Reconstruct what was knowable
Keep research sources, availability times, strategy versions, trigger events and risk decisions. Event time can differ from the time information reached the system. A review must not use later news to rewrite what could have been known. Useful records let a reviewer check whether conditions held, rules ran and uncertainty remained.
Reconcile fills with the account
Drafts, instructions, submissions, partial fills, cancellation confirmations and settlement are distinct states. A cancelled order may have earlier fills; a timed-out request may still have succeeded.
Follow the same operation identity through original fill records, fees, remaining orders and capital reservations before explaining exposure and account changes. Keep simulation records separate from live venue receipts.
Review access is not training consent
The design lets people use the product without contributing private training data. Training participation needs a separate opt-in; payment, trading approval and training consent are distinct. Records enter a versioned dataset only after permission, provenance, quality and timing checks. This is an R&D governance principle, not an available training-consent interface or a completed dataset.
Selection and correction shape the dataset
Check sources, strategy versions, timestamps and outcomes; handle duplicates, missing data, privacy and future-information leakage. Revision reasons, waiting conditions and rejected proposals can explain more than a profit figure. Returns depend on market conditions, exposure and chance, so they cannot directly label reasoning as correct.
Learn offline, evaluate independently
The planned flow uses offline fine-tuning and distillation, then uses an independent, time-separated evaluation set to check task quality, stability and adherence to rules and permission limits. Candidate versions run in shadow mode against a known baseline.
Only qualified versions return to research through a controlled release with rollback. Model versioning, strategy activation and capital authority remain separate.
What each stage should produce
The planned evidence chain is: consent record → samples with provenance and versions → frozen dataset manifest → candidate model → time-separated evaluation report → shadow observations → release and rollback records. Label real usage and simulated cases separately. These are required deliverables for the learning process, not a claim that every artifact already exists.
What each stage should produce This is the planned learning cycle. Consent, model evaluation and release permissions remain separate.
Keep stage labels next to the evidence
Foundational interfaces, bounded research persistence, grid and limited DCA simulations, and membership and Credit simulations have defined-scope evidence. Complete native research workflows, continuous live execution and independent autonomous authorization still require acceptance. Specialist datasets, model weights, evaluated releases and ecosystem feedback remain R&D goals.