Recipe: Grade a Strategy Honestly with VS ungated, VS null

Answer card

To grade a plan honestly, run GRADE plan X OVER <range> FILL <model> with VS ungated, VS null, and BY regime. The two VS clauses subtract the counterfactuals (ungated prices the plan with its regime gate removed, null prices doing nothing) so what remains is judgment, not market drift. Then read the regime cells, not the blended average. Certification is the open recomputation, and self-hosted runs certify the same numbers; a Grade carries the signed Attestation only when the platform's neutral judge runs it against pinned data and a pinned fill model, and without that signature it is provisional.

The recipe

You have a plan. You want to know if its edge is real or if you are staring at a market that went up. Grade it against its own counterfactuals, across regimes, under an explicit fill model.

GRADE plan fade-ladder OVER 2025-01-01..2025-06-30 FILL conservative
  VS ungated
  VS null
  BY regime, session

Read that as four questions in one statement:

• OVER: the date range. If fade-ladder was authored by an LLM, the judge fences this to post-training-cutoff, date-blinded days. Weights leak what code fences can't stop; a date-blinded window is the only window an LLM author can be graded on honestly.
• FILL conservative: the fill model. A quote is not a value, and a Grade refuses to bank expected-$ from any cell flagged extrapolated (priced off an offline ceiling with no live anchor). You cannot hill-climb into a fantasy corner.
• VS ungated: replay the plan with its regime gate removed. This isolates how much of the P&L came from trading only the declared regime versus firing in every regime.
• VS null: replay flat: no position. This is the market's own drift over the window. Beat it or the edge is a beta you paid options premium to rent.

Read the cells, not one average

The BY regime, session clause is the whole point. A single blended number hides where the edge actually lives. The grid does not.

(Illustrative figures, invented for this walkthrough; not a computed result.)

The blended average of those cells might read +0.29R and look like a winner. But the chop · power cell is doing suspicious work, and its support flag says extrapolated: the fill model had no live anchor for those prints, so the Grade already declined to bank that expected-$. The honest read of these illustrative cells: the edge lives in trend and vanishes in chop. An agent could express that by arming the template only in trend regimes and leaving the chop cells on the floor (not a recommendation, just how the grid narrows the choice).

That is what "read the cells, not one average" means in practice. One number is a claim you must trust. A regime grid with support flags is a claim a model can check.

Certified vs provisional

Same numbers, two trust levels:

• Provisional: you ran GRADE in self-hosted Kestrel, or under a trial capability. The math is identical and reproducible; self-hosted Kestrel computes, and certifies, the same numbers. What a provisional run lacks is the signature. Use it to iterate freely.
• Certified: the platform's neutral judge executed the run against pinned data and a pinned fill-model version, then minted the signed Attestation. Only the platform mints that attested receipt. This is the certified Blotter and Grade behind a shareable proof URL: the evidence an agent shows its human before anyone funds anything.

Certification over custody: the platform is the open judge, and what it sells is the Attestation (the signature), not the strategy. You keep the plan; you buy the proof that the grade is honest.

Where this is not the fit

If you want a leaderboard of parameter sweeps, this is the wrong tool: a Kestrel Grade replays committed judgment, it never curve-fits, and it grades judgment, not parameters. And note the status. As of mid-2026: anonymous trial sims, certified Grades, shareable proof URLs, and 402 Offers with Stripe settlement are live; always-on paper presence and the human-signed live path are in build. The free tier needs no signup.

A Grade is never flattering when the syntax forces it to declare its counterfactuals and refuses to bank cash from a cell it could not calibrate.