DECISION ENGINE / PLAYGROUND

Route work with clarity.

Classify a request, draft a tool sequence, and inspect recovery decisions — without executing any actions.

◆

7Intent categories

⇄

HybridClassifier + policy

⌁

0Tools executed

01 / INPUT

Task prompt

MULTI-LABEL

Describe the work to be done

TRY AN EXAMPLE

02 / ENVIRONMENT

Tool availability

SIMULATED

{ }Code executionPropose running tests⌕Web searchPropose source retrieval▤File accessPropose project reads

Simulated tool failure

Retries already used

03 / OUTPUT

Decision trace

Awaiting input

INPUT

DECIDE

ROUTE

Nothing analyzed yet

Enter a task to see the proposed tool sequence, confidence-like scores, and any required human approvals.

NO EXTERNAL ACTIONS WILL BE EXECUTED

◇Suggestions only. Authorization remains with the host application.

TRADE-OFF EXPLORER

Compare with context.

Rank alternatives using criteria you set. The system does not invent option values or make autonomous purchases.

01 / CANDIDATES

Options & attributes

Scores range from 0 to 100. Lower risk and cost are better; higher values are better for all other criteria.

02 / SETTINGS

Decision criteria

Quality Risk Cost Speed Certainty

Max cost

Max risk

03 / RANKING

Results

Select Compare options to view your weighted ranking.

MODEL / TECHNICAL CARD

Built to be inspectable.

Understand what Decisions Flash does, how it was evaluated, and where its limits are.

Training samples248Synthetic and curated

Holdout samples78Small internal set

Micro-F185.5%Internal test only

Exact-match64.1%All labels correct

ARCHITECTURE

Two layers, separate responsibilities.

01

Intent classifier

Character TF-IDF and independently trained logistic-regression classifiers rank potential intents in English and Arabic.

02

Deterministic policy engine

Rules order proposed tools, simulate failures, cap retries, and require external authorization for sensitive actions.

03

Weighted option matrix

Ranks only the values you enter. Weights and hard limits are explicit and inspectable.

BOUNDARIES

What it does not do.

No autonomous tool execution

No calibrated confidence estimates

No learned reinforcement policy

No real-time data or hosted inference

No security enforcement in the browser