Glossary
Theory Delta uses a few precise terms throughout the site. Here's what each one means. Most labels on a finding link straight back to the relevant definition here.
- Finding
-
A single, fact-checked claim about how an agentic tool actually behaves — what works, what silently fails, and what the docs don't mention. Findings are the unit Theory Delta publishes.
- The delta
-
The one-line key takeaway of a finding: what is newly true in practice versus what you'd assume from the documentation or from a model's training data. It's the headline, not editorial opinion.
- Claims
-
The discrete factual statements a finding makes. The “11 claims” counter is how many separate assertions the finding is built from — each one is individually sourced or tested.
- Tested
-
How many of a finding's claims were verified first-hand — by running the tool, reading its source, or independent confirmation — rather than taken from documentation. “0 tested” means the claims are currently backed by source/doc review, not first-party execution.
- Content type
-
The category of a published piece. Two kinds appear in the corpus:
finding— A specific empirical result about one tool or behaviour.landscape— A survey that maps a whole area — multiple tools or patterns compared side by side.
- Confidence
-
The evidence tier of the finding as a whole. It signals how the claims were established, from strongest to weakest:
empirical— Executed and observed first-hand by Theory Delta.medium— Established by reading the tool's source, issues, or changelog directly.secondary-research— Drawn from documentation or third-party reporting, not yet verified first-hand.
- Evidence kind
-
The per-claim label in a finding's evidence table, describing how that specific row was established:
tested— Run in a real runtime by Theory Delta and observed.source-reviewed— Confirmed by reading the tool's source code, issues, or commits.independently-confirmed— Corroborated by a credible third party, not just the vendor.docs-reviewed— Based on the tool's documentation or release notes.
- Staleness risk
-
How quickly the facts underneath a finding tend to change. “High” means the area moves fast (APIs, limits, and behaviours shift between releases) so you should re-check against your own environment before acting. It describes the subject area — not whether this page itself is out of date.
- Lifecycle · Published
-
Every finding carries a lifecycle state, shown as a pill at the top. Only published findings are listed publicly:
published— Live and publicly listed — the normal state.in-flight— In fact-check or pre-publish review; claims may still change. Not publicly listed.retracted— Withdrawn — superseded or found wrong. Kept at its URL with a retraction notice.
- Fact-checked · corrections
-
“Fact-checked” is the date Theory Delta last ran a review pass over the finding. “Corrections” counts factual fixes made to the finding after publication. “0 corrections” means none have been needed since it went live — not that nobody has read it.
- Task hubs
-
Curated groupings of findings organised by what you're trying to do (e.g. “add memory to an agent”) rather than by which tool you're using. Browse them under “By task”.
- By tool · By task
-
The two ways to browse the corpus. “By tool” lists findings keyed to specific tools (the findings index); “By task” lists task hubs keyed to a goal. Same findings, two entry points.
- Scan
-
Point Scan at a repository or agent config and it checks what it finds against Theory Delta's catalog — surfacing the findings that apply to your setup.
- For agents
-
Machine-readable access for AI agents: an MCP endpoint and structured files (llms.txt, agent cards) so an agent can query Theory Delta before it makes a tool decision.
- Independent · evidence-traced · no vendor influence
-
Theory Delta takes no vendor sponsorship and accepts no editorial input from the tools it covers. “Evidence-traced” means every claim is either linked to a primary source or explicitly labelled with the evidence kind behind it — nothing is asserted without a traceable basis.