Companion Chat

What Companion Chat is for

Companion Chat is the AI panel available throughout the workstation. Open AI Companion from the global navigation while you are in Data Manager, Signals, Strategies, Portfolios, Studies, Results, Notebooks, Live, or System & Jobs. You can keep the panel open while moving between pages, resize it on a desktop, or use the same entry from the mobile navigation drawer.

Use it to develop an idea, draft work for review, understand a Result, diagnose a blocker, or ask how a workstation control fits into the larger workflow. It is an assistant inside Arizmic, not a second authority over your data, Results, or trading account.

Start a conversation

1. Connect a model in System & Jobs → AI Models.
2. Open AI Companion from the global navigation.
3. Select the exact provider and model you want to use.
4. Check the context disclosure. Confirm the page, selected object, version, and bounded fields match the question you intend to ask.
5. Write a specific question and choose Send.
6. Stop a streaming response if it is heading in the wrong direction or no longer needs more context.
7. Verify consequential statements in the owning workstation page before using them.

The current baseline provides a global, read-only dock with page orientation, model selection, response usage, and Trust Records. Durable multi-message conversations, selected-object resolution, streaming cancellation, proposals, and action receipts are intended v1 capabilities and remain unavailable until their implementation gates pass.

Know what context the model receives

Being on a page does not grant the model unrestricted access to that page. The intended context system resolves a bounded summary of the object you select and shows it before sending. Supported summaries may include a Dataset or Prepared version, Signal, Strategy, Portfolio, Study, Result, Job, or Notebook record.

The disclosure should name:

• the active page and selected object;
• the exact object or Result version;
• the fields and artifact excerpts included;
• omitted or redacted information;
• the provider and model receiving the request; and
• whether the endpoint is local or hosted.

Raw credentials, unrestricted file paths, account secrets, full plugin source, and unbounded datasets are excluded. Text found in a Dataset, Notebook, Result, or imported file is context, not an instruction that can expand the model's permissions.

Brainstorm ideas

Use Chat as a thinking partner before building anything. Describe the behavior you want to capture, the instrument and timeframe, the data you have, and the constraints that matter. Ask the Companion to:

• turn a rough observation into one or more concrete Signal or Strategy ideas;
• suggest alternative explanations and ways the idea could fail;
• identify the minimum data level and history needed;
• compare a few entry, exit, sizing, or risk approaches;
• propose the smallest useful test; and
• point out assumptions that should become explicit settings.

A useful prompt is: “I want to capture intraday trend continuation after an opening range break. Give me three distinct approaches, the data each needs, where each tends to fail, and the smallest test I should run first.”

Brainstorming is intentionally open-ended. Nothing it produces becomes a saved Signal, Strategy, Portfolio, Study, Notebook, or Result until you review it in the feature that owns that object.

Author with AI Companion

AI Companion is the no-code authoring path for users who want help turning an idea into a structured draft. It can propose:

• a Signal definition with inputs, outputs, parameters, data requirements, and dependency declarations;
• a Strategy with rules, Signal bindings, sizing, exits, and execution needs;
• a Portfolio with members, allocation, shared capital, admission, and risk policy;
• a Strategy or Portfolio Study design with parameters, ranges, validation, and requested fidelity; and
• Notebook source or an investigation outline for a bounded analysis.

The draft must use the same schemas and validators as manual authoring. The Companion may explain a validation error and propose a repair, but it cannot declare an invalid object ready. Review the exact fields and diff, correct any mistakes, and accept the proposal through the owning Signal, Strategy, Portfolio, Study, or Notebook workflow.

Generated code and formulas require the same review as code you received from another person. Never assume a fluent explanation proves that timing, units, dependencies, state, or execution semantics are correct.

Explain Results

Open the exact Run Output or Study Result first, then ask the Companion to explain a bounded selection. Useful questions include:

• “What drove the largest drawdown in this Result?”
• “Explain the relationship between trade count, average trade, and costs.”
• “Which validation blockers prevent this candidate from moving forward?”
• “Summarize the differences between these retained trials without ranking unsupported fields.”
• “Which Result tab should I inspect next, and why?”

Grounded explanations must use the engine-reported values, retained lineage, validation state, and verified artifact excerpts supplied by Results. The model may explain those values; it may not recompute authoritative KPIs, invent a missing trade, strengthen a validation label, or imply that a simulated fill occurred live.

Treat every explanation as a guide back to the underlying Result. Open the cited field, trade, diagnostic, or artifact and confirm it directly.

Recommend Strategy improvements

The Companion can review an exact Strategy version alongside one or more retained Results and propose changes worth testing. Ask it to separate:

• failures caused by the Strategy rules from failures caused by data, execution assumptions, costs, or insufficient validation;
• behavior that is consistent across runs from behavior concentrated in one period, instrument, or parameter choice;
• a repair to an obvious defect from a new Strategy hypothesis; and
• risk-control changes from attempts to improve returns.

Good recommendations are specific and reviewable: change one condition, add a filter with a stated purpose, test an alternative exit, reduce concentration, or request a higher-fidelity validation. Each recommendation should say what it is meant to improve, what it may worsen, what data it needs, and how to test it.

Do not ask the model to “make this profitable” or accept a bundle of changes whose effects cannot be separated. A recommendation is a hypothesis. Save it as a new Strategy version only after reviewing the diff, then compare it through a new Study. Past Results do not guarantee that the change will improve future performance.

Use proposals and actions safely

The intended v1 Companion is read-and-propose by default. When an approved workflow supports an action, the sequence is:

1. the Companion creates a typed proposal;
2. the owning feature validates it;
3. you inspect the exact diff, scope, limits, and warnings;
4. you approve that one action or use an explicit limited autonomy grant;
5. the owning feature performs the change; and
6. Arizmic records an immutable receipt with the result and affected versions.

The model cannot grant itself permission. It has no unrestricted shell, filesystem, network, credential, plugin-source, or arbitrary HTTP access. Failures and partial effects remain visible rather than being hidden behind a successful-sounding response.

AI Companion does not place, modify, or cancel Live orders; arm, pause, resume, flatten, or kill deployments; change broker credentials; or weaken Live risk controls. It may explain a visible Live state or direct you to the appropriate operator workflow, but Live actions remain under their own confirmations and receipts.

Protect private information

Local conversations and history are stored locally in the intended v1 design and can be inspected, exported, or deleted. A local history record does not make a hosted request local: any disclosed prompt and context still go to the selected hosted provider.

Before sending, remove credentials, account identifiers, proprietary datasets, unnecessary Strategy source, local paths, and personal information. Use the smallest context that can answer the question. Review the provider's retention settings independently.

Recover from common problems

• No model is available: configure and test a route in Model Connections.
• The answer ignores the selected object: stop and inspect the context disclosure. Select the exact object/version or provide a smaller verified excerpt through the supported context control.
• The answer contains a number you cannot locate: do not use it. Ask for the source field, then verify it in Results or the owning object.
• A proposal fails validation: keep the validation findings, revise the proposal, and run the same owner validation again.
• A request times out or is interrupted: do not repeat an action until its final state is known. Read the Trust Record or action receipt first.
• The model suggests a Live action: treat it as commentary only. Use the Live page's own readiness and confirmation workflow.
• Sensitive context was sent to a hosted provider: stop the route, rotate exposed credentials if necessary, preserve the disclosure and Trust Record, and follow the provider's incident process.

Next

Use Companion Workspace when you need the intended full-page conversation, history, proposal, and receipt workflow. Return to the relevant Signal, Strategy, Portfolio, Study, Result, or Notebook page to validate and apply any useful proposal.