"Repo reconstruction benchmark ladder"

Repo reconstruction benchmark ladder

Progressive evaluation tiers for retrieval-first, multi-shard repository reconstruction campaigns. Machine contracts live under contracts/orchestration/repo-reconstruction.schema.json and are listed in contracts/index.yaml.

Tiers

TierFocusPrimary KPIs (examples)
issue_repairSingle defect or small patch setPatch applies cleanly; targeted tests pass; no regression on stated paths
subsystem_regenOne bounded module or feature sliceBuild + scoped test suite; docs facts consistent with code
crate_regenFull crate boundarycargo check/equivalent; integration tests for public API
repo_regenWhole repositoryFull CI ladder; cross-crate invariants; verification evidence stored

Gating

  • Advance tiers only when the prior tier’s KPIs meet rollout thresholds for your environment (latency, cost, and trust boundaries are deployment-specific).
  • Prefer retrieval-grounded artifacts (shard briefs, symbol graph, verification evidence) over monolithic prompts; see mens-training-data-contract.md for opt-in training lanes.
  • Remote execution should carry lease and campaign correlation on mesh envelopes where supported; see orchestration-unified.md and ADR 017 (Populi lease / remote execution).

Persistence

Campaign specs, artifact rows, and benchmark KPI snapshots are stored in the orchestrator DB when available (reconstruction_campaign_spec, reconstruction_artifacts, reconstruction_benchmark_kpis in the execution domain schema).