03 / FOUNDATIONS
How models, context, and memory fit together
A practical map of generation, working context, retrieval, tools, and systems of record.
Evergreen explainerA practical field guide to AI systems
Find the fit.
Know the limits.
Understand the layers, ground the work in evidence, choose the simplest reliable system, and count the cost of results you can actually use.
Open the reference cabinetTask What does success require?Evidence How was the claim tested?Limits Where does it stop working?
Open the reference cabinet
Useful methods.
Explicit assumptions.
05 / RESEARCH
A source-grounded research workflow
Move from discovery to claim ledger, synthesis, counterevidence, and publication checks.
Step-by-step workflow06 / SYSTEMS
Retrieval, tools, workflows, or agents?
Match system complexity to path variability, external evidence, and action risk.
Architecture guide07 / RELIABILITY
Diagnose the system before rewriting the prompt
Separate model, retrieval, OCR, schema, tool, permission, and interface failures.
Diagnostic order08 / SAFETY
Safe AI workflows: data, tools, and approval
Set the privacy, permission, confirmation, recovery, and human-review boundary.
Risk guideTOOL / ECONOMICS
Accepted-result calculator
Use your own costs, review time, retries, and acceptance rate. Calculation stays in your browser.
Interactive toolFrom the original research / 2022
See the method
behind the measurement.
A source diagram worth reading alongside the evaluation checklist.
Scroll sideways to inspect the full diagram
Multiple questions for the same scenario.
Open full-size diagram ↗HELM’s original diagram makes the evaluation dimensions visible. The checks describe evaluation coverage in this historical illustration; they are not model scores.
Source: Stanford CRFM · 17 November 2022 ↗What counts as evidence?
A published claim.
An observed result.
Keep them distinct.
The starting collection explains how to evaluate models. We do not imply that an illustrative task is a benchmark run, or that a vendor’s score predicts your workflow.
Read our evidence policy →Retrospective archive
The record
The papers, releases, benchmarks, standards and rulings behind practical AI use, read for what they measured and what they left open.
100 entrys, each prepared 16 September 2026 from documents published earlier and cited in full, and published here on 18 September 2026. No entry was published on the date of the event it describes.
Open the record