# Clef Flash: Decisions at Machine Speed

> A 3:44 film testing Clef Flash on local security triage, support routing, and tool selection, with examples of its strengths and missed threats.

[Canonical HTML page](https://trustcyber.ca/insights/clef-flash-decisions-at-machine-speed/)

- Author: [Junior Williams](https://trustcyber.ca/about/)
- Type: Insight film
- Published: 2026-10-03
- Modified: 2026-10-03
- Topics: AI models, Cybersecurity risk, AI governance

## What this examines

Clef Flash takes a state and a set of typed questions, then returns probabilities for each answer. This film follows a local demonstration across synthetic security events, a ransomware scenario, images, customer support tickets, and tool requests. It shows useful routing decisions alongside missed security cases, then asks where human review belongs.

## Why it matters

Fast classification can help teams sort work, but the cost of a missed threat differs from the cost of a misrouted support ticket. The film makes that difference visible through the model’s recorded responses.

## Key ideas

- Typed questions turn one input into several decisions: yes or no, a choice from a list, or a score on a scale.
- The local runs demonstrate both useful classifications and security misses; speed alone does not establish reliability.
- Test the model on representative cases and use its outputs to support routing, uncertainty review, and escalation to a person.

## Typed decisions in one call

The film introduces three question types: binary decisions, choices from a defined list, and scores on a scale. A state can be a log line, email, ticket, or image. Several questions can be scored together, letting the same input support decisions about maliciousness, team ownership, an attack tactic, and severity.

## Recorded demonstrations and misses

The security sequence processes 32 synthetic events with four questions each. The recorded run produced 128 decisions in 57.551 seconds. A separate ransomware example scores 16 binary questions in one call; the median of three recorded runs was 3.747 seconds.
The film also shows important misses. An invoice-forwarding mail rule and payroll copied to a USB stick were both routed to no action. In the support examples, 23 of 24 tickets reached the expected team, while all 12 tool requests matched their expected tool. Those small test sets illustrate differences between tasks rather than a general accuracy guarantee.

## Where human review fits

The closing recommendation is to choose the model for the task and test it against cases that matter in that setting. Use classifications to sort work and probabilities to identify responses worth reviewing. Give a person the evidence needed to assess ambiguous or consequential cases.
The film also introduces Allow, Hold, or Deny as related work on evaluating AI agent authorization. Routing an alert and deciding whether an agent may act require different questions and evidence; the recorded examples provide a starting point for testing those boundaries.

## Evidence note

The film uses synthetic logs, tickets, requests, and images. Its displayed local timings and model outputs come from recorded project runs on an Apple M4 Pro with 48 GB of memory. These examples describe this demonstration; they do not establish production accuracy or calibrate the reported probabilities.

## Resources

- [Open the original MP4](https://trustcyber.ca/videos/clef-flash-film.mp4)
- [Video](https://trustcyber.ca/videos/clef-flash-film.mp4)