GuardianAI — LAB6 — Math Slip-Collapse Boundary
Maps recoverable arithmetic errors vs irrecoverable trajectory collapse
WebsiteEvidence
Run READY
Provider together
Guardian off
README
Left: Baseline LLM|Right: GuardianAI (Slip Gate)

Run Setup

Raw vs Stabilized

`1` uses the prompt above with live prompt and answer panels. Higher counts run the controlled LAB6 math script so the prompts, expected answers, and numeric checks stay reproducible in the browser. The right lane is GuardianAI only.
LAB6 controlled math stays fixed here so the homepage can reproduce the archived math evidence runs cleanly.
`0` = no guarded retry. Higher values let GuardianAI spend more reopen or verification passes before withholding.
Standard keeps the production Guardian gate. Slip experimental keeps the same signals but swaps in the verification-first decision layer for slip experiments.

Math Benchmark Source

LAB6 math script

50 prompts

LAB6 arithmetic prompts are running in batch mode.

The active math problem, target answer, and lane outputs stay visible in the browser for tuning.

Status: Waiting for the first completed math prompt in this run.

The active math prompt will appear here after prompt 1 completes.
The expected numeric answer will appear here for math-script runs.
The current plain-LLM answer will appear here after prompt 1 completes.

Run Summary

Snapshot

Run the selected mode to compare the raw left lane with the stabilized right lane.

A — Plain LLM

No GuardianAI

The left lane will show the live plain-LLM math answer and batch outcomes after you run the comparison.

B — Stabilized Lane

GuardianAI (Slip Gate)

The right lane will show the live GuardianAI math answer, corrections, and release decisions after you run the comparison.