The Tribunal

Verdict on the build. Jul 30 2026, 2:45 PM CDT.

One judgment seat exists and works. It ruled three times, was graded blind by an independent Fable, and beat both degenerate baselines. Four seats attacked the design, and one found the write path completely broken while an evidence file on disk claimed PASS.

28agents, all explicit model

0inherited, 0 opus

3real rulings on record

0orphaned convenings

What was built

The Tribune of Viability: a Fable-judged seat with a four layer brain, an append only casebook, and eight scripts that turn the convening protocol into two commands instead of thirteen steps. Seats two through five do not exist. Nothing is scheduled.

Layer	Holds	Who writes it
CHARTER	identity, lens, job	Anthony only
BRAIN	researched knowledge	compiler agents
CASEBOOK	rulings and outcomes	the orchestrator
LESSONS	promoted patterns	a consolidator

Graded blind, on mechanism

A Fable that did not write the exam, the key, the seat, or the ruling scored it. Mechanism, not class, because a prior adversary proved class is partly recoverable from writing style alone.

Measure	Seat	Always approve	Always object
Caught, mechanism	4 of 8	0 of 8	0 of 8
False positives	2 of 6	0 of 6	6 of 6
Ruling accuracy	78.6%	42.9%	57.1%
Mechanism gated	57.1%	42.9%	0%

This ruling shows genuine viability reasoning executed through a scar shaped vocabulary that cannot name a failure class it has never been scarred by, with economics as the proven missing class.

Four things that broke, and were fixed

The write path was dead for 24 minutes while an evidence file claimed PASS. A collision fix left a placeholder the write script refused to overwrite, so no convening could ever close. Found by a seat running an open followed by a commit. The race test that proved the collision fix had exercised the open half alone.
The verification accepted garbage. A truncated draft, a double verdict file, and a grocery list reading VERDICT: yes buy the eggs all committed with exit 0.
LC_ALL=C defeated the em dash gate entirely. Any cron or LaunchAgent context.
The atomicity claim named the wrong mechanism. Measured, not asserted: a 1012 byte line gave 0 corrupt of 1800, a 1036 byte line gave 1272 corrupt of 1800.

What this run refuses to claim

Not that the backtest measured anything. The seat caught 2 of 2 non survivors, then declared both as retrieved from its own brain, which was seeded from the same ledger. On the six records where it was genuinely blind there were no real failures to find, and it flagged two live projects as dead.
Not that the exam is clean. Version one was blind classified 12 of 12 from writing style alone. Version two still leaks, and is used only because the surface attack recovers class and cannot recover mechanism.
Not that the seat gets smarter. One lesson exists, marked UNPROVEN, awaiting its control arm.

The gates, each watched firing

lifecycle sweep   canary caught, daemon and Wire correctly ignored
size cap          4,170 of 4,000 rejected
multibyte         one em dash caught by a two count divergence
brain staleness   241 day brain rejected, fresh brain passed
schema guard      a spec format ledger line the audit could not see
brain slice cap   HALT at 37,992 of 32,000, no seat spawned

How it dies

All four stress seats agree, and it is not bad rulings, rot, or zombies. It is non invocation: nobody convenes it, the casebook stays empty, and an empty casebook reads as no recent cases rather than as failure. The system would look intact while the substrate never forms.

Full verdict, stress reconciliation, and every evidence file live at ~/KAIRO/projects/tribunal/. Charter is UNRATIFIED pending Anthony. Built Jul 30 2026.