AssetFrame — engine (assetframe-script)

The AssetFrame forecast engine: market data → analysis → AI research brief → deterministic compiler → confidence → report → scoring → append‑only ledger → publish. It pairs with assetframe-infra (the Next.js site + Clerk auth, Neon, R2, Lemon Squeezy). This repo writes; the website reads.

Two‑repo architecture

Repo	Owns
assetframe-script (this)	Python engine + the publish chain — generation, scoring, the ledger, and writing to R2 + Neon.
assetframe-infra	The website + infra — auth, billing, the DB schema (migrations), and reading R2/Neon to serve users.

They share R2 (report files) and Neon (tables) — no shared code. The contract:

• R2 — private bucket, key format <date>/<slug>/<file> (free/pro × html/pdf, plus preview.png).
• Neon — the engine INSERTs, the web SELECTs: editions, open_calls, open_call_predictions, scored_results (id format AF-YYYYMMDD-<SLUG> / report_ref). Schema migrations live in assetframe-infra; the engine only inserts data.
• Enums — scripts/pipeline/scoring/taxonomy.py (confidence buckets / asset classes) is mirrored in the infra repo's web/lib/content.ts. Keep the two in sync.

Setup

pip install -r requirements.txt     # fpdf2 (PDF), pymupdf (preview)
pip install boto3                   # only for publish.py (R2 upload)
npm install                         # @neondatabase/serverless, for sync-db.mjs
cp .env.example .env                # fill R2_*, DATABASE_URL[_DEV], EODHD/ANTHROPIC

Daily run (the scheduler decides; Claude only writes/criticises)

The engine is a Python package (scripts/ with concern subpackages: pipeline, scheduler, analytics, delivery, coordination). Run every entrypoint as a module from the repo root:

python -m scripts.scheduler.config.validate_config        # validate config/assets.json
python -m scripts.scheduler.run.run_daily --mode dry_run       # resolve today's DUE assets, write nothing
python -m scripts.scheduler.run.run_daily --mode score_only    # score closed windows + refresh memory (idempotent)
python -m scripts.scheduler.run.run_daily --mode generate_only # generate due assets

Modes: dry_run | score_only | generate_only | production (Phase 2 adds approval). Scope with --asset <id> or --asset-class fx. Run manifests land in runs/<date>/run_manifest.json.

Publish chain (engine → site)

python -m scripts.delivery.export_content   # -> content/catalog.json + content/track-record.json
python -m scripts.delivery.publish          # -> uploads report files to R2
npm run sync-db                     # -> writes editions + track record into Neon (both branches)

data/, reports/, content/, runs/ are gitignored working artifacts (regenerated each run). The append‑only ledger/outcome_ledger.csv is the tracked source of truth.

Tests

python -m pytest tests/ -q        # bash / CI (matches .github/workflows/ci.yml)

python -m pytest tests/ -q        # PowerShell

Tests live in tests/ (run from the repo root); tests/conftest.py applies the subpackage sys.path shim, so the relocated modules' flat imports resolve under pytest.

The full report pipeline is documented in .claude/skills/mvp/SKILL.md. Decision‑support only — not regulated financial advice; no auto‑trading; no auto‑publish (a human approval gate gates every report before it goes live on the site).