The Bout Gallery

BRIDGE.EXE

Bass: the unplayed chair vs Challenger: a real model seat in the Hive Bridge harness

Not a numbered skills bout — the league's sanctioned opener: bidding, tempo, and the moment the deal got honest, fought through the bee swarm's contract logic.

Load Man

Bass: load every skill's full file at startup (~29,092 est. tokens, 19 skills) vs Challenger: progressive disclosure — name+description at boot, body on match

Frozen bar: startup cost ≤ 30% of Bass and findability ≥ 9/10 on plain-English queries.

4.5% of Bass's startup cost (~1,295 est. tokens), findability 10/10. Round 1 failed 7/10 — most local skills had no standard frontmatter; the 24-skill mirror was the fix and the rematch cleared the bar. Scored by the builder; the card's numbers stand in the record.

Agency Man

Bass: the 28,614-byte persona file vs Challenger: the 1,336-byte Checklist Kid, distilled from it

Frozen bar: can a fraction of the persona carry the same work?

The lab keeps the Kid: a 21× smaller file that holds the shape. Neither arm takes the other's weapon — the verdict is a discipline, not a belt.

Vet Man

Bass: install on stars and README vs Challenger: deterministic static scan — pipes-to-shell, base64, remote eval, exfil patterns

Frozen bar: trap fixture FAILs with ≥ 3 distinct flags; the clean skill PASSes with 0.

Trap FAIL with 6 flags; clean PASS with 0. Per the record: the scanner is a torch, not a gate — read every skill anyway.

Fetch Man

Bass: clone a whole pack, copy folders by hand vs Challenger: the skills CLI, one vetted skill at a time

Frozen bar: all four checks pass and the installed tree holds exactly 1 skill — no pack bleed, no hooks executed.

Exactly one skill landed (tdd), readable, no scripts run at install. Bass loses on granularity and audit trail.

Durable Man

Bass: the buddy's native setup vs Challenger: the Pi Durable pattern — event-logged agent state in SQLite

Frozen bars: kill-survival, fork correctness, scheduled wake, fresh-process recall.

5/5 locally: SIGKILL mid-tool resumed to exactly S1–S6, fork [S1,S2,B3] vs trunk [S1..S6], a stored-sleep WAKE fired 0.26 s after its due — across the kill. Bass's kill orphans the tool subprocess; no wake primitive exists. Adoption = the pattern, not a migration.

Port Man

Bass: tsc — TypeScript 7.0.2 native binary vs Challenger: tsrs, the Rust port (0.18.0, built from a TS 7.1-dev snapshot)

Frozen bars: exact diagnostic fidelity on the fixtures; ≥ 2.0× speed on the big synthetic; peak RSS ≤ 0.70× Bass.

Under card v1 Bass kept the belt on a single fidelity miss: 61 vs 61 diagnostics, same file/line/col/code everywhere — one message's elided type rendering differed. Speed had passed (3.88× big fixture, 2.82× aggregate) and memory had passed (0.324×). Ketan ruled card v2 — tuple-level fidelity makes the rendering a reported note, not a failure — and the belt moved without a re-run. Both verdicts stand in the record.