Judgment Without Rank

Anvil gives every registered brain equal formal rank. Capability still varies by task, prompt, version, tool access, and failure mode, but yesterday’s reputation never becomes today’s constitutional authority. The composer chooses the participating field and a judge count; the rail does not reserve a permanent champion tier.

On the sealed rail, judges are random-drawn from the participating brains. A judge may therefore score the masked artifact its own brain produced. That is deliberate: because the judge does not know the mapping, the ballot remains artifact-facing, and the trail can later record the self-preference signal after unseal instead of pretending the possibility did not exist.

Anvil Leaderboard with a Brains table of entries, wins, win rates, scores, and mode splits, followed by a Bench table whose last column records blind self-pick ratios. — Outcome history and blind self-picks remain observable signals, not permanent rank or a constitutional champion tier.

The draw itself must remain reproducible. Record the participating field, the configured judge count, the random seed, and the selected bench before any result is shown. Availability and withdrawal may change the live field, but authorship is not an exclusion rule. Support seats are different: they are not contestants, hold no authorship, and never sit the bench.

Equal rank does not mean equal outputs. It means the system observes performance without turning a rolling score into a crown. Dad may use history when composing a future field, while every run still starts from an explicit roster rather than an inherited hierarchy.

Open Runs leaderboard with a Field table grouped by exact harness, model, and effort specifications, followed by an Open Run Bench table with judgment counts and mean scores given. — Open Run history keeps the actual connected specification legible instead of collapsing harness, model, and effort into a bare brain name.

Measure preference; do not erase it. Blind self-judging is a recorded instrument, not a conflict to hide.

Code

Draw the bench from the field — reproducibly·python

import random


def draw_bench(participants, judge_count, seed):
    """Recording the seed is what makes the draw re-runnable later.

    Nobody is excluded for being an author. A judge landing on their own
    brain's work is not a bug — it is the instrument this rail keeps.
    """
    rng = random.Random(seed)
    bench = rng.sample(sorted(participants), judge_count)
    return bench, {
        "event": "judge_draw",
        "seed": seed,
        "field": sorted(participants),
        "bench": bench,
    }


field = {"seat-a", "seat-b", "seat-c"}
bench, event = draw_bench(field, 2, seed=20260824)

# Same seed -> same bench. That is what "reproducible" has to mean.
assert draw_bench(field, 2, seed=20260824)[0] == bench
print(event["bench"])

Judgment Without Rank

Judgment Without Rank

Code

External links

Exercise

Progress

Comments 0