Fpa research loop

Skill JeffBrines/openfpa/skills/fpa-research-loop

The FP&A toolbelt for AI coding agents. Claude Code or Codex does the thinking; openfpa gives it a tested finance kernel, durable company memory, and a Karpathy inspired research loop that improves forecasts against your actuals. By Guiderail and Jeff Brines.

Install

npx -y skills add JeffBrines/openfpa --skill fpa-research-loop

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

5 stars5 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Use after forecasts have scored actual outcomes and you want the AI to run bounded autonomous champion/challenger research epochs, discard weak candidates, and propose only evidence-backed model promotions.

SKILL.md

3.0 KB, as published. Nobody here has run it

Company Research Loop

Purpose

Run an AutoResearch-style loop against the company's own forecast history. The AI may generate, test, and discard challengers autonomously. Only promotion to the active champion requires human approval.

Memory And State

.fpa/research/objective.yaml: company-specific metrics, weights, hard checks, minimum improvement, and complexity penalty.
.fpa/research/*.epoch.yaml: every hypothesis and evaluated epoch, including discarded candidates.
.fpa/models/registry.yaml: current champion, challengers, retired champions, and human-approved promotion history.
.fpa/index.yaml: rebuildable lexical memory index.
.fpa/context-pack.md: temporary task-specific retrieval output, never canonical memory.

Workflow

Discover the company command. Run openfpa entrypoint-list <company-root> --kind research. Use a registered research runner when one exists.
Retrieve context. Rebuild memory with pyfpa.build_memory_index(".fpa"), then create a context pack for the miss being investigated. Read prior failed epochs before proposing a repeated hypothesis.
Load the objective and registry. The objective is CFO-specific. It should include forecast-error metrics by decision importance, hard accounting checks, a minimum improvement, and a complexity penalty.
Run bounded epochs. Default to at most five challengers in one run. For each:
- state one falsifiable financial hypothesis;
- generate the smallest company-specific change;
- use rolling or holdout periods not used to fit the candidate;
- run every hard check;
- call pyfpa.evaluate_challenger;
- persist the final ResearchEpoch.
Discard autonomously. Mark failed or weak candidates discarded. Preserve their code reference, evidence, metrics, and rejection reason so future agents do not repeat them without new evidence.
Propose the strongest challenger. Register only promotion-eligible challengers. Mark the strongest epoch proposed and explain the objective gain, tradeoffs, complexity cost, and relevant memory.
Promote only after approval. On explicit human acceptance, call pyfpa.promote_challenger, update the epoch to promoted, and save both with explicit overwrite. The prior champion moves to retired history.

Guardrails

The AI may experiment autonomously after the initial architecture is approved.
Never tune or score on the same periods.
Hard accounting and reconciliation checks override metric improvement.
Do not promote a larger model unless improvement exceeds its complexity cost.
No evidence means no experiment; no holdout means no promotion.
Human approval is required only for champion promotion, not each epoch.