Phase benchmark

A stopwatch and a set of scales for the policy that most of the native experiments are built on top of: play some complete games, and report six numbers about how they went rather than one number about how they ended.

The intuition

Two policies can average the same score for completely different reasons. One survives a long time while clearing steadily; another lives a shorter life and gets rescued by one enormous chain. On a heavy-tailed score those look the same in a mean, and they behave completely differently when you change anything about them.

So this harness reports the score, and then the mechanism underneath it:

• mean moves — how long the games lasted, which for this game is nearly the same statement as the score;
• clears per move and reveals per move — the flow rates the board has to sustain against the rise (see the family page for why 2.4 and 1.4 are the numbers that matter);
• mean maximum height — how close to the ceiling the policy was living;
• work per move — search cost in the repository's own unit, so a strength change can be separated from a budget change;
• censored games — how many were still alive at the move cap, whose score is therefore only a lower bound.

Every one of those is public information about the trajectory. Nothing here looks under a gray disc or forward in the tape.

How it works

1. Start a headless game from a seed.
2. Ask the shared native behaviour policy for a column. The default configuration is a three-ply search with five imagined chance outcomes per node, a one-million-unit work cap per move, and a large negative value for dying.
3. Play the move, add its cleared and revealed discs to the running totals, record the board's maximum height, and add the search work the move cost.
4. Stop at game over or at the move cap (default 1,000), and mark the game censored if the cap ended it.
5. After all games, print one JSON line: games, mean score, mean moves, clears per move, reveals per move, mean maximum height, work per move, wall seconds, censored count. With --trace it also prints one line per move with the disc, the column, the height, occupancy, cover count, potential and the board.

What happened

No retained result. The experiment index records this source as support-only and repository-verified and states plainly that "no durable standalone result was located"; audit 04 §B.2 lists it among the sources whose question has no lane. This page describes the instrument from its code and quotes no scores for it.

The technical record

Status: support-only — repository-verified. No ledger entry and no task record was located, so no measurement is attributed to this source.

Defaults: 8 games from seed 0x3d700100, --max-moves 1000, --depth 3, --samples 5, --max-work 1000000, --terminal-utility -1000000. The policy is drop7::cfpi::chooseBehaviorAction in src/core/native/public-behavior.hpp — the same shared phase behaviour that several other experiments freeze and build on, including the tail-survival CEM, which records its version as a 165-weight public evaluator inside a depth-3, internal-width-two, three-stratum selective search.

Build it like any standalone experiment: make experiment SOURCE=approaches/baselines-diagnostics/phase-benchmark/phase-benchmark.cpp (see the reproducibility guide; use CXX=clang++).

Two cautions when reading its output. The clears- and reveals-per-move figures are computed as the mean per-game total divided by the mean moves, not as the mean of the per-game rates, so they are not paired per-game statistics. And the default 8 games is a smoke test, not a cohort — benchmarks defines the tiers that count.

Source: phase-benchmark.cpp.

What this taught us, and what is still open

• The vocabulary this project judges policies in — flow, height, work, censoring — is implemented here, and it is deliberately richer than "mean score".
• What that vocabulary said about the shared phase policy is not recorded anywhere in this repository. The instrument survived; its readings did not.
• Any future use should state its seed range and role up front, because this program has no seed-lease bookkeeping of its own and will happily replay any range it is given.