Add two showcases: eval regression demo + de-id/re-id sample results

[Hefrock]

Makes the strongest work legible to someone skimming the repo — no cloning, no Python — and demonstrates the eval-as-CI-gate pattern end to end.

1. agent-eval — worked regression example (skills/agent-eval/examples/)

A support-ticket summarizer, before/after a "be more concise" prompt change that looks like a clear win:

Metric	Baseline	After change
Mean cost	$0.0049	$0.0030 ↓
Mean latency	1312ms	822ms ↓
format pass rate	80%	100% ↑
accuracy pass rate	100%	62% ↓

Every dashboard an engineer usually watches went green — but the change silently dropped an account ID + error code, collapsed reproduction steps, and mangled a timeline. The gate catches all three:

⚠ 3 regression(s) vs baseline (passed before, failing now):
  acc_03: 1.00 -> 0.40
  acc_06: 0.90 -> 0.30
  acc_08: 0.80 -> 0.50
GATE FAILED — exit 1

• results_baseline.jsonl + results_regressed.jsonl (same 20 case IDs)
• README.md walks the scenario with the real output
• CI self-check: a step asserts the regression example still trips the gate, so it can't bit-rot (the step passes when the gate correctly fails)

2. deid-reid-harness — sample results (skills/deid-reid-harness/RESULTS.md)

A verified run rendered for non-runners:

• The privacy-utility frontier (ASCII plot + table): regex-baseline-v0 at 0.449/1.00 vs over-redact-v0 at 0.900/0.60 — opposite corners, neither dominates
• The cross-track headline: 49/50 records (98%, CI 94–100%) remain vulnerable despite a perfect Safe Harbor scrub — 34 re-identifiable, 47 inferable, 32 on both axes
• Track 1/2/3 breakdowns, every figure with a 95% bootstrap CI

Both showcases are linked from their SKILL.md and the root README.

Verification (re-run before merge)

Both showcases were re-verified deterministically, not asserted from memory:

• deid-reid — 15/15 headline numbers reproduce exactly from a fresh pipeline run at seed 20260101 (frontier corners, cross-track 49/34/47/32, inference 0.94, re-id min_k=1 / 6 population singletons, bootstrap CI bounds). test_harness.py (31/31) locks them against drift.
• agent-eval example — 8/8 documented figures match live score_eval.py output (pass rate 16/20, accuracy 62%, cost $0.0030, latency 822ms, the three named regressions), and the gate exits 1 as designed.
• Test suites: scorer 32/32, deid harness 31/31, MCP server 14/14.
• CI self-check step confirms the regression example still trips the gate on every push.

Scope note (good faith): the deid figures are empirical — reproduced from real harness runs. The agent-eval example's per-case scores are illustrative synthetic data — a teaching scenario for a hypothetical summarizer, not a measurement of a real system. What's verified there is the tool behavior (scoring math, per-category breakdown, regression detection, gate exit codes), which is exercised by the run and by the 32 unit tests.

🤖 Generated with Claude Code