Drop7 Policy Protocol (D7P)

D7P is the standard interface between a Drop7 policy and a benchmark harness, similar to UCI for chess engines. It lets any strategy, in any language, run against the same scripted rounds and report results the same way.

D7P has two equivalent layers:

1. a TypeScript interface used by the in-repository registry (src/bench/policies.ts), and
2. a text wire protocol for standalone processes (src/bench/d7p-server.ts).

Both layers express the same idea: a policy is a deterministic function of the public state.

The public state

A policy receives exactly four things, matching the information boundary in methodology.md:

• the 49 visible board cells;
• the visible next disc;
• the number of moves until the next row rise; and
• whether the game is over (equivalently, whether any column is legal).

A policy must not receive or infer the round id, generator seed, future discs, hidden gray-disc values, score, level, absolute move number, or any history it cannot reconstruct from the public state. In the TypeScript registry, strict policies are handed a sanitized state whose score, level, and move counter are fixed constants; policies that deliberately read level or move number are flagged publicInformation: false and badged "extended state" on the leaderboard.

TypeScript layer

interface BenchPolicy {
  id: string;                    // stable kebab-case id
  name: string;
  family: string;                // approaches/<family> grouping
  description: string;
  publicInformation: boolean;    // strict public-state only?
  slow?: boolean;                // excluded from the default bench suite
  chooseColumn(state: GameState): number | null;
}

chooseColumn returns an internal column index from 0 to 6. Interfaces for players display the same columns as 1–7. Returning null or an illegal column is recorded as an illegal decision and the harness plays the first legal column as a fallback. Policies must be deterministic: same public state, same column. Solver seeds, when needed, are derived from the policy id and are never game seeds.

Wire protocol

The wire protocol is line-oriented UTF-8 over stdin/stdout. The harness sends commands; the policy answers. All tokens are space-separated.

Handshake

-> d7p
<- id name Greedy 1-ply
<- id family heuristic-search
<- id public-information true
<- d7pok

Readiness

-> isready
<- readyok

Setting the position

-> position startpos next 3 rise 5
-> position board 00000000000000000000000000000000000000008888888 next 3 rise 5

• board is the 49-character serialization from the engine (serializeBoard): row-major from the top, 0 empty, 1-7 numbered, 8 solid gray, 9 cracked gray.
• startpos is the standard opening (empty board, one covered bottom row).
• next is the visible next disc, 1-7.
• rise is the moves remaining before the next row rise, 1-5.

Asking for a move

-> go
<- bestmove 4

bestmove carries the internal column index from 0 to 6, or none when no column is legal. User-facing tools add one when they display it. The policy may emit info ... lines at any time; harnesses must ignore lines they do not understand. quit ends the session.

Reference server

Any registry policy can be driven over the wire protocol:

node --experimental-strip-types src/bench/d7p-server.ts --policy expectimax-d2

Scripted rounds

Benchmark randomness is predetermined by the drop7-scripted-round-v1 format (src/bench/rounds.ts):

• discs[m]: the visible disc at move m, indexed by absolute move number;
• latentRows[r][c]: the hidden value of the covered disc in column c of covered-row generation r (generation 0 is the opening bottom row).

The engine's latent-board mode (PlayMoveOptions.latent in src/core/typescript/engine.ts) carries each hidden value with its physical disc through gravity and rises, and a reveal can only produce that predetermined value. The standard suite is Gauntlet 01–08 (src/bench/rounds/), generated from the dedicated 0x5eed**** domain, which overlaps no historical or reserved research seed range.

Scripted rounds are a public playground: they consume no research seed lease and produce no tier evidence. Research-tier claims follow benchmarks.md.

Running the benchmark

npm run bench                  # default policies x all rounds
npm run bench -- --all         # include slow reference policies

Results are written to web/data/leaderboard.json with one replay file per game in web/data/replays/, and render in the web console at /leaderboard.