Research

The research has three evidence corpora: four repositories surveyed for workflow, a public-framework corpus read for architecture, and a follow-on corpus separating durable/local products, evaluation/optimization, and interoperability protocols. The corpora have separate methods and share one pattern-catalog.md.

Read in this order

Document	What it answers
`repos/axiom-mesh.md`	Doc-authority + prompt library. 2026-03→04, .NET
`repos/hexguard.md`	Scoped instructions + phase checklists. 2026-06→07, Angular/.NET, 105 packages
`repos/hexguard-templates.md`	Spec-first + decision procedure. 2026-07, reference apps
`repos/rift-forge.md`	Skills + mechanical gates + concurrency. 2026-07→08, 401 branches
`frameworks/README.md`	Public-framework architecture corpus: question, method, template, and index
`frameworks/synthesis.md`	Six-framework comparison · convergences · divergences · unsolved boundaries · catalogue reconciliation
`follow-on/README.md`	Follow-on corpus: shared comparison spine · product/evaluation/protocol methods · eight-subject index
`synthesis.md`	The learning curve · convergences · divergences · the 8 failure modes · the maturity ladder · what nobody solved
`pattern-catalog.md`	Canonical definition of every pattern, with source, rung, and target module. The input to Phase 4
`harness-landscape.md`	Phase 3 addendum: what each harness actually reads, measured 2026-08-14. The four primitives, and what changed under the source repos

Reading only the workflow corpus: synthesis.md then pattern-catalog.md. Reading only the public framework result: frameworks/synthesis.md then pattern-catalog.md. Starting follow-on research: follow-on/README.md, then its shared spine and the relevant track template.

harness-landscape.md was added during Phase 3 and amends synthesis §3.1 in place. The source repos were built 2026-03 → 08 and predate the Agent Skills standardization; where a per-repo file describes a portability problem, check the landscape snapshot before acting on it.

Method

Direct inspection of working trees, 2026-08-14, not of the repos’ own descriptions of themselves. Where a repo’s documentation and its scripts disagreed, the scripts were taken as what actually ran — which is how several stated-but-unenforced rules were identified as unenforced.

Each per-repo file follows one fixed template so the four are comparable column-by-column:

Snapshot — scale, dates, stack, measured counts
The setup — files and moving parts
What works — with evidence
What doesn’t — with evidence
Pain points → how they were solved — a table with a held? verdict per row
How to improve it further — ordered by leverage
Extraction verdict — take / take-as-warning / leave

Conventions

Counts are dated. Every number was measured on 2026-08-14 by a named command. They will go stale; that is expected, and the date is what makes the staleness visible (CLAUDE.md — the evidence rule).
Quoted rules are verbatim from the source repo, in quotation marks. Where a repo’s own wording is sharper than a paraphrase, the wording is kept.
Counter-examples are first-class. A pattern derived from a failure names the repo that demonstrated it — hexguard’s 98 release workflows and hexguard-templates’s “placeholder for future additions” are load-bearing evidence, not criticism.
Patterns are defined once, in pattern-catalog.md. Per-repo files cite ids.

What this deliberately does not do

Judge the source repos’ products, architectures, or code quality. The subject is the workflow: how work was proposed, decided, executed, verified, and remembered — and which parts of that survived contact with an agent.