One state.
Many decisions.

Ask several focused questions about one situation. Get probabilities back in one fast pass, so your code can act without parsing a paragraph.

THE SITUATION

Agent fix, attempt 3

Testssame 2 still failing
Theory predictsHTTP 401
Trace showsHTTP 500
Budget1 step left

Attempt 3 changed api.ts and auth.ts. auth_spec L42 and session_spec L17 still fail on the same assertions recorded after attempts 1 and 2. The token-refresh theory predicts HTTP 401; the trace records HTTP 500.

01Same root cause?

—

02Would another retry help?

—

03Theory ruled out?

—

04Needs a specialist?

—

YOUR CODE DOES

…

See the rule

if any answer < 70% sure → gather_more_evidence
elif same cause, retry won't help,
     theory ruled out → change_strategy
else → continue_plan

Why not just ask a chat model?

A chat model writes an answer that your code then has to parse. Blitz scores the choices directly.

Chat model0 ms

Blitz0 ms

Same root cause?YES 88%
Retry help?NO 73%
Theory ruled out?NO 53%
Needs a specialist?YES 53%

Chat-model lane is illustrative: about 150 tokens at 60 tokens per second. The Blitz lane uses measured warm model time.

Try it on a case.

Pick an example, then run it. Edit the evidence or choices to test your own.

Checking endpoint…

OUTPUT AWAITING RESULT

Press run.

Promised: 20–25 ms.

22 msto answer four questions at once, warm

0tokens generated. Nothing to parse

235/sdecisions per second on one GPU

Methodology & full results

Earlier paired benchmark, warm p50: 91.83 ms model time in one batched job vs 340.83 ms as four separate jobs (3.7×). End-to-end 1,318 ms vs 5,438 ms, three paired runs.
Single decision: 69 ms warm batch-1 model time (live snapshot).
Throughput: 235.8 decisions/s, warm A100, batch 32.
Agreement: +4.4 pp: selected checkpoint 75.9% vs control 71.5%.
High-confidence slice: 96.8% accuracy at 15.4% coverage for confidence ≥ .90.
Brier score: 0.160 over 2,000 confirmation cases.
Model: 0.6B parameters. The panel batches prompt rows in one forward pass; shared-prefix reuse is not implemented yet.