docs(design): hypothesis map + structural counterfactual (Phase 26 — FOR REVIEW, do not merge yet)

[bryanmatthewsimonson]

⚠️ Leave this as a draft — it's for your review, not to merge tonight

Per your decision ("safe wins now, design the rest; leave the design docs as a draft PR for morning review"), this PR is design-only and deliberately unmerged. The two capabilities it specifies touch the epistemic firewall and need your sign-off before any code.

What

Two normative design docs (+ a ROADMAP Phase-26 stub) for the speculative half of the corpus-analysis review. Both are bounded by the constitution: no fused score or probability — structure, not estimation.

• docs/HYPOTHESIS_MAP_DESIGN.md — your "propose hypotheses from multiple points of view" ask. Competing answers to the case question, with captured claims (kind-30040) attached as supporting / undermining structural evidence, rendered side-by-side, model never picks a winner, no score/probability (grep-tested, like CASE_SYNTHESIS §5). Promotes the existing synthesis positions; local-derived first (the dossier posture); LLM edge-suggestion via the codified gate → ground → filter → human-accept firewall. §6 enumerates the three ways it could cross a red line (a fused hypothesis score, a support-count scoreboard, the model picking a side) and the guard for each.
• docs/COUNTERFACTUAL_DESIGN.md — your "Monte Carlo / if a claim were true or false" ask, reframed to clear the firewall (your decision: structural, counts not probabilities). A deterministic "what depends on this claim" trace over the existing case graph — "removing this claim dissolves 1 contradiction knot · leaves 2 claims with no support · drops proposition P's attestation origins 3→2" — every number a count with its derivation, never a probability. No LLM, no wire kind. §4 refuses the probabilistic version explicitly: that's a separate, constitution-amending decision (the project's answer to Rootclaim-style scoring — six Bayesian analyses spanning 23 orders of magnitude — is a deliberate no), flagged not taken.

Each doc ends with open questions for you (wire kind vs local-derived; whether even a neutral per-section count is too scoreboard-like; inline vs dedicated surface for the trace; whether to build the negate variant).

Context (already shipped, separately)

The safe wins from the same review already merged tonight: brief download (#166), provenance surfacing (#166), cross-article links + vocabulary clarity (#167).

Wire / code

None — docs only.

🤖 Generated with Claude Code

https://claude.ai/code/session_017SZCGu5qxuF2QCuGyJrNcH

---

Generated by Claude Code

[bryanmatthewsimonson]

Superseded by #169 (H.0) — the same design docs, post-review, with the approval header and decisions recorded. This copy was explicitly FOR REVIEW / do-not-merge.