[feat] Frontier-tracking: distill authoritative agent-engineering sources + experiment to extend them

[wimaan3]

Problem / motivation

borromeanRings's value is being the enforced distillation of best practices for building software with AI agents. But that frontier moves fast and is spread across providers' docs and courses (Anthropic, OpenAI, Google, …), research papers, and reputable OSS harnesses. Today we absorb this ad hoc — the AI-Fluency 4D work (ADR-0020) and the agent-skills evaluation were both one-off distillations. To stay the best way to build software with agents, this must be a repeatable, evidence-backed process.

And distillation alone isn't enough: copying published guidance makes us a follower. The differentiator is experimentation — validating which practices actually improve agent-built software, and discovering improvements the sources missed.

Proposed change

A defined pipeline (process + supporting skill/docs) with five stages:

1. Source registry — a curated, versioned list of authoritative sources (provider docs/courses, key papers, strong OSS harnesses), so we notice when they change and can re-distill.
2. Distillation — extract candidate practices and classify each: enforceable invariant → candidate check; agent guidance → candidate skill/playbook; vocabulary/principle → docs/ADR. (Worked examples: AI-Fluency → ADR-0020 + 5 skills; the agent-skills eval.)
3. Experimentation — don't adopt on authority. Run controlled evals to test whether a practice measurably improves agent output, and probe for gaps the source missed. Produces evidence — consistent with borromeanRings's thesis of auditable, evidence-backed engineering.
4. Integration — graduate only validated findings into the harness via the normal spec → ADR → gated-PR path. Experimentation gates adoption (no cargo-culting).
5. Provenance & licensing — every adopted practice traces to its source (registry → ADR trail); restrictively-licensed sources are re-authored + attributed (generalize ADR-0020's CC BY-NC-SA handling into a standing rule).

Scope / non-goals

• Distinct from the deep-research enhancement (ADR-0014): that improves the governed agent's research; this is borromeanRings's internal R&D to improve itself ("recursive, but shipping" — Manifesto).
• Not a general scraping/research product (that's a separate product built with borromeanRings).
• Never adopt-by-authority: provider claims are hypotheses until an eval supports them.
• Respect source IP — re-author and attribute; never relicense incompatible content into the repo.

Relationship to #79

The app-type capability-profile system (#79) is the consumer of this process: this pipeline discovers and validates how best practices change; #79 is where validated changes land — as versioned profile/playbook updates rather than ad-hoc rewrites.

Alternatives considered

• Keep doing it ad hoc — rejected: inconsistent, lossy, no provenance, no experimentation; we drift behind the frontier.
• Distill-only (skip experimentation) — rejected: importing practices that may not help; loses the "discover what they missed" upside.
• Automate the whole pipeline now — premature: start with a documented process + source registry + the licensing rule; automate once the manual loop is proven.

Fit with the roadmap

A process-level epic, partly already happening manually. Lean start: (1) a docs/ source registry + the re-authoring/licensing rule + a lightweight distillation→ADR workflow; (2) an eval substrate for the experimentation stage; (3) broaden sources; consider automation. Sequence after current increments are proven.

Open questions

• Cadence/trigger: manual, scheduled, or on major provider releases?
• Eval substrate: how to measure "this practice improves agent-built software" reproducibly? (The hard part — ties to evaluation work.)
• Registry home: a doc, a skill, or structured data?
• Conflicting guidance across providers — adjudication (experiment wins)?