RESEARCH / PAPERS
Teardowns, reports
and measured runs.
Teardowns of how agent systems are built, a source-by-source review of open-weight cyber capability, and a track on small models that make typed, calibrated decisions without generating text.
Agents and the field, examined.
Infraze: agentic DevOps on an engine you choose
A proposal for a vendor-neutral infrastructure agent that routes each step to a cheap or an expensive engine, remembers what it built, and gates what it breaks.
GLM-5.3 and open-weight cyber capability: what the evidence supports
A source-by-source reading of Anthropic's GLM-5.3 cyber report: what three parties agree on, where the framing outruns the data, and what open weights change.
OpenClaw: voice, FaceTime and where to host it
How OpenClaw does phone calls, device Talk, iMessage voice notes and experimental FaceTime, and every documented way to host the Gateway and its Macs.
OpenClaw: a self-hosted agent gateway for your chat apps
How OpenClaw works: a local Gateway that routes chat apps to an agent loop, with tools, sandboxes, plugins and SQLite state. Product, infra and code.
Muse Code: Meta's terminal coding agent, taken apart
How Meta's Muse Code works: a Rust CLI on Muse Spark with an event log, OS sandbox, subagents, MCP and the Muse Session Protocol.
Instinct: a personal agent you text, with its own computer
How Instinct, the invite-only personal agent from Spear Street Technology, works: channels, cloud computer, Vault, payments and safety.
Dots: OpenAI's always-on agent with its own cloud computer
How OpenAI's dots work: a persistent GPT-6 Astra agent in ChatGPT with its own cloud computer, plugins, auto-review and memory.
Decision Bench v0, first run: three systems on Banking77
First measured results: a TF-IDF baseline, Laya zero-shot, and GLiClass-small zero-shot on 77-way Banking77, scored on accuracy, calibration, flips, and cost.
Decision Bench experiments: temperature refit and the option-count sweep
Two follow-ups: a train-only temperature refit fixes a trained baseline's calibration, and zero-shot models hold at 5 options but degrade sharply by 77.
Decision Bench experiments: near-miss distractors, per-bucket temperature, and shortlists
Similar distractors erase most of Laya's small-option edge, per-bucket temperature beats one global value, and a shortlist cuts order flips from 55% to 9%.
Laya: An Open Typed-Decision Engine, Examined
A technical teardown of Laya's encoder architecture, calibration path, public benchmarks, and an Apple Silicon smoke test.
Jev and the Typed-Decision Landscape
A survey of TypeSafe Jev, the three meanings of RLCD, and open alternatives for non-generative decisions.
Typed Decisions Without Generation: A Survey
A survey of models that answer closed, typed questions without generating text: their lineage, two working systems, and the open problems that remain.
Abstention and Honest Uncertainty for Small Decision Models
A survey of how classifiers and language models say they do not know, and how to score that behavior in typed-decision systems.
Which Encoder Backbone Should a Small Decision Model Start From?
A survey of open encoder backbones for typed-decision models, comparing ModernBERT, mmBERT, Ettin, DeBERTaV3, and retrieval-tuned alternatives.
Maestro Flash: A Non-Autoregressive Multilingual Model for Typed Decisions and Structured Outputs
Design proposal for Maestro Flash, an untrained multilingual encoder for typed decisions: architecture, training recipe, evaluation plan and compute.
Option Cardinality Limits in Small Decision Models
Why jointly encoded options degrade past about 20 labels, why Jev caps at 255, and which encodings avoid both limits.
Why Option Order Flips Typed Decisions, and How to Measure It
A report on permutation sensitivity in typed-decision models: why reordered options flip answers, how to measure it, and what fixes exist.
Calibrated Decisions via Proper-Scoring-Rule RL for Small Encoders
A proposal to train small typed-decision encoders with reinforcement learning against proper scoring rules, so calibration is trained, not patched.
What a Million Decisions Costs: APIs, Small Encoders, and Local Hardware
A cost comparison for typed decisions on commercial APIs, self-hosted encoders, and local hardware, using only fetched list prices and published measurements.
Sub-1B Decoders as Decision Models: Constrained Generation Against Encoders
A survey comparing small decoder models with constrained decoding against encoder classifiers on typed decisions.
Distilling LLM Decisions into Small Encoders
A survey of how soft targets, hidden state transfer, and rationale supervision carry decisions and probabilities from large models into small encoders.
Zero-Shot Label Scoring in the GLiNER Family
What GLiNER, GLiClass, and GLiNER2 teach about scoring arbitrary labels with small encoders.
Calibrating Sigmoid Label Scorers for Selective Automation
Per-option-count temperature fitting fixes confidence scale, but sigmoid scorers still need an explicit decision rule and a ranking check.
Constructing Unanswerable Items for Typed-Decision Benchmarks
How to build underspecified, contradictory, and out-of-language items so abstention precision and recall grade exactly, without a judge.
Fixing the Denominator: A Cost-per-Accepted-Decision Benchmark Design
A design that fixes prompt shapes, token accounting, error budgets, and utilization so one cost figure compares APIs, rented GPUs, and local hardware.
An Indic-First Typed-Decision Benchmark: Dataset Proposal
A proposal for a human-verified benchmark of typed decisions in Indic languages, with native-script and transliterated pairs for each core example.
Stage Zero: A Fixed Six-Axis Harness for Choosing a Decision Backbone
A measurement design testing four backbones on Banking77 accuracy, ordinal error, Brier score, calibration error, median latency, and option-order flip rate.
A Standard Flip-Rate Protocol for Typed Decisions
A measurement design that fixes the seeded permutation count, option-count buckets, and distributional readout for flip-rate reporting.
Teacher Temperature and KL Direction for Calibrated Distillation
A survey of how teacher temperature, calibration, and divergence direction shape post-refit Brier score and calibration error on high-cardinality choice heads.
Cardinality Break Points: Context Length or Head Ceiling in Single-Pass Decision Encoders
A measurement design that sweeps option cardinality per trunk to test whether accuracy breaks follow native context length or a shared head ceiling.
What Multilinguality Costs on Indic Decisions, and How to Measure It
A report bounding the cost of running mmBERT-base instead of an English ModernBERT control on native-text Indic decision benchmarks, per language.
Velofy Decision Bench v0: One Frozen Protocol for Small Decision Models
A frozen benchmark design that unifies accuracy, calibration, abstention, flip rate, cardinality, and cost per accepted decision into one comparable run.
Maestro Flash: Architecture, Roadmap, and First Skeleton
A proposal for a multilingual typed-decision model, its 90-day evaluation plan, and a minimal CPU-only pipeline experiment.
Permutation-Group Training for Small Encoders: Does a Consistency Reward Cut Flips?
A report on whether permutation-group training with a consistency reward cuts encoder flip rates without hurting accuracy or calibration.
Do Calibrated Teachers Produce Calibrated Students After the Refit
A survey of whether calibrating the teacher before distillation improves the student's calibration error after the refit, or only its accuracy.
Where Refusal Should Be Learned in the Training Stack
A survey of where refusal is learned (pretraining, tuning data, selection head, post-hoc threshold) and how each interacts with proper-scoring-rule calibration.
Rationale Supervision Versus Soft Targets for Encoder Decision Heads
Whether filtering, loss weighting, or auxiliary rationale supervision beats plain soft-target distillation for single-pass encoder decision heads.
Self-Host Crossover Throughput: When a Rented Encoder Beats the Cheapest API
A crossover analysis of the sustained decisions per second at which a Maestro-class encoder on a budget GPU beats the cheapest measured API.
How Many Examples per Label Do Small Decision Encoders Need?
Published few-shot evidence on Banking77-scale intent sets, and whether the GLiClass 8-shot recovery needs fine-tuning or works with frozen encoders.
Shortlist Then Rerank for High-Cardinality Typed Decisions
A protocol pairing a bi-encoder shortlist with a joint rerank pass on Banking77, reporting recall, conditional accuracy, and calibration separately.
Beating Max-Softmax: The Cheapest Abstention Triggers for Small Decision Encoders
Negative energy and margin beat max-softmax thresholding at the cost of one reduction over logits, and each wins on different slices.
The Smallest Selection Head That Beats Thresholding
A survey of minimal selection-head and abstain-class designs, and the smallest configuration that beats per-bucket post-hoc thresholding on risk-coverage.
Grouping Calibration Parameters by Option Count
Per-count temperature grouping flattens calibration error as label sets grow, while flip rate needs a separate order-robustness lever.
Symmetric and Reverse KL for Tail Dark Knowledge After the Refit
A survey of whether the direction of the distillation loss changes what tail dark knowledge survives on high-cardinality choice heads after recalibration.
Usable Decision Context in Single-Pass Encoder Heads
A report on how much ModernBERT-class context a single-pass Laya-style head can use before option text crowds out state, and what longer truncation must prove.