Controlled context
Give facts, policies, and sources explicit identities so the model reasons from an inspectable information base.
AI · FINANCE · INSTITUTIONAL DESIGN
The Harness Ladderfrom capable models togoverned institutions
Six experiments in reliability, authority, deliberation, action, adaptation, and evidence.
A pedagogical excursion into the architecture that makes AI institutionally usable. Explore how controlled data, tools, permissions, validation, and provenance turn model intelligence into accountable financial processes.
The pedagogical objective
Intelligence needs
an architecture.
An AI harness is the engineered environment surrounding a model: controlled data and context, deterministic tools, output contracts, validation, permissions, workflow, memory, escalation, and observability.
The series makes these responsibilities visible. A model may interpret a transaction or recommend a decision; the surrounding system determines which evidence is authoritative, which checks must pass, and whether the recommendation may become an action.
Deterministic verification
Assign exact calculations, reconciliations, and hard constraints to software that checks the model independently.
Explicit authority
Represent decision rights and escalation rules separately from the model’s confidence or recommendation.
Preserved disagreement
Retain specialist perspectives, minority arguments, and uncertainty within a structured deliberative process.
State-dependent autonomy
Tighten permissions as conditions deteriorate and require sustained evidence before relaxing controls.
Traceable evidence
Preserve provenance and distinguish verified facts, management assertions, contextual signals, and manipulation.
A deliberate learning progression
Each layer adds
responsibility.
Begin with the reliability of an accounting answer. Add authority, deliberation, action, and adaptation. Then move upstream to the evidence on which every judgment depends. The six layers are complementary dimensions of institutional design.
00
Reliability · Corporate accounting
Can we verify the process behind the answer?
Pedagogical objective
01Distinguish a correct-looking answer from a controlled, auditable financial process.
Authority · Banking regulatory approvals
Who may turn a recommendation into a decision?
Pedagogical objective
02Separate analytical capability from delegated institutional authority.
Deliberation · Central-bank monetary policy
How can specialists disagree responsibly?
Pedagogical objective
03Discipline evidence and procedure while preserving minority arguments and uncertainty.
Action · Institutional trading desk
Which proposals may become executable orders?
Pedagogical objective
04Separate freedom to reason from permission to act through an independent execution boundary.
Adaptation · Bank liquidity crisis
How should permissions change during a crisis?
Pedagogical objective
05Make autonomy depend on institutional conditions, with rapid escalation and disciplined recovery.
Evidence · Syndicated-credit committee
What deserves to count as decision-grade evidence?
Pedagogical objective
Distinguish verified facts from assertions, forecasts, signals, and adversarial instructions.
The conceptual synthesis
From model
to institution.
The companion paper connects the six experiments into a coherent framework. It explains how the architecture around a model governs what it may see, recommend, decide, and do—and what the institution is entitled to treat as knowledge.
Companion paper · NB00–NB05
The Harness Ladder
Six Experiments in Reliability, Authority, Deliberation, Action, Adaptation and Evidence
Each notebook section explains its place in the ladder, implementation approach, financial setting, and the relevance of harnessing. The paper also examines control design, failure modes, evaluation, and three analogous applications per layer.
The broader design discussion extends beyond the compact mechanisms implemented in the notebooks, helping readers transfer the lessons to other financial workflows.
The computational laboratories
Six notebooks.
One connected journey.
Each notebook contains ten code units with pedagogical explanations. Read the financial case, inspect the intermediate records, and follow the boundary between model reasoning and system control.
NB00–01 · Foundations
From a reliable answer to legitimate authority.
Corporate accounting
The Reliability Harness
Meridian Analytics plc purchases a three-year cloud analytics subscription with an implementation fee. The same model analyzes the transaction first without a harness and then with structured data, a synthetic accounting policy, deterministic calculations, a JSON contract, reconciliation checks, and provenance.
Learning objective. Distinguish a correct-looking answer from a controlled, auditable financial process.
Inspect the journal entry, validation results, system disposition, and audit record.Banking regulatory approvals
The Regulatory Harness
NovaBank Digital seeks authorization to provide wholesale settlement and liquidity-management services. Strong prudential metrics coexist with cloud concentration and interconnectedness concerns. Deterministic tests and an authority gate require human committee escalation when the specified joint triggers are met.
Learning objective. Separate analytical capability from delegated institutional authority.
Inspect prudential tests, the recommendation, mandatory escalation, and the downstream draft letter.NB02–03 · Judgment and action
From structured disagreement to controlled execution.
Central-bank monetary policy
The Monetary Policy Harness
A fictional central-bank board chooses a 25-basis-point increase, a hold, or a 25-basis-point reduction. Seven specialists examine common evidence from distinct perspectives. Validated memoranda, a Deliberation Ledger, and a dissent map support board synthesis without a mechanical majority rule.
Learning objective. Discipline evidence and procedure while preserving minority arguments and uncertainty.
Inspect evidence references, the distribution of views, preserved dissent, and the policy statement.Institutional trading desk
The Trading Desk Harness
Seven specialists contribute to a desired portfolio in a synthetic market. Deterministic controls apply restricted-security rules, position limits, minimum cash, turnover constraints, and liquidity participation limits. Permitted orders produce simulated fills and profit and loss; adverse results can activate a kill switch.
Learning objective. Separate freedom to reason from permission to act through an independent execution boundary.
Inspect proposed versus permitted allocations, control results, simulated fills, and the kill switch.NB04–05 · State and evidence
From changing permissions to the foundations of knowledge.
Bank liquidity crisis
The Adaptive Crisis Harness
Meridian National Bank experiences a synthetic 30-day liquidity episode. A deterministic state machine governs NORMAL, WATCH, STRESS, and CRISIS. Eight specialists provide advisory assessments on milestone days. Permissions tighten as conditions deteriorate; hysteresis delays relaxation, and crisis exit requires simulated human approval.
Learning objective. Make autonomy depend on institutional conditions, with rapid escalation and disciplined recovery.
Inspect the governance trajectory, autonomy charts, permission matrix, and Crisis Governance Report.Syndicated-credit committee
The Epistemic Harness
Orion Infrastructure Holdings requests a USD 2 billion syndicated credit facility. An Evidence Ledger preserves source identity and reliability. Contradiction checks and quarantine rules govern the evidence supplied to seven specialists and a credit committee. A final gate can require further diligence despite a favorable recommendation.
Learning objective. Distinguish verified facts from assertions, forecasts, signals, and adversarial instructions.
Inspect source classifications, contradictions, quarantine, the final gate, and the diligence request.The principle across all six layers
The model
reasons inside
the institution.
The model provides intelligence. The harness determines how that intelligence becomes an accountable institutional process.
Capability and reliabilityA fluent answer still needs independent checks, controlled inputs, and an audit trail.
Judgment and authorityA useful recommendation does not confer the right to approve or execute.
Autonomy and statePermissions change with conditions, while human authority remains explicit.
Information and evidenceSources retain their identity, reliability, uncertainty, and permitted use.
Begin the excursion
Read. Run.
Inspect. Adapt.
For financial practitioners, executives, students, researchers, and builders.
- Read the paper’s opening framework, then work through NB00–NB05 in order.
- Use a notebook’s Open in Colab link and save your own copy.
- Add
ANTHROPIC_API_KEYto Colab Secrets and grant notebook access. - Run the ten code units sequentially, examining intermediate records and governance gates.
- Change a synthetic input or threshold and observe the effect on conclusions, permissions, and dispositions.
The notebooks configure claude-sonnet-4-5-20250929. An Anthropic account with model access is required; API usage may incur charges. Adjust the centralized MODEL setting if needed. Model responses can vary across runs.