Feedback request: structured strategy representation and provenance layer for financial agents

[waswrsis]

Background

I maintain an early-stage open-source project called QST.

QST tries to represent quantitative trading strategies in a structured format that is easier to transmit, validate, store, compare, and process by agents or automated research systems.

It is not a reinforcement learning environment, backtesting engine, trading engine, broker adapter, or replacement for FinRL / FinRobot / AI4Finance components. Its scope is narrower: strategy representation, validation, and provenance.

Core idea

Many strategies are currently stored as Python code, configs, notebooks, or informal descriptions. These formats are flexible, but they are not always easy to validate, compare, audit, or process reliably by agents.

QST explores whether strategies can be represented as structured, canonical, hashable records, with outputs such as validation diagnostics, deterministic canonical JSON, graph / parameter / instance hashes, and provenance metadata for experiments or agent workflows.

Possible integration path

QST could start as a sidecar tool or external adapter, without requiring changes to existing FinRL, FinRobot, or AI4Finance workflows.

A first step could simply be:

strategy idea or config -> QST record -> validate -> canonicalize -> hash -> provenance report

This may be useful for agent-generated strategies, strategy memory, experiment reproducibility, or tracking strategy evolution.

Questions

1. Does this kind of structured strategy representation layer seem useful for financial agent or quant research workflows?
2. Would strategy memory, agent-generated strategy validation, or experiment provenance be a reasonable first use case?
3. Are there existing papers, standards, or systems I should study before continuing?
4. Are there any design risks or wrong assumptions I should be aware of?

Project link: https://github.com/waswrsis/Quant-Strategy-Tokenizer

[HarperZ9]

This does seem useful for financial-agent workflows, especially if it stays as a sidecar record layer rather than trying to become a backtester or execution runtime.

The split I would preserve is:

• strategy_hash: identity of the canonical strategy record itself
• experiment_hash: strategy_hash + data snapshot + backtest/eval adapter + parameters + costs/slippage assumptions + seeds
• agent_receipt_hash: experiment_hash + agent/model/tool versions + prompt or task ref + approvals + produced recommendation

That separation matters because the same strategy can be evaluated against many universes or time windows without pretending the strategy identity and the experiment result are the same object.

For a FinRobot-style integration path, I would start with a read-only tool rather than a runtime adapter:

1. Accept a QST/GKR record or proposed strategy text.
2. Validate vocabulary, ports, parameters, and capability declarations.
3. Return canonical JSON plus graph/parameter/instance hashes.
4. Emit diagnostics such as unsupported_token, custom_token_requires_approval, missing_data_binding, missing_risk_constraint, not_executable_by_adapter.
5. Let the financial agent cite those diagnostics before it writes research notes, backtest requests, or strategy-memory entries.

A useful first acceptance test would be: an agent-generated strategy may be stored as memory only when it has a canonical strategy hash and explicit non-goals; it may be described as "backtested" only when it is also bound to an experiment receipt with data snapshot, date range, costs, adapter version, seed set, and metric definitions.

This is close to the receipt pattern we use in Telos/Gather: keep the canonical object, the evidence binding, and the agent action receipt as separate layers. That prevents a clean strategy representation from being accidentally promoted into a performance or execution claim.

[waswrsis]

@HarperZ9

QST 1.0 Research Branch Publication
Published: July 5, 2026
Branch: research/qst-1.0-agent-provenance
Version: 1.0.0a2
Tag: v1.0.0a2-agent-provenance
Commit: 8dc90e9e9cc9c1fdebbac688dc80c6c887f0d3f0
Branch
Tag
We have completed and published the QST 1.0 agent-provenance research branch.
This candidate develops QST into a deterministic record, evidence, receipt, claim, and
governance layer for financial agents. It includes canonical strategy identity,
receipt-backed admission, content-addressed evidence, authority modes, token-gap
governance, FinRobot-facing read-only tools, report provenance, and a partial Qlib
workflow importer.