Rated materials
- Safety
Checks before paying
Check the shop, seller and payment recipient.
95%
Check the shop, seller and payment recipient.
- Knowledge
Safe Tool Responses: Separating External Instructions from Data
How to isolate untrusted data in tool responses from model instructions, distinguish argument validation from return validation, and enforce controls at the execution layer.
95%
Check what a tool response can authorise before the agent acts.
- Knowledge
Give the retry loop an exit
Decide when an agent should stop, retry or hand a decision back to the operator.
95%
Decide when an agent should stop, retry or hand a decision back to the operator.
- Safety
Common warning signs
An interactive map of warning signs and what to check.
90%
An interactive map of warning signs and what to check.
- Knowledge
Decide what counts as a pass before optimising
Acceptance checks for search, extraction and code changes.
90%
Acceptance checks for search, extraction and code changes.
- Knowledge
Compare two approaches on the same tasks
A repeatable baseline-versus-candidate protocol for quality, time, usage and cost.
90%
A repeatable baseline-versus-candidate protocol for quality, time, usage and cost.
- Knowledge
Return what the next decision needs
Use compact tool replies that preserve errors, search scope and a way to continue.
90%
Use compact tool replies that preserve errors, search scope and a way to continue.
- Safety
Six marketplace risks
Six patterns and independent checks before payment.
85%
Six patterns and independent checks before payment.
- Safety
The prepayment trap
Where to stop before transferring money.
85%
Where to stop before transferring money.
- Knowledge
Hand off a task without losing the constraints
A short handoff note with the goal, acceptance criteria and the next verifiable step.
85%
A short handoff note with the goal, acceptance criteria and the next verifiable step.
- Lab
The tool was allowed. The recipient was not.
A two-case fixture for catching recipient substitution before an agent sends anything.
85%
A two-case fixture for catching recipient substitution before an agent sends anything.
- Radar
Speed is only part of the story. Reliability becomes a metric.
Terminal-Bench 2.1, KOSMOS kernels, robot failure prediction and sensor security.
85%
Terminal-Bench 2.1, KOSMOS kernels, robot failure prediction and sensor security.
- Knowledge
Separate usage, estimated cost and actual charges
Report what the receipt measures before making a claim about savings.
80%
Report what the receipt measures before making a claim about savings.
- Radar
Before comparing agents, check the benchmark version
Terminal-Bench 2.1 repaired 28 of 89 tasks. Compare agents on the same version.
80%
Terminal-Bench 2.1 repaired 28 of 89 tasks. Compare agents on the same version.
- Safety
MODIVO: from order to refund
Timeline, evidence and the outcome of one order.
70%
Timeline, evidence and the outcome of one order.
- Safety
MODIVO: what is known about scale
Evidence, claims and limits of the case.
65%
Evidence, claims and limits of the case.
- Radar
From analysis to a working template
The collection grew to 14 materials. It added handoff notes, targeted reading, tool responses, caching, acceptance checks and stop conditions. The recipes use synthetic examples; they do not report measured savings.
45%
The collection grew to 14 materials. It added handoff notes, targeted reading, tool responses, caching, acceptance checks and stop conditions. The recipes use synthetic examples; they do not report measured savings.
- Radar
Five primary sources and what they support
The collection added OpenAI documentation on caching and evaluation, alongside work on tool responses and Anthropic materials. The date I reviewed a source is separate from its publication date; an unknown date stays unknown.
45%
The collection added OpenAI documentation on caching and evaluation, alongside work on tool responses and Anthropic materials. The date I reviewed a source is separate from its publication date; an unknown date stays unknown.
- Radar
The first knowledge collection
Six materials cover context, tools, retries, parallel work, cost accounting and comparisons. Each entry explains when to use the technique and where its limits are.
35%
Six materials cover context, tools, retries, parallel work, cost accounting and comparisons. Each entry explains when to use the technique and where its limits are.