PRD: Agent-led CEX HFT research pipeline (GP → Factor Bank → MCTS)

[proerror77]

Problem Statement

Monday already has strong building blocks for governed CEX research: authenticated Binance replay materialization, point-in-time dataset admission, typed FactorAst formulas, label-isolated purged walk-forward evaluation, durable checkpoints, sealed-holdout evidence, immutable promotion records, and signed Paper/Shadow deployment gates.

Those components are not currently composed into the required research pipeline. The durable mission execute path constructs a loose ResearchMission from command arguments and immediately runs MCTS as the only supported engine. Both GP and MCTS can generate formulas, no canonical typed Factor Bank connects discovery to combination search, no deterministic Ridge or shallow-tree baselines establish joint predictive power, and no combination walk-forward produces a four-stage strategy artifact.

The existing event backtest can replay Snapshot, L2Update, and Trade events, but it is disconnected from alpha-harness research evidence. Queue position, partial fills, market impact, and true capacity are not modeled. A successful research result therefore cannot currently prove the complete sequence from an Agent-authored hypothesis through governed factor discovery, subset selection, event replay, one sealed holdout, and signed Paper/Shadow compatibility.

The system needs one bounded, deterministic first vertical slice without giving the Agent, GP, MCTS, or research evaluator any trading authority.

Solution

Make the Agent/LLM the research control plane and alpha-harness mission execute the single deterministic execution seam.

The Agent selects one admitted CEX instrument and horizon, emits a content-bound typed Research Mission with falsifiable hypotheses and frozen policies, and schedules the experiment. The deterministic harness then runs:

Typed single-instrument Mission
→ controlled GP formula generation
→ leakage/data-quality and purged walk-forward screening
→ immutable Factor Bank
→ Ridge and shallow-CART baselines
→ Factor-Bank-only MCTS subset selection
→ frozen normalized equal-absolute weights
→ Signal / Sizing / Risk / Execution artifact
→ combination walk-forward
→ event-level L2 replay with honest capability flags
→ final immutable precommit
→ zero-or-one sealed-holdout opening
→ StrategyBundle + PromotionRecord
→ existing signed Paper / Shadow intake

Training/validation evidence and verified signed Paper/Shadow feedback may inform a later Mission. Sealed-holdout evidence cannot alter the active search. A later Mission influenced by an exposed holdout must bind a fresh unopened holdout.

LiveSmall remains disabled.

User Stories

1. As a quantitative researcher, I want the Agent to select one admitted CEX instrument and horizon, so that the first experiment has a bounded and reproducible scope.
2. As a quantitative researcher, I want each Mission to contain typed hypotheses, so that research intent is machine-verifiable.
3. As a quantitative researcher, I want every hypothesis to include observable falsification conditions, so that failed ideas cannot be rationalized after evaluation.
4. As a quantitative researcher, I want Mission inputs and policies content-bound, so that a resumed experiment cannot silently change identity.
5. As a quantitative researcher, I want the Agent to read structured training and validation failures, so that later Missions can avoid known dead ends.
6. As a quantitative researcher, I want verified Paper and Shadow evidence available to later Missions, so that runtime behavior can improve future research.
7. As an auditor, I want sealed-holdout evidence excluded from the active search, so that final evaluation remains honest.
8. As a quantitative researcher, I want GP to generate formulas only from controlled fields, operators, windows, and constants, so that formula search remains bounded.
9. As a quantitative researcher, I want candidate generation to remain label blind, so that future information cannot enter discovery.
10. As a quantitative researcher, I want temporal leakage and point-in-time data checks before scoring, so that invalid factors fail before ranking.
11. As a quantitative researcher, I want numerical and coverage failures recorded explicitly, so that rejected candidates remain useful evidence.
12. As a quantitative researcher, I want purged and embargoed walk-forward screening, so that factor admission reflects causal out-of-sample behavior.
13. As a quantitative researcher, I want accepted factors stored in an immutable typed Factor Bank, so that later search consumes stable identities rather than formulas in motion.
14. As an auditor, I want rejected factors and rejection reasons retained, so that negative research evidence is not lost.
15. As a quantitative researcher, I want deterministic Ridge and shallow-CART baselines, so that joint predictive power is measured before subset search.
16. As a quantitative researcher, I want baseline policies frozen before fitting, so that baseline results are not tuned after validation.
17. As a quantitative researcher, I want MCTS to select only Factor Bank subsets, so that discovery and combination search remain separate.
18. As an auditor, I want MCTS actions limited to add, remove, and swap, so that it cannot generate formulas or tune continuous parameters.
19. As a quantitative researcher, I want fixed normalized equal-absolute weights in the first slice, so that subset evidence is not confounded by weight optimization.
20. As a quantitative researcher, I want deterministic MCTS checkpoints and traces, so that selection can be reproduced.
21. As a strategy reviewer, I want a typed Signal stage, so that factor references, orientation, weights, and thresholds are explicit.
22. As a strategy reviewer, I want a typed Sizing stage, so that position mapping is deterministic and bounded.
23. As a risk reviewer, I want a typed Risk stage, so that exposure limits remain distinct from predictive logic.
24. As an execution reviewer, I want a typed Execution stage, so that venue, order semantics, and cost assumptions are explicit.
25. As a quantitative researcher, I want combination walk-forward metrics, so that factor interaction is evaluated rather than inferred from single-factor scores.
26. As a quantitative researcher, I want the selected strategy replayed against event-ordered L2 data, so that research evidence includes market-state behavior.
27. As an auditor, I want replay capability flags to report missing queue, fill, impact, and capacity fidelity, so that an L2 replay cannot be misrepresented as full L3.
28. As an auditor, I want one final precommit before holdout access, so that the selected strategy cannot change after seeing final evidence.
29. As an auditor, I want at most one sealed-holdout opening for the precommitted candidate, so that retries cannot become repeated testing.
30. As a runtime owner, I want passing artifacts to reuse existing signed Paper and Shadow gates, so that research cannot bypass deployment governance.
31. As a runtime owner, I want LiveSmall rejected, so that this research feature cannot enable real-money autonomy.
32. As an operator, I want one exact-main bounded run and independent immutable readback, so that CI success is not confused with operational proof.
33. As a maintainer, I want no new database or service, so that the first slice reuses existing stores and deployment boundaries.
34. As a maintainer, I want Prediction Market evaluation isolated, so that event-settlement semantics cannot enter CEX return/IC evaluation.
35. As a maintainer, I want RL, multi-agent swarms, distributed DAGs, and continuous joint tuning deferred, so that the first slice remains reviewable and reversible.

Implementation Decisions

• The Agent/LLM is a logical research controller, not an execution authority. Its output has no effect until it validates as a typed Mission.
• The first slice supports exactly one already-admitted Binance Spot or USD-M venue, symbol, and horizon per Mission.
• Mission input is versioned, fetched through an existing authorized artifact path, and checked against an asserted content hash. No new authentication service is introduced.
• Each hypothesis has a stable identity, statement, target and horizon, required feature/template families, falsification tests, and source-evidence references.
• Mission identity binds dataset, snapshot, partition, source revision, feature policy, GP policy, screening policy, baseline policy, subset-search policy, weight rule, evaluation protocol, replay policy, and holdout identity.
• Evidence admission distinguishes training/validation, signed Paper, signed Shadow, and exposed holdout evidence.
• Exposed holdout evidence cannot feed the same Mission or its resumed search. A later Mission influenced by that evidence requires a new unopened holdout identity.
• Controlled GP is the only formula generator in this pipeline.
• GP consumes registered feature fields and a frozen grammar policy. Arbitrary generated Rust or Python code is not permitted.
• Candidate generators receive label-free metadata only. Labels remain evaluator-owned, and sealed rows remain inaccessible before final opening.
• Screening combines provenance, type/AST, point-in-time causality, numerical validity, coverage, and purged walk-forward predictive gates.
• Accepted factors are stored in a create-once typed Factor Bank revision using existing immutable registry and artifact stores.
• Rejected candidates remain in immutable candidate/evaluation lineage with structured rejection codes and evidence hashes.
• Each Factor Bank entry binds its canonical AST and hash, orientation, source features, policy identities, dataset and partition identities, fold evidence, and admission verdict.
• Ridge regression and a depth-bounded shallow CART regression tree are mandatory baselines. Both fit only fold-training rows and are evaluated only on the corresponding validation rows.
• Baseline hyperparameters are frozen in the Mission. No baseline hyperparameter search is performed.
• Both baseline artifacts must validate and the Mission's predeclared baseline sufficiency gate must pass before MCTS starts.
• Baselines are research evidence only in the first slice; they do not determine deployable weights.
• The shared domain-neutral UCT kernel is reused with a new Factor Bank subset adapter.
• MCTS state contains sorted Factor Bank identities and immutable content hashes. Its only actions are add, remove, and swap.
• MCTS cannot emit a Formula AST, invoke GP, change factor orientation, change weights, or tune a continuous parameter.
• Legacy Formula-MCTS checkpoints are never silently reinterpreted as subset checkpoints. They remain readable for history or fail closed at the new seam.
• The first slice derives weights mechanically from frozen Factor Bank orientations and normalized equal absolute magnitude.
• The deterministic strategy artifact contains separately typed Signal, Sizing, Risk, and Execution stages with explicit parent hashes.
• Combination evaluation uses the frozen walk-forward partitions. Model fitting uses training rows; subset reward and selection use validation evidence; sealed rows remain excluded.
• Combination evidence keeps predictive metrics separate from post-mapping trading and cost metrics.
• Event replay consumes a content-bound CEX event tape and the exact strategy artifact. It uses the existing event-level L2 engine rather than creating a second replay service.
• Replay evidence explicitly binds clock semantics, available modalities, book depth, trade-tape availability, and support flags for queue position, partial fills, impact, and capacity.
• Queue position, partial fills, market impact, and true capacity remain false until separately implemented and proven.
• Replay results must pass Mission-frozen gates before a final candidate can be precommitted.
• Final precommit binds every upstream semantic identity, selected subset, fixed weights, strategy hash, combination evidence, replay evidence, policies, source revision, and implementation/configuration identity.
• Precommit makes the search terminal. No further proposal, baseline fitting, subset search, or resume is allowed.
• Sealed-holdout opening is atomic and idempotent for the same precommit. A conflicting or concurrent opening fails closed.
• A passing holdout produces the existing immutable promotion and bundle lineage. A failed holdout produces no promotion.
• The four-stage bundle becomes eligible only for existing signed Paper/Shadow intake. It conveys no automatic activation authority.
• Signed Paper/Shadow feedback may be admitted into a later Mission through existing memory and signature verification.
• LiveSmall remains rejected.
• Existing DuckDB journals, registry revisions, memory records, artifact references, checkpoints, bundles, promotion records, and signed envelopes are reused. No new database or service is introduced.
• CEX and Prediction Market missions, artifacts, evaluators, and receipts remain typed as separate lanes and reject cross-lane input.

Testing Decisions

• The sole top-level behavioral seam is alpha-harness mission execute.
• Tests assert accepted inputs, emitted immutable artifacts, hashes, receipts, terminal state, and fail-closed behavior rather than private helper structure.
• Existing mission-execute snapshot-binding, publish-once, cost-evidence, and sealed-receipt tests are prior art.
• Existing prepared-dataset label-isolation, purged-walk-forward, holdout-mutation, and checkpoint-drift tests are prior art.
• Existing GP determinism and AST-depth tests are prior art.
• Existing UCT checkpoint/restore and deterministic-search tests are prior art.
• Existing event-level L2 replay tests are prior art.
• Existing signed Paper/Shadow envelope integration and LiveSmall rejection tests are prior art.
• A happy-path E2E must emit typed hypotheses, accepted and rejected GP evidence, Factor Bank revision, both ML baselines, MCTS subset and checkpoint, fixed weights, four-stage strategy, combination walk-forward, replay receipt, precommit, and zero-or-one sealed receipt.
• An MCTS counterexample E2E must reject unknown Factor Bank IDs, Formula AST mutation, GP invocation, weight changes, and continuous parameters.
• A holdout-isolation E2E must prove that holdout-only mutation cannot alter upstream semantic identities; same-payload retry returns the existing receipt; conflicting or concurrent access fails; and search cannot resume after precommit.
• A deployment-boundary E2E must prove that the passing bundle works with existing signed Paper/Shadow intake, tampering fails, and LiveSmall remains rejected.
• Prediction Market artifacts must be rejected at the CEX Mission seam.
• The final runtime issue must use a fresh materialization bound to the exact selected source revision and independently read back the immutable result bundle.

Acceptance Criteria

• One content-bound typed single-instrument CEX Mission drives the complete alpha-harness mission execute pipeline.
• The Agent can provide falsifiable hypotheses and admitted prior evidence without gaining evaluator, holdout, deployment, or trading authority.
• Controlled GP is the only formula generator.
• Leakage, data-quality, and purged walk-forward gates produce auditable accepted and rejected factor evidence.
• Survivors are stored in a typed immutable Factor Bank revision.
• Deterministic Ridge and shallow-CART artifacts complete and pass the frozen baseline gate before MCTS.
• MCTS selects only immutable Factor Bank subsets using add/remove/swap and cannot change formulas, orientations, weights, or continuous parameters.
• The first slice uses frozen normalized equal-absolute weights.
• The selected subset emits a deterministic Signal/Sizing/Risk/Execution artifact and combination walk-forward evidence.
• Event-level L2 replay is connected to the artifact and reports unsupported fill/capacity features honestly.
• One immutable precommit terminates search before an atomic zero-or-one sealed-holdout opening.
• Holdout evidence cannot feed the active search or a later Mission that reuses the same exposed holdout.
• Passing evidence materializes valid StrategyBundle and PromotionRecord lineage.
• Existing signed Paper and Shadow intake accepts the new bundle; invalid signatures, identities, or payloads fail closed.
• LiveSmall remains disabled.
• No new database/service, Prediction Market evaluator reuse, RL, swarm, distributed DAG, continuous joint tuning, or automatic live activation is introduced.
• One separate exact-main runtime issue completes a bounded run and independent immutable readback.
• Every direct child closes with its own evidence, followed by a separate parent acceptance audit.

Out of Scope

• Full-universe or multi-symbol simultaneous search.
• More than one venue, symbol, or horizon in one Mission.
• Multi-agent Macro/Micro/Cross runtime orchestration.
• Distributed workflow DAGs, schedulers, or MCTS.
• A new database, registry service, vector store, model service, or message bus.
• Arbitrary LLM-generated source code.
• Python-authoritative research or runtime dependencies.
• RL or offline-RL strategy training and selection.
• Continuous factor-weight, threshold, sizing, or execution-parameter optimization.
• Deep models, Transformer/TCN models, or new ONNX training.
• Full L3 queue-position, partial-fill, market-impact, or true-capacity simulation.
• New collectors, production collector mutation, or historical data reconstruction.
• Prediction Market event-settlement or token-execution evaluation.
• Automatic Paper/Shadow activation.
• LiveSmall or any real-money activation.
• Cleaning unrelated branches, worktrees, or the dirty primary checkout.

Further Notes

• This PRD supersedes the earlier CEX MCTS-as-formula-generator design previously published in Issue PRD: Agent-led CEX HFT research pipeline (GP → Factor Bank → MCTS) #580. GP now owns formula generation, while CEX MCTS owns only Factor Bank subset selection.
• The earlier Issue PRD: Agent-led CEX HFT research pipeline (GP → Factor Bank → MCTS) #580 history remains available in GitHub's edit history; this replacement does not claim that the superseded design was implemented as the target architecture.
• The existing domain-neutral UCT kernel remains reusable.
• Open Issue PRD: unify MCTS search control plane across research lanes #184 still describes a continuous Formula adapter. Its wording must be reconciled separately if this PRD proceeds; this publication does not modify PRD: unify MCTS search control plane across research lanes #184.
• GitHub Issues remain authoritative. No local PRD file is required.
• Child issues are created in dependency order so their human-readable Parent and Blocked by sections match native GitHub relationships exactly.
• Each implementation issue must record its exact origin/main base SHA when work begins.