Skip to content
Bahman Shadmehr Independent AI Systems & Automation Engineer

System patterns

11 guarantees behind reliable AI systems.

Choose what must stay true.

These are design guarantees: properties each pattern is built to keep. They are not a service warranty.

Each guarantee names something that must never fail. Next to it is the pattern that holds it: the decisions, authority, and failure modes behind it. The first six are built into my open-source honeworks packages and marked "Open source"; the others are reference designs, marked as such.

  • A crashed run continues where it stopped, with its evidence.

    Item 37 of 50 failed and the script starts again from item 1.

    Resumable AI Workflow Runs

    A run is a folder you can open. Open source

  • A failed score is missing, never zero, and every win can be explained.

    A judge that crashes gives a good candidate a score of 0.

    Gated Best-of-N Selection

    No single judge picks the winner. Open source

  • A prompt that doesn't fit fails loudly instead of being truncated.

    A long JSON prompt was silently cut at the context limit.

    Capability-Checked Model Calls

    Refuse the call that can't work. Open source

  • Every taste score says whose taste it learned and why it scored.

    The output is correct and still nobody would want it.

    Stand-in Human Judgment

    Ask people rarely, and ask them well. Open source

  • A cause stays suspected until a replay test confirms it.

    No single run looks broken, but the outputs keep getting worse.

    Cross-Run Failure Analysis

    Some bugs only exist across a thousand runs. Open source

  • Agents decide small things and log them; design changes wait for approval.

    The agent keeps asking questions, or stops asking and quietly changes the design.

    Spec-Driven Agent Development

    The agents write the code. I own the decisions. Open source

  • Accepted work always ends in an owned state.

    Failures have no final state.

    Durable, Replayable Execution

    Accepted is a promise. Reference design

  • Merge decisions stay reversible and evidence-based.

    Merge decisions cannot be reversed.

    Entity Resolution with Reversible Canonicalization

    Two records. One company. Maybe. Reference design

  • Conclusions never outrun their evidence.

    Citations do not entail conclusions.

    Evidence-Bounded Inference

    No evidence. No conclusion. Reference design

  • A confident answer never skips required review.

    Confident outputs bypass required review.

    Risk- and Novelty-Aware Model Cascades

    Use the smallest model that knows when to stop. Reference design

  • Sensitive data never exits, even during fallback.

    Public fallback activates during capacity pressure.

    Private Inference with No-Egress Fallback

    Private means the fallback stays private too. Reference design

Will be published soon

Three more patterns from the same packages, written up next.

  • Will be published soon

    Missing Is Not Zero

    Keep 'could not score', 'could not answer' and 'could not call' as distinct, typed outcomes through a whole pipeline, so failures never turn into numbers that decisions are made on.

  • Will be published soon

    Offline-First Tests for AI Code

    Test model-calling code with public fakes, recorded HTTP and contract checkers by default, and run real-model tests only behind a lock and a marker.

  • Will be published soon

    Plan, Approve, Then Run

    Declare a model or prompt comparison as a file, show its plan and cost, get a person's approval, and run it under recorded conditions so the result is worth trusting.

Not sure which fits?

Let's build something real.

Tell me what must stay true. I will suggest the right pattern.

Let's build something real Email me