Viability controller

A policy with no scoring function at all. It first classifies the situation into one of five modes, then plays the best move for that mode, using exact facts about the position rather than a tuned number.

The intuition

Human players do not evaluate seven columns on a numeric scale. They decide what the position calls for — "I need to open some grays", "I need to dump this stack before the rise" — and then look for the move that does it. This approach makes that literal:

• charge — the board is comfortable; add to the stored structure;
• dig — there are covered discs that can be attacked; attack them;
• release — the board is filling, or a big clear is available right now; spend the structure;
• repair — a specific clog is forming, such as a low number capping a column; fix it;
• emergency — the board is nearly dead; survive.

The appeal is that it removes the thing that goes wrong everywhere else in this repository: a fitted number that is slightly miscalibrated and therefore ranks two columns in the wrong order. There is no fitted leaf here at all.

How it works

1. Build an exact certificate for the position: for all 7 × 7 combinations of "which disc arrives" and "which column it lands in", work out conservatively what would happen — how many discs clear, how many gray discs are damaged or revealed, how many chain waves follow, and whether the column is productive at all.
2. Pick the mode from thresholds that come from the game's mechanics rather than from fitting. As frozen in the source: emergency if the tallest column reaches six, or at most one column is safe, or the last drop before a rise finds 27 cells occupied; repair if clog debt reaches 16 with a capped low column; release if projected occupancy reaches 32, or 30 cells are already occupied, or the known arriving disc can clear at least three discs in at least two waves; dig if any damage to a covered disc is available; otherwise charge.
3. Apply a survival shield that removes actions which the certificate cannot show to be survivable.
4. Rank whatever is left lexicographically on eighteen ordered fields — first criterion first, ties broken by the second, and so on — rather than by a weighted sum.

Everything read is public: the visible board, the next disc, the rise clock and the terminal flag. There is no gameplay history and no fitted scalar.

What happened

It was clearly worse than the simple comparator. On its 32-game Stage A the controller averaged 132,537.1 points and 43.3 moves against fair D1's 172,697.6 and 54.0, and its flow fell as well — 1.511 clears and 0.729 reveals per move against 1.741 and 0.936. Score, survival and flow all regressed, so it went no further.

Both directions of that flow number matter. The controller was built partly to increase reveal flow, through its explicit "dig" mode, and it produced less of it than the policy it was competing with. Deciding to dig is not the same as finding a move that digs.

The technical record

Status: rejected — ledger-recorded. Experiment history §"Explicit reservoir and constructive-cycle policies (rejected)": 32-game Stage A, 132,537.094 points / 43.281 moves versus fair D1 172,697.625 / 53.969; clear and reveal flow 1.511/0.729 versus 1.741/0.936 per move. Corrected 17,000-point scoring, internally consistent per audit 03 rows 4409–4410. Source and artifact SHA-256 values are retained in the ledger.

Frozen configuration, from the source: Stage A seeds 0x3d65c000–0x3d65c01f (32 games), 1,000-move cap, seven successor samples, policy seed VRC1, 30-minute wall limit, 256 MiB RSS limit, eighteen ranking fields. Its admission gate — never reached — required 700,000 mean points, 200 mean moves, 2.20 clears and 1.20 reveals per move, a 120-move lower-quartile lifetime and 24 joint wins.

Thirty-two paired games is a screen, not a confirmation cohort, and the deficit here is large enough that the screen is decisive about the frozen configuration; it says nothing about a different set of thresholds.

Source: viability-reservoir-controller.cpp.

What this taught us, and what is still open

• Removing the fitted number does not remove the problem. Discrete modes with mechanically justified thresholds ranked moves worse than a simple fair search did.
• The gap between intent and execution is where it failed: an explicit dig mode produced fewer reveals per move than a policy with no modes at all.
• The experiment index draws the same conclusion for the family's other option-based work: a persistent option policy is worth revisiting only after a one-step evaluator demonstrates robust score, survival and flow signal.