S1-L2 adversarial challenge plan (depth contract item 8)

Written BEFORE the final numbers land, so the challenge targets are chosen on their importance rather than on which claims turned out convenient to defend.

Rule: each headline claim gets an attempted refutation from the raw files, not a reread of my own prose. A claim that survives is reported with its surviving evidence; a claim that fails is corrected in place, not footnoted.

H1 — "Supply is inverted: thin where demand concentrates (case/workflow, portal)."

Attempted refutation: query the repaired OSS survey for permissive + candidate rows tagged case_workflow or portal. If several genuinely domain-general, maintained, permissive spines exist, H1 is false and the build case weakens (falsifier F2). Status at draft time: pre-repair the file held 2 permissive+candidate case_workflow rows (flowable/flowable-engine Apache-2.0, moasq/production-saas-starter MIT) and 6 for portal — but 4 of the 6 portal rows are admin/doc frameworks (react-admin, refine, redoc, appsmith), not portals in the archetype sense of an untrusted external identity with scoped read. That distinction must be checked, not assumed: if the portal rows are mis-tagged, portal scarcity is WORSE than the count suggests, not better.

H2 — "Rights class should gate disposition, not membership (D-03)."

Attempted refutation: if nearly all copyleft rows are also unusable on other grounds (archived, unmaintained, wrong stack), then keeping them is clutter rather than optionality. Test: cross-tabulate strong-copyleft rows against maintenance signals and archived status.

H3 — "The corpus interface is resolved but unavailable; A20 splits three ways."

Already independently re-derived by the lane owner (ls + foundry status --json, exit 0 despite failure). Remaining challenge: is the foundry-corpus skill definitely the SAME asset the operator remembers as "1.3M repos", or a different index that merely resembles it? SKILL.md quotes the operator's own words about "1.2M repos on the mini", which is strong but is a documented claim, not a live count. Report as: interface resolved to this asset with that quoted basis; counts unverified.

H4 — "User-provided list count = 0."

Attempted refutation: absence found by one agent checking a bounded path set. The claim is falsifiable by any list living somewhere unchecked (client comms were deliberately not read — private, burned credential). Report as: none on disk in the paths checked, with the paths named, and route the request to CENA rather than asserting universal absence.

H5 — "23-25 capability kinds span the demand surface."

Attempted refutation: walk the 17-industry atom table and look for a demanded capability with no kind. Known risk areas: trust accounting, statutory deadline clocks, PHI consent — my own taxonomy calls these custom-delta rather than kinds, which could be a convenient dodge. If they are genuinely recurring across industries, they deserve kinds.

H6 — Survey-quality challenge (turned on my own subagents)

Already executed and acted on, not deferred: three defects found by re-deriving counts from the files (under-tagged scheduling/e-sign, degenerate pattern shape, recall-based commercial claims), all sent for repair. The residual risk is defects I did NOT find. Mitigation: smoke.py checks structure mechanically, and every count in the final report is re-run from the files immediately before it is quoted.