Glossary
The words these pages use in one particular way, and the acronyms they assume you know, each with the chapter that says the rest. A term here means the same thing on every page; when a page seems to use one differently, the page is wrong, not the glossary.
Terms
- claim — the finder's account of a finding as a voter receives it: the file, the line, the category and the summary, fenced and introduced as unverified. It holds the only text a model wrote that a voter reads. Prompts
- confidence — the highest level,
high,mediumorlow, whose threshold a finding's kept votes reach, computed in code by the policy. No model reports one: a vote has no confidence field. Lenses and policy - data notice — the part of every system message that names each fenced block of the request by its exact tag, and says what it holds. The code writes it, not the profile. Prompts
- discarded — a verdict that reached no confidence level, and whose failed votes, had they been keeps, would not have reached one either. Lenses and policy
- evidence — the diff's lines around a finding's cited line, in its hunk, read from the diff by the harness and never taken from the finder. The JSON result
- fence — the pair of tags around one block of data in the user message,
<untrusted-diff-1a2b3c4d>and its closing tag, with a random nonce for each block of each request, so the data cannot close its own block. Friction, not a control. Prompts - finder — the model that reads the diff in the research stage and proposes the findings, through the profile's research provider. Start here
- finding — a candidate issue the finder reports. The harness checks each one against the diff: a finding passes the check only when it cites a line inside a hunk and, when the diff has per-file headers, one of its files. A finding that passes has a file, a line, a category, a summary and its evidence, and goes to the panel unless it repeats an earlier one; the others are dropped with a reason, or recorded as gaps. The JSON result
- flag — an advisory signal the harness reads from the diff by code, before any model sees it. It marks hidden characters, or text addressed to an AI reviewer, and tells a human where to look; it changes no verdict and no exit code. Output handling
- gating — whether a profile's kept
highfindings set exit code1, and its gaps exit code4: true unless the profile says otherwise. Profiles - guidelines — reference text a program loads and sets on a profile, fenced in every request of that profile and described to the model as criteria. Profiles
- harness — quorum's own code around the models: it reads the diff, builds every request, checks every reply and computes every verdict. Architecture
- hidden character — a character a reader cannot see or a terminal may act on. That covers the control and format characters, unpaired surrogates, the other default-ignorable code points, and the line and paragraph separators. Output handling
- incomplete — said of a review in which a finding is unreviewed, or
meta.incompleteholds a gap. The CLI exits4for an incomplete review of a gating profile, unless a gating profile kept ahighfinding. The JSON result - inert — said of text printed so that nothing in it acts on a terminal or a GitHub Actions log. Its hidden characters are shown as
<U+XXXX>in a line, and escaped in JSON. Output handling - kept — a verdict whose kept votes reach a confidence level. Lenses and policy
- lens — one angle of the panel: an id and the question its voter is asked. Each finding gets one vote per lens. Lenses and policy
- panel — the stage that sends every finding to every lens of the profile, one blind vote each, and computes its verdict. Also the voters themselves. Lenses and policy
- policy — a profile's rules for how many kept votes each confidence level needs:
unanimous,majority,any,off,{ minVotes }or{ minRatio }. Lenses and policy - profile — what a review looks for, as data: the research and vote instructions, the lenses, the policy,
gatingand optional guidelines.securityis the default, and the CLI runs every profile its config selects. Profiles - provider — how quorum reaches a model: an object with an id, a model and a
completefunction that takes a request and returns the reply's text. Providers - quote — the code the finder says a finding is about. It is only compared with the evidence, which sets
quoteMatched, and is never sent to a voter. The JSON result - research — the first stage of a profile: one call to its research provider, whose reply lists the candidate findings. Architecture
- sink — two things, told apart by what surrounds them. The event sink is the function a program passes as
ports.events, which receives every event of a review. In the security profile's instructions, the sink is the code a vulnerability ends in, which the finder quotes and the reachability lens asks about. Library API · Profiles - stage — one step of a profile's review: research, then dedup, which drops a finding that repeats the file, line and category of an earlier one, then panel. Architecture
- threshold — the number of kept votes a confidence level needs, once the policy is resolved for the profile's lens count;
nullfor a level that is off. Lenses and policy - unreviewed — a verdict that reached no confidence level while its failed votes, had they been keeps, would have reached one: neither kept nor cleared. It makes the review incomplete. Lenses and policy
- verdict — a finding once the panel has voted on it: its votes, the kept and total counts, and its confidence when it reached one. The JSON result
- vote — one lens's judgment of one finding,
keepandreasoning; or a failed vote,nullwith its error, which counts as a reject. The JSON result
Acronyms
- API — application programming interface: the functions and types
src/index.tsexports, and the HTTP interface of a model endpoint. Library API - CI — continuous integration: the jobs that run on each change, which gate on quorum's exit code. The CLI
- CLI — command-line interface:
quorumtyped at a shell. The CLI - CWE — Common Weakness Enumeration, the numbered weaknesses (
CWE-89) the security profile asks the finder to name as a finding's category. - JSON — JavaScript Object Notation: the format of the result on stdout and of every model reply quorum reads. The JSON result
- LLM — large language model: the finder and the voters are LLMs, reached through providers. Providers
- PR — pull request: a proposed change, whose diff quorum reviews and whose author is the attacker of the threat model. Threat model
- TTY — a terminal. With no file argument and stdin a TTY, quorum prints its usage instead of waiting for a diff. The CLI
