Glossary

This page defines, in one sentence each, the terms the site uses for the workflow, its records and the decisions about its work.

Term	Meaning
Acceptance	An authorized decision to accept, return or refuse a candidate, recorded by an actor independent of production, with a named researcher remaining accountable when the decision is delegated.
Actor independent of production	Who may accept, return or refuse a candidate: a person, or an agent or process a named researcher has authorized, but never an actor that executed the work or selected a choice recorded in it.
AI-assisted research	Research in which a generative AI system produces or evaluates material that enters the work or informs a research decision.
Boundary transition	A point where material enters or advances into work others rely on: admitting an input, moving a draft into the shared manuscript, or adopting an instruction as a default.
Brief	One trace record on one screen: what was asked, what was chosen and why, what was checked and what was not, and the version it shows, which a decision then covers.
Candidate	A unit of work offered for checking or acceptance: the exact work a check, review or decision is about.
Check	An action that produces evidence about whether a candidate meets a criterion, and supports only what it examines.
Check row	One line in a record’s checks.jsonl naming the candidate, the criterion, the method, the result and the limits.
Closed	A finding that a later review did not raise again, which is an observation and does not establish that the correction worked.
Criterion	The explicit standard a check compares a candidate with.
Decision account	The record, in a trace record’s decisions.jsonl, of who proposed and who selected a consequential choice, why it was taken over the alternatives, under whose authority and with what evidence.
Default adoption	Adopting an instruction, such as a line in AGENTS.md or a skill, as a default for later tasks, one of the three boundary transitions.
Default owners	The people whose prompts or comments direct a change, who own it and its account unless an explicit assignment states otherwise.
Defect ledger	The file ai/defects.jsonl, which keeps each finding under a stable id with the outputs it reaches, its fix and rechecks, and keeps its raiser, fixer and rechecker apart.
Delegated actor	An actor independent of production (an agent, process or person) whom a named researcher has authorized to accept, return or refuse a candidate, with that researcher remaining accountable.
Direction	The human side of a trace record: written instructions, choices, expressed preferences and objections.
Disclosure review	The read of every record file for private or third-party material before the record is promoted, which is not acceptance.
Execution	The machine side of a trace record: what the agent ran and produced.
Execution marker (Execution-agent: line)	The last line of a commit a model executed, naming that model, which marks production and not authorship.
Faulty default	A standing instruction that contributed to a failure, whose correction verifies only through a recheck of a later run under the revised instruction.
Finalize	What you say once the draft record says what happened, so the agent completes the record in your name before it commits.
Finding	A detected failure entered in the defect ledger, from a review or raised by a person, which keeps a stable id while it is corrected and rechecked.
Fix claim	Fixes and a ledger id in the message of the commit that fixes a finding, by which the governing-actions report joins the correction to its recheck.
Governance routines	The six trace-* skills, which specify which instructions, choices, checks and decisions to retain while performing research tasks; see Skills.
Governing-actions report	The report saying, for every commit and each of the five governing principles, whether the action is recorded, missing, not applicable or unknown.
Incoming folder	incoming/ inside the shared folder, where outside material waits until someone admits, returns or refuses it.
Intake record	The row in sources/intake.jsonl that records the decision to admit, return or refuse material from outside, who made it, and what was checked first, including when nothing was stated.
Later run	A trace record produced later under a revised standing instruction, which someone examines to see what the instruction did.
Limits	What a check did not cover, stated in its row.
Model review	A model’s critique of a candidate, which is advisory and never acceptance on its own.
Named researcher	The researcher who authorizes a delegated actor and remains accountable for that delegation and for relying on the accepted work.
Needs you	The queue label for a record that waits on a person, shown with its reasons, and always for a record that logs no choice.
Principal	A person whose prompts or comments direct a change, and so the party an agent acts for.
Profile	What setup installs: academic (the default, with paper and slides) or core (without them).
Promotion	Moving a draft from workspace/sandbox/ into the shared manuscript in paper/, one of the three boundary transitions.
Queue	The numbered table of records with commits nobody has accepted yet, with each record’s owners, followed by the unaccepted commits that name no record.
Refuse	Deciding against a candidate after examining it, a recorded decision that leaves the work unaccepted.
Registry	One of the shared files that hold acceptance decisions, intake records, transitions and findings; see Registries.
Research skills	Skills for substantive research tasks, such as processing a source or auditing claims, of which TRACE-OS ships four as worked examples you can replace with your own.
Return	Sending a candidate back for correction with an objection, a recorded decision that leaves the work unaccepted and opens a finding in the defect ledger.
Review report	What a model review run writes to the tracked ai/reviews/<run>/: the report, the findings as returned, the prompt and the rows it recorded.
Review state	Where a unit of work stands, from not reviewed to accepted, returned or refused; see Review states.
Routine	The queue label for a record with no listed reason to need a person, which triages and never accepts (not to be confused with the governance routines).
Sandbox	workspace/sandbox/, where drafts live until you ask for them to move into paper/.
Self-check	A recheck by the person who made the fix, labelled as such.
Self-review	A decision recorded by an actor that executed the covered work or selected a choice recorded in it, which is not independent of it, so the decision is recorded as self-review and never as acceptance.
Shared instructions	The instructions every agent in the project reads, kept in AGENTS.md.
Skill	A written procedure in .skills/ that tells an agent how to do one part of the work and is itself a standing instruction; see Skills.
Staging	workspace/record-staging/, the ignored private folder where a record is drafted before it is promoted.
Standing instruction	An instruction that applies to later tasks by default, such as AGENTS.md, WORKFLOW.md or a skill, whose adoption is recorded as a default adoption.
Targeted recheck	A check of a corrected object against the original criterion, recorded with its method, evidence, checker and limits, without which no finding reaches the verified state.
TRACE	The five governing principles: track records (T), review against sources (R), accept independently (A), control transitions (C) and evaluate corrections (E).
Trace record	The record kept for an agent session that commits work: a record of human direction and machine execution.
Trace stats	TRACE-STATS.md, the Markdown snapshot of project health rendered from the workflow check.
Transition	A row in sources/transitions.jsonl recording a promotion into paper/ or a default adoption.
Undeclared actor	Someone who recorded a decision without saying whether they were a person, a model or a delegate, which is never acceptance.
Unit of work	What is produced, checked and decided on, such as a commit, a draft, a table or a passage.
Verified	A finding whose correction a targeted recheck supports, recorded with its criterion, method, outcome, checker and limits.
Workflow	The arrangement of work, checks and decisions by which researchers assign tasks to an AI system, check its outputs and decide on them.
Workflow check	The command that runs the project’s mechanical checks and writes what it found to workflow-check.json.