EXPERIMENT 002.1 · AI SAFETY
Testing reward
compatibility.
Can a checked example establish that two rewards can be satisfied together?
LUNA · INDEPENDENT API CALLS
One experiment · up to eight calls in parallel
Development & safety gate
Waiting for the first checkpoint
Rough elapsed-pace estimate, refreshed every 15 seconds. Different phases may take longer.
- Phase jobs completed
- Awaiting first checkpoint
- Model / reasoning effort
- gpt-5.6-luna / none
- Isolation preflight
- Awaiting verification
- Last checkpoint
- Awaiting connection
SHARED BUDGET
— committed / $40 cap
The $40 ceiling is shared, not reset for this run. Pending requests retain reservations; failed requests with unknown charges remain conservatively accounted. Usage is application accounting, not a reconciled provider invoice.
WHAT IS BEING MEASURED
Two rewards. Independently checked.
rout
Does the proposed answer or finite program meet the outcome requirement?
Exact coin-state or bounded integer checks.
rCoT
Does the visible trace meet the specified reasoning constraint?
A check on public structured output, not hidden chain of thought.
Joint witness
A single valid proposal that satisfies both reward requirements.
Failure to find one is not a proof of conflict.
This stage validates compatible versus conflicting requirements within a declared finite language. It does not yet establish all three aligned / orthogonal / in-conflict categories in unrestricted tasks.
FIXED COMPARISON
Description only
Judge the reward pair from its specification.
Unguided search
Propose examples without verifier feedback.
Verifier-guided search
Use exact feedback to improve the next proposals.
80 development cases · 320 held-out cases · 40 held-out templates · two domains. The backdoor task uses a restricted affine-trigger JSON language, not arbitrary generated code. Fixed sample size; no stopping for significance.
EVALUATION
Held-out results are sealed
Development logs are public. Held-out calls and outcomes are withheld until the fixed run completes.
LIVE QUEUE
0 active / 8 maximum
Connecting to the queue…
RESEARCH RECORD
Development calls
Public explanations and proposals, not private reasoning. Five records per page; no full-ledger download is needed to watch progress.
Loading records…