No curve can price a hold

The negotiation layer gives the player real levers: promise a delivery date against the org's expectation band, stake a multiplier on the outcome, renegotiate mid-run, or walk. Every lever is an exploit surface. Before freezing any pricing value, I ran the candidate strategies through a brute-force grid — every fixed policy plus the obvious adaptive ones, 1,000 replications per row — and let the argmax speak.

Exploit 1: lock tight, then renegotiate

The grid's best row under the draft pricing: promise an aggressive date at a high stake multiplier, then renegotiate outward once the stake is locked. Expected standing gain +9.0 per run while every honest row sat near zero or negative. The draft had priced renegotiation by lateness only, so the stretch was nearly free once the multiplier was frozen.

The structural finding: no multiplier curve can price a hold

The deeper problem: holding a deeply padded promise could not be priced by ANY curve on the stake multiplier. At a near-certain date the expected outcome delta is positive and the multiplier is non-negative — so the expected value of holding is non-negative no matter what curve you draw, while every counter-accepting row goes negative. Even a multiplier of zero leaves the hold weakly best.

If a behavior can't be priced through an existing lever, it needs its own price. Shipped: a flat sting for holding at commit, plus a renegotiation stake term scaled by relative stretch. After one tuning iteration the exploit row measures −3.0, and the policy-renegotiation row lands at −1.0 — a usable escape valve, not a dominant strategy.

Exploit 2, found at the next layer: the free outward stretch

Adding the kickoff meeting (the team blurts out its own estimate — the "take" — before the run starts) re-opened the seam in a new shape: promise the org's anchor, then renegotiate to the team's number on day 0 at par stake. Nearly free again, for the same root cause — the base curve priced lateness, the stake term priced only the premium above par, and nothing priced the stretch itself. Fix in the same named lane: a stretch price independent of the multiplier. After it, every take-following row is negative but still far better than holding a doomed date, which is the intended shape — the kickoff ask reads as priced insurance, not an oracle and not a sucker button.

The method note

Both exploits were found by the grid before any human played them, and both fixes were verified by re-running the same grid — the pricing debate became measurement. The acceptance tests now pin the grid itself: no strategy row may dominate the intended play. One wording lesson: the natural clause "the honest strategy weakly dominates" was unsatisfiable as written, because the win probability is structurally zero for every row in this fixture — ties on a dead metric fail any candidate. Acceptance criteria for a balance property have to name the metric that actually discriminates.