# Harkonnen special-Karama inspection and forced returns The existing Advanced development power draws one through four random, unseen cards from another player during Bidding. Harkonnen spends its Karama and once-per-game use, inspects the combined hand, and returns the same number of cards. Newly drawn cards may be returned. Larger hands keep their real choice. When the combined hand contains exactly the number owed, every held card must return. The engine completes that return automatically and emits a house-colored notice explaining the lack of a choice. A manual return also announces the completed exchange. The table's existing cosmetic toggle, reduced-motion behavior and short notice queue apply. ## Private saved observation **Your Harkonnen inspection** below the board retains the actual randomly drawn cards, including their readable rules and enlarged inspectors. It identifies the target and turn and describes historical knowledge: the cards may have changed hands since. Only Harkonnen's owning seat receives this record. Allies, the target and other seats receive no inspection record or suspended private auction data. The copied card descriptions are observations, outside physical card custody. They persist through ordinary returns, later card movement and JSON recovery. An internal integrity receipt binds the original card faces and provenance; changed history is rejected before disclosure. The receipt is not projected. Existing forced-return saves without a recorded draw retain a truthful snapshot of the cards held before returning them; the interface does not claim these were all originally drawn from the target. Legacy exchanges with a real choice keep their selector without inventing prior knowledge. Both manual and automatic returns use the same settlement, restore the suspended decision/response and keep the existing auction funding recovery. Higher-priority interruptions finish first. Restoring a completed exchange must never redraw, repeat the return, spend a second Karama or consume the power twice. ## Authority and limits This composes the existing contract recorded under **Emperor and Harkonnen special Karama checkpoint** in [implementation history](IMPLEMENTATION_STATUS.md): GF9 base rules page 14; November 2020 FAQ page 3 (phase and once-per-game limit) and page 6 (blind draw, inspection, equal return and temporary hand limits). Automatic resolution of a uniquely determined action follows the user's explicit interface requirement. Persisting information already inspected does not reveal the target's current hand. The existing recovery when an exchange removes the only Karama backing an unfunded auction bid remains provisional. This checkpoint does not resolve that ruling or certify all timing and expansion combinations. Advanced and publication gates remain closed. ## Verification scope Focused engine tests cover forced and real choices, saved interruption priority, legacy observations, private projection, rejected-action immutability, physical custody and minimal legal AI continuation. An authenticated in-memory SQL test races two real special-card actions, accepts one, reloads every seat and rejects stale replay without changing the inspection or cards. It also races two restored legacy continuations through the production recovery scheduler: one return is saved with no new random draw. Genuine choices remain waiting without writes. Component tests distinguish drawn cards from legacy held-card history and exclude non-owner output. Required checks and browser evidence are recorded in the private source-bound checkpoint report. These checks establish bounded behavior, not full Advanced acceptance or AI difficulty calibration. Browser QA used genuine Advanced setup in a new isolated development room, then conserved Bidding positions. Playing the real special-Karama control drew and returned four cards with no return prompt; bidding stayed available. The private history and enlarged Shield inspector survived refresh, which retained the exact saved row hash. At a 390-pixel viewport, the 339-pixel history panel fit within the 375-pixel document without horizontal overflow; each inspector button was 44 pixels high. The viewport was reset. Screenshot capture timed out, so these interaction and DOM checks do not establish full visual acceptance.