Verifying the app via MCP (golden-expectation suite)

This is a live regression check an agent can run end-to-end through the MCP bridge’s dev tier: a small set of facts computed independently of the app (from public EVE data and the formulas documented below), diffed against what the running app’s tools return. Run it after changes to sde/, market/, or any exposed capability — it’s the fastest way to catch a silent regression in the deterministic engines without a human at the keyboard.

Setup

  1. Enable Start MCP bridge on launch and Expose compute engines to MCP on the Plugins page’s MCP bridge card (or start the bridge and flip both toggles for the session).
  2. Poll <app data dir>/mcp.json for {url, token} (see mcp.md) and configure your MCP client.
  3. Call tools/list and confirm production_profit, fitting_stats, and reprocessing_yield are present — if they’re missing, the dev-tier toggle didn’t take effect (re-check step 1).

Golden facts

1. SDE lookup

sde_type_info for typeId 34 (Tritanium) returns name: "Tritanium", group "Mineral". This is a fixed SDE fact (Tritanium’s typeID has been 34 since EVE’s launch) — any other name/group means the SDE bootstrap or query path is broken.

2. Route jump count

route from "Jita" to "Amarr" returns a jumps count that should be a small integer > 0 and reachable: true (both are empire hubs on the stargate network — always connected). route from a system to itself is always jumps: 0.

3. Appraise arithmetic

appraise with items: [{ name: "Tritanium", quantity: 1000 }] must return a sell value == 1000 × (per-unit sell price from market_price for typeId 34) and a cargo volume == 1000 × 0.01 (Tritanium’s packaged volume is 0.01 m³/unit, an SDE fact) — the appraisal is a pure multiply-and-sum over market_price/sde_type_info, so any drift there is a bug in the aggregation, not the price feed.

4. Production profit — required material quantities

Formula (from CLAUDE.md’s “Profit / EIV” section):

required_qty(mat) = max(runs, ceil(base_qty * runs * (1 - ME/100)))
material_cost     = Σ required_qty * price(mat)
job_fee           = (Σ base_qty * adjusted_price) * system_cost_index * (1 + facility_tax)
profit            = product_qty * product_price − material_cost − job_fee

Pinned golden case — Rifter blueprint (typeId 691), base (ME 0) materials per run: 32 000 Tritanium (34), 6 000 Pyerite (35), 2 500 Mexallon (36), 500 Isogen (37). At ME 10, 1 run, required_qty = ceil(base * 0.9): 28 800 / 5 400 / 2 250 / 450. Call production_profit with { blueprintTypeId: 691, runs: 1, me: 10 } and diff its per-material required quantities against these numbers — a Rust regression test (golden_production_profit_required_quantities in capabilities.rs, gated on EVE_SDE_PATH) pins the same values against the live SDE.

5. Fitting stats — hull layout

An empty [Rifter, Golden Test] EFT string (no modules fitted) still resolves at the all-V skill basis fitting_stats always uses: 3 high slots, 3 mid slots, 4 low slots, 162.5 CPU output (130 base × 1.25, CPU Management V), 51.25 PG output (41 base × 1.25, Power Grid Management V) — those two skills bonus the hull unconditionally, no module needed to “trigger” them. Call fitting_stats with that EFT text and diff layout.{high,mid,low}Slots and layout.{cpu,powergrid}Output against these — pinned by golden_fitting_stats_rifter_layout in capabilities.rs.

Auth boundary (must always fail to reach)

assets, corp_assets, and my_orders must not appear in tools/list under either tier, and a direct tools/call for any of them must return a “method not found”/unknown-tool error, never data. This is pinned structurally by the mcp_never_reaches_auth_gated_data Rust test (capabilities.rs), which fails at compile-test time if any auth-gated capability is ever flipped onto mcp/mcp_dev — an agent verifying the app doesn’t need to check this live, but may as a smoke test.