Skip to content

Latest commit

 

History

History
75 lines (59 loc) · 2.66 KB

File metadata and controls

75 lines (59 loc) · 2.66 KB

Context concept coverage v1

Decision

Adopt the benchmark and its label boundary. Do not claim that current context retrieval has solved task-level coverage.

The frozen regression thresholds passed, but selected evidence covered only 6/12 concepts. The result is development evidence from a consumed prospective validation set, not a promotion gate or generalization result.

Identity

  • Harness revision: 3d96f14a5fd7798dd1aca8603ec4a28e4e5d4a69
  • Harness worktree dirty: false
  • Source manifest BLAKE3: 5991d8a643a873ef61d5a4122f52abd7f589a5403d13ab609ebc6b9428e73d9a
  • Concept-label BLAKE3: f496deaadd5098b1faf9f7519cb9e92cdcf14ebf09d73f5beebd675b889c9763
  • Full report BLAKE3: 371ed1a2ac598693fb8884729ff18801d4cf5ec34ac8c54cc19e3093af3d14d8
  • Tokenizer: exact cl100k_base
  • Source budget: 1,200 tokens per task

Accuracy

Metric Result
Candidate relevant-file recall 11/11 (100%)
Candidate concept recall 9/12 (75%)
Selected relevant-file recall 7/11 (63.6%)
Selected concept recall 6/12 (50%)
Concept selection retention 6/9 (66.7%)
Labeled line-anchor recall 19/38 (50%)
Labeled-file precision 7/27 (25.9%)

The difference between 100% candidate file recall and 75% candidate concept recall is the central result: seeing some candidate from every labeled file does not prove that candidate generation reached every required behavior in those files.

Generation missed Flask's run-server-name-ipv6 and run-ipv6-regression concepts and Tokio's blocking-pool-shutdown-regression. Selection then lost Gin's non-bmp-regression plus Express's empty-extension-resolution and trailing-dot-regression, despite candidates reaching those anchors.

Cost

  • Selected source: 4,073 tokens
  • Complete first-response JSON: 6,999 tokens
  • Complete two-turn JSON, including known-hash request and response: 14,756 tokens
  • Dead-end source: 1,703 tokens
  • Exact known-hash resends: 0
  • Estimated repeated-range source: 549 tokens

The four warm context medians ranged from 90.7 ms to 256.6 ms on this host. Cold indexing ranged from 535 ms for Express to 4.78 s for Tokio. These timings describe this machine only and are not acceptance thresholds.

Reproduction

cargo run --release --example representative_benchmark -- \
  --manifest benchmarks/validation.json \
  --concept-labels benchmarks/context_concept_coverage.json \
  --require-concept-thresholds \
  --repos-root target/validation-repos \
  --output target/context-concept-coverage.json

The full JSON report retains per-concept matched anchors, candidate diagnostics, selected evidence, exact token accounting, pinned revisions, and host metadata.