# E86 Frontier Paired Run — Status

**Status: BLOCKED_INVALID_CREDENTIAL**
**Real frontier calls: 0. Real spend: $0.00. Hard ceiling: $0.25.**

## What was attempted

- Same pre-registered suite as the E86 local evaluation: **20 held-out tasks**, 10 families,
  suite `e85-xdomain-v1`, task-set hash in `preregistration.json`.
- One frontier arm: `B_frontier_single`, Anthropic adapter, model
  `claude-sonnet-4-5` (configurable), temperature 0, max 64 new tokens.
- A credential was present in `~/.config/cain/anthropic.env`, but the provider probe returned
  **401 authentication_error: invalid x-api-key**. No inference request was made after the probe.
- The hard dollar ceiling was enforced before any message request. Nothing was purchased.

## What was verified

- The adapter path is implemented in `cain/cain2027/e86/frontier.py`.
- A mock integration ran on 3 tasks and every row is labelled `MOCK`; it is **not** a frontier
  result and is not counted as one.
- No credential value appears in any artifact; probes record status only.

## What is required to turn this into a result

Inject a valid `ANTHROPIC_API_KEY` through the approved secret environment, set
`E86_FRONTIER_MAX_USD` to an explicitly approved ceiling, and run:

```bash
cd /root/cain-e86-worktree
E86_FRONTIER_MAX_USD=0.25 python3 scripts/evolution86/run_frontier_eval.py
```

The script will:
1. re-check the pre-registration hash;
2. probe authentication without spending;
3. run only the 20 pre-registered tasks;
4. stop before the ceiling;
5. write raw rows and paired McNemar against the local `A_single` rows.

Until then, **no frontier comparison exists**. The local baseline remains 5/20 (25%) on these
20 tasks, with the E86 learned arm also 5/20.

## Honest classification

- This is a **BLOCKED** gate, not a negative result and not a positive result.
- The provider rejected the credential; the adapter did not silently fall back to a mock.
- The mock rows are integration evidence only.
