# VAGI Behavior Suite v1 — model baseline pre-registration

Frozen: 2026-11-18, before any model call.

## Question

Can a raw frontier model, given the **same published task data and the same policy card as the
library arm**, solve the suite's 500 held-out tasks?

This is an external baseline for the suite. It is **not** a test of CAIN's governed procedure
runtime, and it is **not** evidence of AGI or VAGI.

## Arms

| Arm | Model | Tasks | Selection | Max output |
|---|---|---|---:|---:|
| `haiku-5-5-full` | `claude-haiku-5-5` | 500 | all held-out (seed 9402) | 400 |
| `fable-5-2-stratified` | `claude-fable-5-2` | 100 | 10 per domain, deterministic seed 9411 | 400 |

## Protocol

- One attempt per task. Wrong answers are never retried. Invalid JSON is incorrect.
- The prompt contains the task data, a domain instruction, the authoritative policy card
  (authority table, governance rules, recovery mapping, newsvendor table) and the exact JSON schema.
- The prompt **never** contains `fixes`, `hidden`, `bad`, a reference answer, or another task's answer.
- No tools, no memory, no code execution beyond the suite checker (software answers pass a static
  screen first: no imports, no `eval`/`exec`/`open`, no dunder access).
- Concurrency 6. API temperature is omitted because the API rejects it for these models.

## Scoring

`vagi_suite.suite.check(domain, task, answer)` exactly. The analysis reports overall and per-domain
counts, exact Wilson 95% intervals, invalid-JSON rate, API errors, cost, and latency.

## Budget and stopping rule

Hard cap **$1.50 per arm**, **$1.91 total**. A run stops before exceeding the cap and reports every
skipped task. List prices used: Haiku 5.5 $0.10/$0.50 per MTok; Fable 5.2 $8/$40 per MTok
(Anthropic pricing page, read 2026-11-18).

## What would make this useful

- A **low** raw-model score shows the suite tests procedure application, not raw model knowledge.
- A **high** raw-model score shows the suite is easy for a frontier model and weakens any claim that
  the library's 500/500 demonstrates a special capability.
- Either result is published, with its limits.

No AGI/VAGI/production/certification claim. No independent verifier. The suite is constructed and
the library is hand-written for these families.
