EvidencePacks without ## Structured Fields keep reappearing — fix the class, not the instances

[explosivebit]

Found 2026-09-08 while bringing docs/ROADMAP.md back to reality after v0.37.0.

The recurrence

Three EvidencePacks currently score CL0 (0.1) because they declare no ## Structured Fields:

EVID-033  R_eff = 0.10
EVID-034  R_eff = 0.10
EVID-035  R_eff = 0.10

EVID-035 is why RFC-004 sits at R_eff 0.10 and shows up as at-risk in forgeplan health.

This is not new. The April 2026 backlog carried an item "F6: Fix evidence blind spots" naming EVID-015 / EVID-025 / EVID-026 / EVID-027 — the same defect, different IDs. Those were fixed. Five months later it is back on fresh IDs.

Fixing the instances has now failed twice. Each round closes the specific packs and leaves the door open.

Why it keeps happening

Nothing stops a pack without structured fields from entering the graph. forgeplan new evidence scaffolds a body, the author fills prose, and ## Structured Fields is easy to omit — the artifact validates, links, and activates fine. The consequence only shows up later and indirectly: the pack silently scores CL0, and whatever it informs drops to 0.1 for a reason that reads as a scoring problem rather than a missing section.

v0.37.0 made this louder — the scorer now warns explicitly (warn: evidence EVID-035 declares no structured fields — scored CL0). But a warning at score time is the wrong end of the pipeline: by then the pack is in the graph and something downstream is already wrong.

Fix direction

A validate-time rule for kind: evidence requiring ## Structured Fields with verdict: and congruence_level: present and recognised. Severity is the interesting question:

• Must would block activate on any pack lacking them. Defensible — CLAUDE.md RED LINE Release v0.2.0 — Phase 3 complete #7 already states these fields are mandatory, and v0.37.0 made the scorer fail closed on their absence. A Must rule would simply make the validator agree with the scorer and the documentation.
• Should is the safer first step if the existing corpus would light up. Measure before choosing: count how many of the 167 existing packs would fail. If it is 3, Must is fine. If it is 30, Should first and a migration.

Whoever takes this: check the count first, then pick the severity — do not pick the severity and then discover the count.

Also fix the three current instances

EVID-033, EVID-034, EVID-035 need their real verdicts recorded. Do not paste verdict: supports in to make the number go up — read what each pack actually measured and write the honest verdict, including weakens or refutes if that is what it found. A pack that scores low honestly is worth more than one that scores high falsely, and this whole class exists because the field was treated as a formality.

Refs: RFC-004, EVID-033, EVID-034, EVID-035, PRD-086, v0.37.0