Consolidate AutoPredict around an auditable evaluation-to-shadow path

[howdymary]

Objective

Make AutoPredict a trustworthy forecast-evaluation and shadow-deployment framework with one canonical path:

record/ingest -> validate -> evaluate -> report -> shadow -> promote

The product should prove that a user-provided forecast adds held-out value over the market baseline before real-money execution is exposed.

Prioritized issue groups

P0 — repository and evaluation integrity

• AP-001/AP-002: remove broken legacy backtest surfaces and the root package marker.
• AP-003: converge overlapping CLI and compatibility entrypoints.
• AP-004: add compile, source-test, installed-wheel import, format/lint, and type gates.
• AP-005: define one versioned dataset contract and route evaluation through one engine.

P1 — credible end-to-end product

• AP-006: replace the no-op paper loop with real shadow execution.
• AP-007: align live-product messaging and fail closed.
• AP-008/AP-009: make risk direction-aware and persist/reconcile live order state.
• AP-010: build a first-party point-in-time Polymarket recorder and deterministic replay.

P2/P3 — defensible learning and product clarity

• AP-011–AP-013: aggregate out-of-fold evidence, minimum event counts, uncertainty, and multiple-testing protection.
• AP-014: replace optimistic passive-fill assumptions with event replay and sensitivity bands.
• AP-015: describe the current capability as recalibration/parameter search until a forecast-provider interface exists.
• AP-016: rebase maturity and roadmap claims on measurable release gates.

Implementation sequence

1. Repository cleanup and release-integrity CI.
2. Canonical schema and evaluation CLI.
3. Forecast-provider boundary.
4. Polymarket recorder and deterministic replay.
5. Real shadow trading with durable state.
6. Statistical promotion hardening.
7. Live-safety readiness review.

Definition of done

• Source and installed-wheel integrity checks pass.
• One versioned data-to-report workflow is documented and implemented.
• Shadow and future live execution share normalization, decision, risk, and logging paths.
• Promotion uses aggregate held-out evidence against the market baseline.
• Live submission remains gated until reconciliation and operational safety requirements pass.

Detailed evidence, acceptance criteria, and work packets are included in the repository-consolidation specification package on the cleanup branch:

• Issue audit
• PRD
• Technical PRP
• Task plan