Blitz RESEARCH PREVIEW

SHARED-STATE DECISION MODEL

One state.
Many decisions.

Give Blitz one view of the situation. Ask a panel of focused questions. Get explicit distributions that software can compose into its next action—without generating a paragraph.

See the workflow → Try one decision
shared state→question panel→ distributions→code action

01 / SHARED-STATE FANOUT

One trace. Five focused decisions.

Blitz acts like a semantic sensor panel. Each question reads the same evidence independently; ordinary code combines the resulting distributions.

SHARED STATE

Agent recovery trace

Attempt three changed the same files. The same two tests still fail. A reviewer says the working hypothesis no longer explains the evidence.

  • Attempts3
  • Repeated failureYes
  • BudgetMedium
01Same root cause?

YES91%

02Would another retry help?

NO86%

03Hypothesis invalidated?

YES78%

04Needs specialist review?

NO64%

05What should happen next?

CHANGE STRATEGY82%

CODE COMPOSES

change_strategy

Stop repeating the failed approach and form a new hypothesis.

Illustrative probabilities. The live playground below exposes the atomic decision primitive; shared-state batched serving is the next measured milestone.

02 / INSPECT THE PRIMITIVE

Try one atomic decision.

Checking endpoint…

A workflow is built from small, focused judgments. Start with a real checkpoint example, then change the evidence, question, or options.

INPUT / CHOICE
03 Choices

Research checkpoint. Probabilities are not deployment-calibrated guarantees.

OUTPUT AWAITING RESULT

Your decision appears here

Ready when you are.

03 / MEASURED, NOT PROMISED

A compact model with visible receipts.

Fresh confirmation results and a warm A100 throughput measurement from the current research checkpoint.

0.6Bparameters
235.8decisions/s · batch 32
75.9%preferred-choice agreement
0.160Brier score
2,000fresh confirmation cases

Synthetic held-out evaluation; not a claim of real-world accuracy or calibration.

04 / THE MODEL'S JOB

A model for the uncertain part.

Blitz is a compact 0.6B model adapted to score bounded choices in one forward pass per atomic question. It handles the semantic judgment between observed evidence and a software system's next action.

0.6B parameters→one forward pass→choice probabilities
  1. 01
    Share the evidence

    Present one concise state that every decision can inspect.

  2. 02
    Fan out focused questions

    Evaluate independent routing, evidence-availability, risk, and recovery signals.

  3. 03
    Compose the action in code

    Use explicit distributions to continue, retry, review, reroute, or stop.

WHERE BLITZ FITS

Repeated decisions inside software workflows.

01

Agent recovery

Detect repeated failures and choose whether to retry, review, or change strategy.

02

Incident operations

Judge severity, route ownership, and decide when ambiguous evidence needs escalation.

03

Evaluation pipelines

Score retrieval relevance, context coverage, semantic equivalence, and experiment outcomes at scale.