docs(phase-14/39): specialist pool, bias mitigation, calibration set

phases/14-agent-engineering/39-reviewer-agent/docs/en.md

@@ -74,6 +74,20 @@ python3 code/main.py
 
 Output: two review reports written to disk and a console table of dimensional scores.
 
+## Production patterns in the wild
+
+The receipts: Cloudflare's April 2026 AI Code Review system ran 131,246 review runs across 48,095 merge requests in 5,169 repos in 30 days. Median review completed in 3 minutes 39 seconds. Up to seven specialist reviewers (security, performance, code quality, docs, release management, compliance, Engineering Codex) ran in parallel under a Review Coordinator that deduplicated findings and judged severity. Top-tier model reserved exclusively for the coordinator; specialists ran on cheaper tiers.
+
+Four patterns make this work at scale.
+
+**Specialist pool, not one big reviewer.** One reviewer with a 5-dimension rubric works for solo repos. Once the codebase has security-critical, performance-critical, and docs surfaces, split into specialists with smaller prompts. The coordinator does deduplication; the specialists never run the full rubric. Model-tier separation falls out: cheap specialists, expensive coordinator.
+
+**Bias mitigation as design requirement, not optimization.** LLM judges show four reliable biases (Adnan Masood, April 2026): position bias (GPT-4 ~40% inconsistent on (A,B) vs (B,A) ordering), verbosity bias (~15% score inflation toward longer outputs), self-preference (judges prefer outputs from the same model family), authority (judges over-rate references to known authors). Mitigations: evaluate both orderings and only count consistent wins; use 1-4 scales that explicitly reward conciseness; rotate judges across model families; strip author names before scoring.
+
+**Calibration set, not vibes.** A 10-20 task historical set with known correct verdicts. Run the reviewer over it on every prompt change. If agreement with the historical record falls below 80%, the rubric needs revision before the reviewer ships. This is what every team eventually rediscovers; better to start with it.
+
+**Hybrid norm with the gate.** Verification gate (Phase 14 · 38) handles the deterministic checks (did acceptance run, did tests pass, did scope hold). Reviewer handles the semantic checks (was this the right work, are assumptions documented, is the handoff usable). Anthropic's 2026 guidance is explicit on this split: don't ask the reviewer to redo what the gate already proves.
+
 ## Use It
 
 Production patterns:
@@ -110,6 +124,14 @@ The reviewer is the second pair of eyes the workbench grows when humans cannot d
 
 - [OpenAI Agents SDK handoffs](https://platform.openai.com/docs/guides/agents-sdk/handoffs)
 - [Anthropic Claude Code subagents](https://docs.anthropic.com/en/docs/agents-and-tools/claude-code/sub-agents)
+- [Cloudflare, Orchestrating AI Code Review at Scale](https://blog.cloudflare.com/ai-code-review/) — 7-specialist + coordinator architecture, 131k runs / 30 days
+- [Agent-as-a-Judge: Evaluating Agents with Agents (OpenReview / ICLR)](https://openreview.net/forum?id=DeVm3YUnpj) — DevAI benchmark, 366 hierarchical solution requirements
+- [Adnan Masood, Rubric-Based Evaluations and LLM-as-a-Judge: Methodologies, Biases, Empirical Validation](https://medium.com/@adnanmasood/rubric-based-evals-llm-as-a-judge-methodologies-and-empirical-validation-in-domain-context-71936b989e80) — the 4 biases and mitigations
+- [MLflow, LLM-as-a-Judge Evaluation](https://mlflow.org/llm-as-a-judge) — production tooling for separated builder/evaluator
+- [LangChain, How to Calibrate LLM-as-a-Judge with Human Corrections](https://www.langchain.com/articles/llm-as-a-judge) — calibration-set workflow
+- [Evidently AI, LLM-as-a-judge: a complete guide](https://www.evidentlyai.com/llm-guide/llm-as-a-judge)
+- [Arize, LLM as a Judge — Primer and Pre-Built Evaluators](https://arize.com/llm-as-a-judge/)
 - Phase 14 · 05 — Self-Refine and CRITIC (single-agent self-review baseline)
+- Phase 14 · 30 — Eval-driven agent development (calibration set generator)
 - Phase 14 · 38 — the verification gate the reviewer reads
 - Phase 14 · 40 — the handoff packet the reviewer report feeds