Release notes: 2026-05-27 — bench-csv-validity-passed¶
Status: Ready for review
Repo: li-langverse/lic
PR: (branch cursor/bench-validity-scorecard-ce9b)
PH / REQ: PH-7e (tier-1 correctness / dashboard validity)
Author: agent
Summary (one sentence)¶
Tier-1 harness CSV passed now matches the same rtol/atol verify gate as verify_benchmark_results, and horner_pure_li oracle aligns with x=1.1 in horner_core.c.
Agent continuation (required)¶
- Read:
benchmarks/results/README.md(verify columns),benchmarks/scripts/ingest/build_summary.py(validity_for_benchmark). - Run:
cmake --build build && ./scripts/bench-verify-results.sh 1;cd ../benchmarks && python3 scripts/ingest/ingest-lic.sh. - Then: confirm
data/latest/summary.jsonhashorner_pure_liandreduce_sumwithvalidity_status: pass; re-runpython3 scripts/ecosystem-quality-grade.py. - Blocked on: none for harness fix; dashboard ingest needs lic build on agent VM or nightly.
Changed (specific)¶
| Area | What | Evidence |
|---|---|---|
| Harness CSV | verify_csv_rows sets passed from float_close(rtol, atol) not 1-ULP only | benchmarks/harness/bench.py |
| Horner spec | HORNER_BENCH_X=1.1; horner_pure_li iterative oracle (closed form overflows at 5M steps) | benchmarks/harness/reference.py |
| Report pick | primary_report_for_label for iterative + analytical exports | benchmarks/harness/reference.py, bench.py |
| Tests | test_horner_tier1_oracle_matches_bench_x | benchmarks/harness/test_reference_analytical.py |
Not changed (scope fence)¶
- LLVM / MIR codegen for horner loop — not in this PR
benchmarksingest orsummary.json— refresh in follow-up after bench run- 56 failing open PRs / ecosystem scorecard
failed_prs— not fixable here
Breaking changes¶
None.
Security¶
N/A — harness correctness only.
Performance¶
N/A — no kernel timing change; expect dashboard reds cleared after re-ingest (validity gate only).
Downstream¶
| Repo | Action |
|---|---|
| benchmarks | Re-ingest after merge; nightly Linux publish should show 0 tier-1 validity reds |