Proposal (rationale)

The eight follow-on subjects are not eight interchangeable frameworks. Letta Code, Aider, goose, and Google ADK expose runtime or product mechanisms; Inspect AI and DSPy make claims about evaluation and optimization; MCP and A2A define interoperability contracts. Reusing frameworks/TEMPLATE.md unchanged would force evaluation and protocol evidence into headings such as “the core loop” and “composition”, then make the resulting cells appear comparable in synthesis.

This item defines the method before the first follow-on extraction establishes one accidentally. It preserves WI-010’s successful fixed-template discipline while admitting that separate subjects need separate questions.

Decision

accepted — 2026-08-15. The user directed the follow-on research to start. WI-017 is done, and WI-018 retains this method as the prerequisite for all eight extractions.

Plan

Requirements

• Define a shared comparison spine covering source/claim type, state and identity, external boundary, human authority, durable evidence, operating cost, strongest counter-evidence, and catalogue consequence.
• Define three thin track templates: durable/local products, evaluation/optimization, and interoperability protocols.
• State when evidence must come from implementation or executable tests, when a normative specification is the authority, and how documentation-only claims are labelled.
• Require a pinned SHA, licence, read date, measurement commands, explicit read boundary, and bounded absence checks in every template.
• Define how a cross-track synthesis may say not commensurable instead of forcing a common verdict.
• Decide the directory and index layout without changing the meaning of the existing docs/research/frameworks/ corpus.

Impacts

• New method, templates, and index scaffolding under docs/research/.
• A pointer from research/README.md once the follow-on location exists.
• WI-020 through WI-028 consume the method; no extraction content is written here.

Approach

Start from WI-010’s evidence prompts, not its eight headings. Keep fields that are genuinely common and create track addenda for the rest. Exercise each template against one named question from its planned children without reading the source deeply enough to produce findings; the first actual extraction may propose corrections, but must record the reason.

Acceptance criteria / tests

1. The shared spine and all three track templates exist and name the claim types they admit.
2. Each template requires SHA, licence, date, scope, evidence labels, and counter-evidence.
3. The protocol template distinguishes normative requirements, optional capabilities, reference implementation behaviour, and application policy.
4. The evaluation template distinguishes task definition, execution environment, evidence log, scoring, aggregation, and optimizer feedback.
5. The product template distinguishes conversation memory, recovery state, documentary intent, repository state, and agent-managed long-term memory.
6. rungs check passes and the site reports no broken internal links.

Out of scope

• Performing any of the eight extractions.
• Editing the existing six framework documents or retrofitting them to the new templates.
• Catalogue, module, or CLI changes.

Execution

Completed 2026-08-15 on feature/WI-019-follow-on-research-method.

• Created research/follow-on/ as a third corpus location, leaving the fixed six-framework corpus unchanged.
• Defined one shared comparison spine and separate product, evaluation, and protocol addenda.
• Made claim authority explicit through Normative, Implemented, Executed, Measured, Documented, and Opinion labels, including a cross-track rule that admits analogy and non-commensurability.
• Exercised each template against one accepted selection question without reading the source or creating premature findings; the fit checks are recorded in the follow-on index.
• Validation: node src/cli.ts check — 20 pass, 0 fail; site npm run build — 79 routes; site npm run check — 0 Astro diagnostics and 779 internal links, 0 broken.

Review

Self-review completed 2026-08-15 against every acceptance criterion:

1. Pass. The shared spine exists, and all three track templates name the evidence labels they admit and the subjects that consume them.
2. Pass. Every track composes with the shared Snapshot and counter-evidence sections, which require a full SHA, licence, read date, included/excluded scope, measurement commands, explicit evidence labels, and the strongest falsifying result.
3. Pass. The protocol authority table separates normative requirements, negotiated optional capabilities, reference implementation behaviour, and application policy; requirement levels and independent pins are mandatory.
4. Pass. The evaluation contract has distinct required rows for task/sample, execution environment/tools, evidence log, scorer, aggregation/uncertainty, and optimizer feedback, with a reasoned not-applicable route for evaluation-only subjects.
5. Pass. The product continuity matrix keeps conversation history, recovery state, documentary intent, repository state, and agent-managed long-term memory separate and requires traced write/read and retention boundaries.
6. Pass. Repository gates, site build, Astro diagnostics, and internal-link validation all pass with the counts recorded above.