Meta-Feature Coverage Steering

Preset-driven steering workflows and diagnostics outputs.

Meta-feature coverage steering combines existing missingness, shift/drift, and noise controls into one preset-driven workflow. Use it when you want a reproducible progression through harder regions of the dataset space and a clear diagnostics trail showing what the run actually realized.

Steering resolves onto the same generation surface used by the rest of dagzoo. It is a way to author a staged run, not a separate generator.

When to use

Why it matters for your prior

  • You want one discoverable, deterministic harder-front workflow instead of hand-authoring separate missingness, shift, and noise runs.
  • You need auditable evidence that a run actually moved through the intended curriculum rather than only setting a config knob on paper.
  • You want to reuse current missingness, shift/drift, and noise controls without introducing a parallel curriculum subsystem.
  • You want a benchmarkable preset that can be compared over time with the same smoke-sized CPU workflow.

Operational triggers

  • You want requested-vs-realized steering evidence in diagnostics artifacts.
  • You want one preset that composes missingness, graph drift, mixed drift, and mixture noise into a single run.
  • You need a documented steering smoke workflow before iterating on new presets.

Shipped preset

The built-in preset is anti_memorization_piecewise_v1:

  • missingness_ramp
  • graph_excursion_out
  • graph_to_noise_handoff
  • mixture_noise_ramp

Those stages progressively move the run from light missingness into graph drift, then a graph-to-noise handoff, and finally a mixture-noise ramp. The preset is resolved per dataset with fixed-seed determinism.

This preset is built on top of the existing missingness, shift, and noise controls. The diagnostics output shows the requested stages alongside the realized run summaries.

Preset workflows

Generate smoke run:

dagzoo generate \
  --config configs/preset_steering_anti_memorization_generate_smoke.yaml \
  --num-datasets 25 \
  --out data/run_steering_smoke

The generate smoke preset already enables diagnostics, so the run writes coverage_summary.json and coverage_summary.md alongside the generated data.

Benchmark smoke run:

dagzoo benchmark \
  --config configs/preset_steering_anti_memorization_benchmark_smoke.yaml \
  --preset custom \
  --suite smoke \
  --diagnostics \
  --no-memory \
  --out-dir benchmarks/results/smoke_steering

Benchmark diagnostics artifacts still require --diagnostics. The benchmark summary stays on the existing contract; steering audit evidence lives in the diagnostics artifact pointers rather than a new steering_guardrails field.

What to inspect

  • coverage_summary.json:
    • steering.enabled
    • steering.authoring_form
    • steering.preset
    • steering.stage_count
    • steering.resolution_checks
    • steering.stages[*].requested
    • steering.stages[*].requested_effective
    • steering.stages[*].realized
    • steering.stages[*].metrics
  • coverage_summary.md: condensed requested-vs-realized stage movement for fast human review.
  • Benchmark summary.json:
    • preset_results[*].diagnostics_enabled
    • preset_results[*].diagnostics_artifacts.json
    • preset_results[*].diagnostics_artifacts.markdown

Open the benchmark diagnostics artifact path first, then inspect the top-level steering object in coverage_summary.json. That is the canonical audit surface for requested-versus-realized steering movement.

For field definitions, see output-format.md. For the diagnostics artifact workflow, see diagnostics.md.