plasm-catalog-adversarial-verify

Plasm catalog adversarial verify

Programming agent certifies catalog machine-spirits (CGS/CML + teaching + discovery seating). This is outside CUGA / ladder harnesses — do not treat ladder pass rates as catalog certification.

Mandatory tiers (full cutover)

Complete all tiers. Do not stop after schema validate or seeded HTTP execute.

1. Schema / cardinality

• plasm-cgs schema validate on the catalog (hard-fail dual-wire pagination, missing blocks).
• List cardinality: ≤1 query + ≤1 search per entity (no twin-query heresy).

2. Teaching + dry/live (seeded or execute open)

• Open an execute / MCP session that seats the entities under test.
• Audit the teaching card: auth gates, shelf discriminants (e.g. LikedSong vs Song), critical mutators/polarities expressible.
• Dry (plasm plan / ?mode=plan) then live a minimal program against Hermit or the real backend.

3. Semantic auto-seed (mandatory)

Product path — not optional, not a harness shim:

1. Start plasm-mcp (or hosted equivalent) with catalogs packed and PLASM_DISCOVERY_SEMANTIC_AUTO_SEED=1 plus OPENROUTER_API_KEY (same env family as CUGA treatment / intent-discovery docs).
2. Confirm auto-seed host shape: tools/list exposes plasm_context / plasm / plasm_run and omits discover_capabilities; plasm_context session_mode: "new" rejects seeds.
3. Call plasm_context with session_mode: "new" + intent only (no seeds) for representative intents mined from eval/cases.yaml and ladder-relevant scenarios.
4. Assert:
   • Teaching card seats the right entities (auth co-seed, shelf discriminants, federated supervisor passwords when AppWorld).
   • No hard_miss / unjustified clarify / brand-lock routing errors.
   • Critical mutators/polarities remain expressible on the card.
5. Dry (then live when backend is up) a minimal program from that card using taught e# forms — not hand-invented seeds.

Fix CGS/discovery when seating is wrong (discovery.names, descriptions, seed_class, co_seed_with, shelf copy). Do not paper over with harness hacks, forced seeds, or prompt-only patches.

Brand-lock / named-catalog matching must be whole-token (short ids like phone must not fire inside headphones).

4. Evidence

Per API, write/update apis/<api>/e2e/adversarial-verify-<date>.md including a Semantic seed section:

• pass/fail
• intent used
• what was seated (e# → entity)
• dry verdict / program used

Roll up multi-API work in apis/<family>/e2e/adversarial-roll-<date>.md with a Semantic seed column.

Anti-patterns

• Certifying from manual {api, entity} seeds only.
• Skipping semantic auto-seed because “schema validate passed”.
• Teaching non-executable sugar ([…] legacy projections) without noting runtime rejection.
• Dual-path shims “for later” — full cutover.

Related

• Authoring core: plasm-authoring
• Intent discovery / auto-seed: repo docs/intent-discovery.md
• E2E transport loop: plasm-catalog-e2e-test