"]
    ranked_by_context = report["ranked_by_context"]
    score_rank = next((index + 1 for index, item in enumerate(ranked_by_score) if item["id"] == model_id), None)
    context_rank = next((index + 1 for index, item in enumerate(ranked_by_context) if item["id"] == model_id), None)
    score_record = next((item for item in ranked_by_score if item["id"] == model_id), None)
    if score_record is None:
        return {
            "policy_version": report["policy_version"],
            "model_id": model_id,
            "known_to_roadmap": False,
            "score_rank": None,
            "context_rank": None,
            "weighted_score": None,
            "context_tokens": None,
            "promotion_rules": report["promotion_rules"],
        }
    return {
        "policy_version": report["policy_version"],
        "model_id": model_id,
        "known_to_roadmap": True,
        "score_rank": score_rank,
        "context_rank": context_rank,
        "weighted_score": score_record["weighted_score"],
        "context_tokens": score_record["context_tokens"],
        "highest_score_candidate": report["highest_score_candidate"],
        "highest_context_candidate": report["highest_context_candidate"],
        "highest_context_tokens": report["highest_context_tokens"],
        "promotion_rules": report["promotion_rules"],
    }