Small decisions.
Working applications.

Good retrieval combined with candidate assessment helps most. Classifier-only navigation through the table of contents performed poorly in this probe. Finding a provision does not replace interpreting it.

Adjust the review threshold to your product. The 86.4% result includes a code rule; it is not basal alone. Harder cases still need verification.

A lower threshold automates more cases but increases error risk. Your rules and a labelled message sample help select the right operating point.

The same mechanism can drive a character's reaction, select a simulation action or respond to a new event — from an explicit list of permitted moves.

At the 0.90 threshold, the application automatically accepted none of the 24 answers: 12 went to revision and 12 to review. Despite 69 correct assessments out of 72, this operating point supported checking rather than automated publication. Validate thresholds and rubrics on your own data.

basal can choose the next move from a real page's text and links. The Copernicus demo is one run; the table's 10/10 result belongs to a separate probe in our catalog. A business deployment needs testing on its specific sites, layouts and actions.

Narrow data with ordinary SQL first. This implementation evaluates every remaining row, so cost grows with row count. A few dozen messages do not establish performance on millions of records.

For all nine questions about existing code, the right function ranked first. However, the 0.80 threshold hid five of those hits: after applying it, 5 of 10 results were correct, including the SMTP rejection. This distinguishes good ranking from a well-chosen threshold. A large repository would need candidate retrieval and separate validation.

36 of 39 assessments were correct: all budget checks, 12/13 fit checks and 11/13 readiness checks. We did not measure sales conversion. Fit rules, weights and thresholds to your offer and treat the score as help organising work.

The cost matrix distinguishes the most probable category from the best action. We measured the router only, not downstream answer quality or cost, so we do not claim savings for the whole system.

Similar eligibility scores need not produce the same action. For PLN 590.40, expected incorrect-refund cost was PLN 3.28; for PLN 19,680, it was PLN 105.28. With review costing PLN 25, act chose refund preparation and human review respectively. This follows a simplified matrix and model probabilities, not measured savings. act and evidence are experimental; validate procedures, thresholds and costs on your data.

A classifier need not have a fixed label list. Changing the customer introduces new devices; selecting an order changes the next question; a new request can replace choice with act. This demo created 12 distinct request definitions served by one model, without separate weights or additional training. The application sets boundaries; uncertain cases go to people. Experimental act depends on probability quality and the chosen costs.

Value comes from combining two small decisions. “Invoice” alone cannot separate adjacent invoices or preserve a multi-page document. basal assesses meaning and boundaries; code combines the signals, and a PDF library splits the file. Production testing must also cover OCR quality, missing pages, unreadable scans and unusual layouts.

The recording also shows a real model mistake. Application checks stop the uncertain price before import. Thresholds and a separate question improve the procedure without guaranteeing every error is caught. Automated ERP imports need validation data, unit and currency rules, and a change preview.

Similar names do not necessarily mean the same company; different names do not always mean different companies. A small model can assess meaning under organisational rules. Begin with suggestions for approval and measure incorrect-link risk before automatically changing data.

Content classification becomes useful when combined with permissions. PLN 30,000 here is negotiating room, not measured savings or a certain avoided loss. We do not add it to ROI. Deployment needs company access rules, leak tests and false-alarm assessment.

The model links text to catalog meaning. Code calculates consequences from known dependencies. PLN 246,000 is the value of orders at risk of delay, not a certain loss or guaranteed saved revenue. Technical judgement and purchasing remain with people; production use needs the full schedule, stock reservations and supplier availability.

The team gets short, verifiable excerpts and an exception queue instead of just links. A quote is the source's statement, not basal's confirmation that an event is true. A factual price-increase report need not imply a critical tone. Sources, thresholds and topics can be adapted to your company. This probe sent no alerts or replies and attributes no savings to ORLEN.

One model helps compare campaign terms across media. The team gets findings tied to specific files before spending the ad budget. Combined input alone failed on the voice in this probe; a separate channel check exposed it. basal-omni is experimental and separate from basal-2.0; the released model may answer differently and media support may change. Business value is in preparing reviews faster; publication remains a human decision.

In this probe, early basal-omni-1.0-max read specific image fields and routed an unreadable document for rescanning. The application used the answers for quantity checks. This does not replace full invoice verification, authenticity checks or accounting approval.

On these calls, 21 of 25 fields matched the scripts, with 5/5 correct handling paths. Cards remain proposals for listening and approval. The selected examples demonstrate the application; they do not establish quality on arbitrary company calls.

One multilingual message can supply several fields needed in a workflow. Change the questions and options without separately training the classifier. Validate on your own voices, languages and recording conditions.

basal chooses the next step, n8n connects systems, code calculates and people approve exceptions. This division automates different workflows through one decision interface. None of the shown stages authorises payments.

A short model decision can guide an agent's next move. The application maintains allowed actions, permissions and confirmations. You see both the path to the goal and the point where automation must stop.

We did not train a separate model for each station. Questions, options and application logic define the task. The same principles transfer from games to ticket routing, image checks and workflow next-step selection.