research
Why a Trading Research System Needs an Evidence Ledger
When models, strategies, and evaluations change over time, an append-only evidence trail helps distinguish genuine prospective learning from a history rewritten after outcomes are known.
Learning creates a provenance problem
An adaptive system changes versions, generates hypotheses, rejects candidates, and sometimes promotes new behavior. If only the latest state is preserved, it becomes difficult to reconstruct why a change happened or what evidence was available when the decision was made. The resulting history can look cleaner than the process that actually produced it, especially when failed experiments disappear.
An evidence ledger addresses this by recording important events as durable facts rather than overwriting one mutable summary. The exact storage design can vary, but the useful properties are stable identity, timestamps, version references, and enough context to connect a decision with the later outcome used to evaluate it.
Append-only records make correction explicit
Append-only does not mean mistakes can never be corrected. It means a correction becomes another event instead of silently changing the original record. This preserves the distinction between what the system believed then and what was learned later. That distinction is essential when evaluating prospective claims because hindsight should not be allowed to alter the evidence that supposedly existed before the outcome.
Hash chaining or signatures can strengthen tamper evidence, but cryptography is not a substitute for good semantics. A perfectly signed record can still be misleading if the metric definition changed without being versioned or if an outcome was linked to the wrong decision. Integrity mechanisms are most useful when combined with explicit contracts for what each record means.
The ledger should record failures too
A research history containing only promoted candidates creates survivorship bias. Rejected revisions, inconclusive samples, data-quality failures, and contradictory outcomes are part of the evidence about how the learning process behaves. Keeping them makes it possible to measure how often proposals fail and whether the system is becoming better at generating useful candidates rather than merely generating more candidates.
Failure records also improve debugging. When a later model behaves unexpectedly, reviewers can trace whether the problem began in data ingestion, hypothesis generation, evaluation, promotion, or execution. Without intermediate evidence, many different failure modes collapse into the same final observation that performance changed.
Public content should consume a narrower projection
The internal evidence ledger can contain information that should never be exposed publicly, including operational state, account details, or control information. A public content system should therefore consume a deliberately sanitized, versioned projection rather than reading the research database directly. This reduces the chance that a publishing feature becomes an accidental path into execution systems.
The public projection can still retain valuable provenance: stable evidence identifiers, delayed aggregate metrics, methodology version, limitations, and the status of an experiment. That is enough to create useful research content without exposing current positions or credentials. The design principle is simple: preserve the evidence needed to support the public claim, and omit everything the claim does not require.