[2026 AI] Build an evaluation harness and published benchmarks for every shipped ML model

[Nanle-code]

Objective

Back every model accuracy claim with reproducible evaluation.

Target area

src/ml, src/lib/feePredictor.ts, phishingDetector.ts, liquidityModel.ts

Context

The README claims the fee predictor is '95% within 10% of actual fees' with '< 50ms' latency, but there's no benchmark script or dataset in CI to back it up. Several other models (phishing, liquidity, build, backup) have no stated evaluation.

Acceptance criteria

• A pnpm ml:eval command runs each model on a held-out, versioned dataset and outputs metrics as JSON and markdown.
• Metrics are compared to a naive baseline (e.g. median recent fee), and the model must beat it to ship.
• The README and docs quote only numbers produced by the harness, with date and dataset version.
• A scheduled CI job re-runs evaluation and posts the results as a workflow summary.
• Automated tests cover the primary flow, at least one boundary case, and at least one failure case.
• User-facing documentation or developer guidance is updated, including compatibility, security, or migration notes where applicable.

Pull request requirements

• All continuous integration checks must pass before a pull request may be merged.
• The branch must be free of merge conflicts with the target branch at the time of merge.
• Do not request merge with failing workflows, skipped required checks, or unresolved conflicts.

[drips-wave]

@Nanle-code This issue could not be added to the Stellar Wave Program because the per-repo points budget would be exceeded (25000 / 25000 points already used). The label has been automatically removed.

The budget resets at the end of each Wave. You can review the remaining budget for your repos on the Orgs & Repos page.

ℹ️ Learn more about points budgets