AI Surrogates in Finance

Constrained Surrogacy Under Governance
Synthetic-First
Auditable · Reproducible
Human-Accountable

This repository is the governed home of AI Surrogates in Finance: a synthetic-first, reviewable laboratory for designing surrogate decision systems in finance under explicit constraints.

The objective is not “a smarter model.” The objective is a defensible controller: a surrogate that behaves like a contract—explicit state, explicit constraints, explicit termination, explicit artifacts—so an independent reviewer can answer: what happened, why it happened, which constraints were active, how close the system ran to boundaries, what is fact vs assumption vs open item, and what required human escalation.

This repo is intentionally explicit about limits: it does not promise alpha discovery, it does not treat backtests as evidence, it does not claim production readiness by default, and it does not claim correctness of external facts.

Core premise: Capability ↑ ⇒ Risk ↑ ⇒ Controls ↑
Read Book (PDF) Open Notebooks Governance Contract How to Use

What this repository is

Surrogates are proxies. Proxies take actions (or drive decisions). Once a model becomes a proxy for a decision, “accuracy” is not the main question. The main question is whether the proxy is reviewable under constraint. This repository treats surrogacy as a professional control problem: governed, bounded, and reconstructable from artifacts.

Repository structure

/book/ — The governing spec: definitions, mechanisms, failure modes, acceptance gates.

/notebooks/ — The governed laboratories: synthetic data, diagnostics, mandatory artifacts.

See Book

What a reviewer must be able to answer

What happened · why it happened · which constraints bound · how close the system ran to boundaries · what is fact vs assumption vs open item · what required escalation.

See Governance Contract

Explicit limits

No alpha promises · no “backtest = evidence” posture · no default production claims · no unverified external facts presented as truth.

Read Disclaimers

The book structure (3 chapters)

This repository is organized to match the book exactly. Each chapter has a corresponding governed notebook in /notebooks/.

Chapter 1 — Why Surrogacy Needs Governance

Surrogacy as a proxy contract: what the surrogate may and may not claim. Enforces “generation ≠ verification”, and makes facts vs assumptions vs open items a first-class output requirement.

Deliverable expectation: a reviewer can reconstruct the logic from artifacts, not from persuasion.

Chapter 2 — Mechanisms: Constrained Surrogate Controllers

Surrogates designed like controllers: explicit state, explicit constraints, explicit routing, explicit termination reasons, explicit escalation paths, and measurable constraint margins (distribution, not point estimates).

Deliverable expectation: the system can explain why it stopped and why it escalated.

Chapter 3 — Evaluation & Acceptance: Reproducible Stage Gates

This repo is not a leaderboard. It is an acceptance framework: compliance distributions across seeds, structural-break stress tests, boundary-time targets, and exportable advance / revise / reject artifacts.

Deliverable expectation: promotion is earned via artifacts and gates, not vibes.

The book structure (3 chapters)

This repository is organized to match the book exactly. Each chapter has a corresponding governed notebook in /notebooks/.

Chapter 1 — Why Surrogacy Needs Governance

Surrogacy as a proxy contract: what the surrogate may and may not claim. Enforces “generation ≠ verification”, and makes facts vs assumptions vs open items a first-class output requirement.

Deliverable expectation: a reviewer can reconstruct the logic from artifacts, not from persuasion.

Chapter 2 — Mechanisms: Constrained Surrogate Controllers

Surrogates designed like controllers: explicit state, explicit constraints, explicit routing, explicit termination reasons, explicit escalation paths, and measurable constraint margins (distribution, not point estimates).

Deliverable expectation: the system can explain why it stopped and why it escalated.

Chapter 3 — Evaluation & Acceptance: Reproducible Stage Gates

This repo is not a leaderboard. It is an acceptance framework: compliance distributions across seeds, structural-break stress tests, boundary-time targets, and exportable advance / do-not-advance decisions.

Deliverable expectation: promotion is earned via artifacts and gates, not vibes.

Notebooks (mapped to the 3 chapters)

The notebooks are not “examples.” They are the operational proof of the book. Notebook names and numbering live in /notebooks/ and should be read as the lab companion to the corresponding chapter.

CHAPTER 1.ipynb

Proxy-contract discipline: explicit separation of facts / assumptions / open items, reproducible synthetic setups, and an early termination + escalation posture that refuses persuasive output when evidence is missing.

Open Notebook

CHAPTER 2.ipynb

Constrained control loops: explicit state machines and routing, constraint enforcement with margin tracking, bounded retries, safe failure, and explicit termination reasons that make “stop” a governed feature.

Open Notebook

CHAPTER 3.ipynb

Evaluation harnesses: compliance distributions across seeds, structural break stress testing, comparative baselines (conservative + robust references), and exportable advance / revise / reject artifacts.

Open Notebook
Notebook invariant: synthetic-first (deterministic seed) · explicit state · bounded loops · explicit termination reasons · facts/assumptions/open separation · export artifacts every run · verification_status="Not verified".

What every notebook enforces (non-negotiable)

This repository follows a strict laboratory contract. These are not “nice-to-haves.” They are what makes surrogacy reviewable.

Laboratory Contract
A surrogate is acceptable only if a third party can reconstruct, challenge, and supervise it from artifacts.
Stop is a feature
Not verified by default
  • Run synthetic-first (deterministic under seed).
  • Define explicit state (TypedDict or equivalent).
  • Implement bounded loops (no unbounded retries).
  • Implement explicit termination reasons (why the system stopped).
  • Enforce facts vs assumptions vs open items separation.
  • Export review artifacts every run, including at minimum:
    • run_manifest.json (seed, versions, config hash, environment fingerprint)
    • final_state.json (what the system believed at the end)
    • decision.json (advance / revise / reject + reasons)
    • risk_log.json (risks triggered + controls applied)
    • artifacts/ folder for generated deliverables
Shared governance spine: generation ≠ verification · facts are not assumptions · scope/boundary control · auditability by design · human accountability.

How to use this repository

Recommended posture:

1) Start with mechanism

Read the relevant chapter in /book/ (mechanism + acceptance rules).

Open Book

2) Run the lab as-is

Run the corresponding notebook(s) in /notebooks/ without edits to reproduce baseline behavior.

Open Notebooks

3) Inspect artifacts, not vibes

Read termination reason, constraint margins, open items, and the stage-gate decision.

See Artifact Minimums

4) Stress structurally

Change regimes, tighten constraints, reduce evidence, and observe which failures emerge and why.

See Constraints & Stops

5) Modify one mechanism at a time

Change a single mechanism, re-run, and compare artifacts. If it “works” only when you ignore gates, it does not work.

See Stage-Gate Discipline

IMPORTANT DISCLAIMERS (read before use)

Educational / Non-Reliance: All materials are provided for educational and research purposes only. Nothing here constitutes investment, trading, legal, tax, accounting, audit, or compliance advice.

Not verified: Unless explicitly stated otherwise in a specific artifact, treat all outputs, claims, calculations, citations, and conclusions as Not verified.

Confidentiality and data hygiene: Do not paste confidential, proprietary, regulated, or personally identifying information into external systems. Use anonymization/redaction and minimum-necessary inputs by default.

No fabricated sources or claims: Zero tolerance for invented citations, performance claims, fees, terms, or consequences. When evidence is missing, the correct output is a verification task list, not persuasive narrative.

Licensing, Governance & AI Use Disclosure

Copyright © 2026 Alejandro Reynoso. All original text, structure, and pedagogical design remain the intellectual property of the author.

License: MIT License. You may use, copy, modify, and distribute provided copyright and license notices are preserved.

AI use disclosure: Generative AI tools were used to assist drafting and editing. Conceptual design, governance decisions, validation, and final approval are performed by the author. Responsibility for interpretation and use remains human-led.

Contact: Alejandro Reynoso · Email: areynoso@yahoo.com · GitHub: alexdibol