Validation

What Validation is for

Validation states what the selected Run Output can support. It verifies identity, data coverage, execution fidelity, replay integrity, reproducibility, and the limitations that must travel with the result.

A completed job is not automatically a validated conclusion.

Read the status grammar

Each validation component is explicit:

• Available means the required evidence was retained and read successfully.
• Not applicable means the component does not apply to this result.
• Blocked means a missing input, unsupported route, insufficient sample, failed artifact, or invalid state prevents the claim.

Warnings identify important interpretation limits without converting a blocked component into a pass. Results never fabricates missing evidence or replaces it with a generic success label.

Verify immutable lineage

Confirm that Validation binds the same identities shown in Overview:

• Strategy or Portfolio version and member Configurations;
• parameter vector or Portfolio candidate;
• Dataset, Prepared Data, source, instrument, and time window;
• execution-model id and fingerprint;
• requested and effective fidelity;
• Summary or Robust depth and retained-detail profile;
• parent Study, trial, fold, or promotion identity; and
• artifact checksums and schema status.

Readers use the stored execution model and artifacts. They do not reinterpret a historical run using today's preset or current engine behavior.

Check data and replay coverage

Review coverage, ordering, continuity, and quality for the exact inputs used. Important blockers can include missing windows, stale Prepared Data, unresolved gaps, invalid anchors, sequence regressions, incompatible sources, or an execution model whose data requirement was not met.

Requested fidelity and effective fidelity must agree with the route that actually executed. A higher label is never awarded because compatible-looking data existed nearby.

Check execution assumptions

Validation records the assumptions that turn intent into orders, fills, and positions:

• order and time-in-force behavior;
• activation and same-timestamp ordering;
• partial fills, cancellations, rejections, and contingent orders;
• fees, slippage, impact, and latency;
• position, margin, and sizing policy;
• session, calendar, and instrument rules; and
• queue and liquidity-consumption policy where applicable.

Modeled fills are not live or exchange fills. The result is only as strong as the execution path and retained evidence support.

Validate L2 and L3 market replay

Depth results distinguish price-level L2 replay from order-level L3 replay.

For L3, review separate states for:

• order-book reconstruction;
• venue and source FIFO eligibility;
• source actions, snapshots, resets, gaps, and epochs;
• queue coverage and orders ahead;
• hypothetical maker and taker behavior;
• liquidity consumption and partial fills;
• temporal integrity; and
• limitations or blocked intervals.

Reconstruction can be Available while FIFO queue simulation is Blocked for a non-FIFO or incomplete source. An unrecovered gap prevents publication of that interval as valid L3 evidence.

Order-level replay does not claim hidden liquidity, a full exchange matching engine, calibrated HFT latency, exact fills, or a market response to hypothetical Strategy orders.

Confirm rerun readiness

Rerun checks establish whether the workstation can resolve the exact versions, inputs, parameters, execution model, and artifacts again. Randomized methods preserve seeds and sampler configuration. Promotion from a Study preserves the exact candidate and execution model rather than silently validating a different configuration.

Reproducibility means the same workflow can be reconstructed under retained inputs. It does not promise the same future market outcome.

Follow Study selection evidence

When a Run Output came from a Study, Validation links to its parent candidate. Use Study Results for rankings, parameter surfaces, folds, holdouts, multiple testing, and selection evidence. Use this page for the promoted run's execution, fidelity, lineage, and reproducibility.

Resolve blockers

For each blocker, record the exact message, affected claim, required corrective action, and Run Output identity. Open Outputs for missing artifacts, Diagnostics for failed phases, and System & Jobs for the producing job. Do not promote, deploy, or compare the result beyond the evidence state that remains available.