Blog
Compare local decision models with text classifiers
Compare local decision models with TF IDF and frozen embedding classifiers using explicit label budgets, matched inputs and an untouched final test.
Detect duplicate tickets with local decision models
Separate candidate retrieval, same incident classification and cluster merging. Use a local decision model to score pairs while auditing false links.
Evaluate classifiers when human annotators disagree
Keep human judgments and evaluate local decision probabilities against a declared reviewer population. Separate majority accuracy from distribution distance.
Recalibrate local decision probabilities on your labels
Fit and freeze a task specific isotonic map for local decision probabilities. Preserve raw scores, save its contract and evaluate on independent labels.
Enforce logical constraints on local model decisions
Validate implications, exclusions and label counts after local classification. Preserve raw probabilities and send contradictory assignments for review.
Evaluate complete agent decision workflows
Score complete agent runs by final state and required replies. Track recovery and stopping reasons, then distinguish repeated success from a lucky attempt.
Audit pairwise rankings from local decision models
Compare candidates in both orders with local decision questions. Preserve ties and disagreements, detect preference cycles and audit the final display order.
Cache local decision results without stale answers
Build an exact cache around local decision probabilities. Preserve request semantics, version model inputs, expire stale entries and recompute actions.
Build hierarchical classifiers with local decision models
Build a caller owned label hierarchy with local decision questions. Trace parent and child choices, validate paths and interpret branch probabilities.
Select local classifier examples for human review
Build a review queue from local decision probabilities while keeping a separate random audit sample. Record selection reasons and protect evaluation.
Build weak labels for local decision models
Create candidate labels from fallible rules with explicit abstention and conflicts. Inspect vote coverage and keep weak labels separate from evaluation.
Version classifier labels and their descriptions
Treat classifier label meanings as a versioned contract. Write option descriptions, preserve dispatch IDs and test label splits with fresh model requests.
Build conformal sets from local decision probabilities
Turn local choice probabilities into conformal label sets. Implement the finite sample rank and understand coverage, empty sets and exchangeability.
Select few-shot examples for local decision models
Compare a baseline, selected examples and reordered demonstrations with fixed routing labels. Keep test labels outside prompts and report failures.
Shadow-test local decision models before rollout
Compare a candidate with the current decision model on matching requests. Keep its actions separate, count missing pairs and set promotion criteria.
Test interference between bundled decision questions
Trace how Ardana isolates question rows, then compare bundled and single question probabilities with stable IDs, fixed options and explicit tolerances.
Test the evaluator before scoring a decision model
Check question and option mapping, probability losses and failed-answer denominators with authored controls before comparing local decision models.
Benchmark local decision API latency under load
Separate cold loads from warm requests, measure complete decision responses, and report queueing, tail latency and successful throughput under load.
Choose local decision actions by expected loss
Turn local model probabilities into actions with explicit error costs, derive a binary threshold, and compare multiclass policies on held out data.
Evaluate rare events with local decision models
Read precision and recall at realistic event prevalence, calculate review volume, and evaluate average precision without hiding rare event failures.
Budget tokens for local decision requests
Measure rendered decision requests, distinguish reported usage from row limits, and test input length before sending long states to a local model.
Minimize private data in local decision state
Choose the fields a local model needs, trace downloads and request handling, and test whether redaction preserves the decision before sharing inputs.
Monitor decision probability calibration drift
Compare mature labels and stable score groups to monitor calibration after rollout, separate traffic shifts, and decide when thresholds need review.
Keep AI tool selection separate from authorization
Let a local model choose an action, then check identity, resource access and exact arguments in application code before executing the tool.
Handle local decision API failures
Separate uncertain answers from failed inference. Handle validation errors, queue overload and runtime failures before applying any action threshold.
Multi-label classification with local decision models
Use separate yes or no questions for overlapping text labels. Read each probability, set per label policies and avoid inventing a joint score.
Interpret ordinal scores from decision models
Read ordinal decision scores as expected level indices. Inspect the distribution, level fits and rubric before mapping a score to an action.
Record what produced a local model decision
Preserve requests, model revisions, quantization and runtime settings before an upgrade. Separate a useful audit record from identical inference.
Rerank retrieved text with a local decision model
Score query and document pairs through a local decision API, preserve candidate IDs and compare the resulting order with retrieval and a reranker.
How to handle requests outside your classifier's labels
Define unsupported and ambiguous requests separately, give each an explicit outcome, and test whether your classifier sends unknown requests to a handler.
How to evaluate a prompt-injection detector
Measure prompt injection detectors against harmless inputs, freeze thresholds before the final test, and check what happens to the protected agent.
What missing information does to AI decisions
Separate model mistakes, missing evidence and uncertain labels. Plan a controlled input audit before assuming a larger model will fix a decision.
How to test whether a model follows changing rules
Build paired tests that change a rule or a relevant fact, check both expected answers, and distinguish correct adaptation from a model that merely flips.
How to test text classifiers across languages
Evaluate classifiers with native, translated and mixed language inputs. Keep label meanings stable and compare results within each task and language.
How to build a trustworthy classification test set
Build classification evaluations with traceable labels, grouped splits and leakage checks. Keep model inputs separate from the answers used to grade them.
Fixing Gemma 4's oversized KV cache
Ardana v0.3.2 fixes an oversized Gemma 4 cache allocation. Model weights alone cannot tell us how much memory a run needs.
Run safetensors models locally with Ardana
Pull supported Hugging Face safetensors checkpoints into Ardana, run typed decisions locally, and understand where GGUF still fits.
Run THX-01 as a local decision model
Run THX-01 through Ardana's CLI and HTTP API. See how its encoder answers typed questions and where its token and option limits apply.
Quantizing a 4B decision model, q4_k_m against q8_0
We ran decider:4b at q4_k_m, q8_0 and bf16 on JevBench's 231 public items. q8_0 is six items ahead, the 2B gap is reversed, and neither clears the noise.
Mixture-of-experts models as local decision models
Two mixture-of-experts models on JevBench's 231 public items through ardana: what 3 to 4 billion active parameters decide and what the file costs in RAM.
Run a decision model in the browser tab on WebGPU
What ardana.ai/playground downloads into your tab, how onnxruntime-web runs decider:0.8b on WebGPU or WASM, and why the tab answers exactly like the server.
Instruct model vs decision model on the same questions
We sent the same typed questions to gemma-4, qwen3.5 and decider through ardana, shuffled the options, and measured where a zero-shot instruct model breaks.
Cascade routing with a local decision model's confidence
How to put ardana serve and decider:2b in front of a chat model, which returned field to threshold, and how to check the threshold on your own labelled tickets.
Does q4_k_m hurt a decision model's calibration?
We ran JevBench's public items against decider:2b at q4_k_m, q8_0 and bf16 through ardana serve on a CPU and compared accuracy, Brier score and calibration.
Ardana, Ollama 0.35 and llama.cpp for local decision models
A support ticket through ardana run, the documented requests for Ollama 0.35 and the llama.cpp server, what each returns, and which models each one can run.
Why a decision model answers with probabilities, not text
How Ardana's choice, score and noul questions turn a support ticket into a probability per option, and how an agent thresholds that instead of parsing text.
