Opening your Chat workspace.
Loading your game…
- 1 · Set up
- 2 · Train
- 3 · Check
- 4 · Run
Know the grounds. Set the goal. Then train.
- Environment & assumptions
- Measurable goal
- Reference strategy
- Training phases
- Checks & limits
Objectives & optimization
Example: Success at least 90%; score at least 5; steps at most 500; prioritize efficiency. Separate clauses with semicolons. Use supported metrics shown in Use values. Choose game rules and the task in World; text does not create new mechanics.
Chat
One environment. Repeatable decisions. Visible evidence.
What you teach
Route an inbox to Billing, Technical support, Sales, Delivery or Human review. One decision labels and routes one message; eight messages complete a round. Ambiguous, unsupported and instruction-conflict examples ask for review. Try your own message and context, then inspect the matched words.
A local English vocabulary teaching classifier. It observes message/context text and engineered whole-word features, never the expected label. The small learned action scorer imitates explicit rules; it is not a generative chat model, semantic language understanding, calibrated probability or a guarantee of 99.999% accuracy. Seeded templates share vocabulary across splits; held-out seeds test new combinations, not unseen language. No real message is dispatched.
- World: try the reference, then control a round yourself.
- Train: fit a small action-scoring network away from the animation.
- Results: compare paired seeds, inspect losses and activate checked weights.
- World: test a new seed. Save project to keep your model.
Chat
Watch decisions change the outcome.
Try your own message
Classify, label and route text using the same inputs as the neural scorer. Unknown or conflicting requests belong in Human review.
Custom rounds check the rules. Custom text is excluded from saved and exported evidence and is not added to training.
A larger context model, when you need one
The built-in teaching classifier runs free in your browser without a model download or API key. It is intended for this routing vocabulary.
Qwen3-4B-Instruct-2507 is a local language-model candidate for richer labels and longer context. Its publisher documents 4 billion parameters, Apache 2.0 licensing and a native context window of 262,144 tokens. It is not installed or benchmarked by this lab; running it consumes your machine's memory and compute. Maximum context depends on the runtime and available memory. A lighter Qwen3-0.6B candidate has a documented 32,768-token context and Apache 2.0 licence, with less capacity for complex classification. Neither candidate provides free hosted compute.
For reliable routing, validate output against allowed labels and send uncertainty to review. Deterministic decoding can repeat an answer, but does not establish that it is correct. A 99.999% reliability goal needs representative labelled data and independent measurement.
Manual buttons work with You · manual. Step advances the selected policy. One step is one decision, including the reference opponent response where applicable. Reset after a round ends.
Rules and what the network sees
Route an inbox to Billing, Technical support, Sales, Delivery or Human review. One decision labels and routes one message; eight messages complete a round. Ambiguous, unsupported and instruction-conflict examples ask for review. Try your own message and context, then inspect the matched words.
A local English vocabulary teaching classifier. It observes message/context text and engineered whole-word features, never the expected label. The small learned action scorer imitates explicit rules; it is not a generative chat model, semantic language understanding, calibrated probability or a guarantee of 99.999% accuracy. Seeded templates share vocabulary across splits; held-out seeds test new combinations, not unseen language. No real message is dispatched.
An Alkamy Labs experiment · Explore our labs ↗
From features to an action.
6 learned weights → one score for each legal action.
A single linear neuron scores each action. It receives these engineered features, not a teacher’s chosen action:
- Supported route words
- Multiple supported routes
- Route has no evidence
- Review needed
- Review of a single supported route
- Instruction-conflict teaching flag
Ridge least squares fits the weights from synthetic reference scores. Rules supply legal actions, feature extraction and opponent behaviour. This is imitation learning. It does not discover an optimal strategy through self-play.
A local English vocabulary teaching classifier. It observes message/context text and engineered whole-word features, never the expected label. The small learned action scorer imitates explicit rules; it is not a generative chat model, semantic language understanding, calibrated probability or a guarantee of 99.999% accuracy. Seeded templates share vocabulary across splits; held-out seeds test new combinations, not unseen language. No real message is dispatched.
Train. Freeze. Compare.
Active weights stay in place while the candidate is checked.
64 fitting seeds, 12 validation seeds and 24 final seeds. Validation selects between two ridge strengths; final tests cannot change the selected weights. Maximum 25 seconds of computation, 30-second worker timeout and two attempts per training round.
Ready · no provider or API key needed.
Activation gate: ≥80% near-best reference-score agreement, finite compatible weights and final mean score at least the untrained baseline and a checked incumbent, when present. Passing this small test does not guarantee future wins.
Uses fresh teacher examples and the current two-attempt training round. Review each candidate before activation. Run keeps active weights fixed.
Off. Train manually, or start another training round when both attempts are used.
What changed?
Compare the same seeds. Keep the losses.
Train or check saved weights for a measured comparison.
Complete a round or training run to see observations and a controlled next test.