The five principles

TRACE names five governing principles that allow research teams to track, review, accept, control and evaluate AI-assisted work. They govern the process between adopting AI and the research outcomes it leads to.

AI-assisted research cycles through prompts, generated candidates, evaluation, selection, revision and reuse. A candidate and its evaluation can reshape the next instruction, and selected material can become an input to dependent work. The principles govern the instructions, evidence, decisions and dependencies surrounding each unit of work, whether a person or an agent performs the action. For each principle, a workflow should specify when an action is required, who or what may perform it, and what evidence or decision it must retain.

The principles concern work that one party produces and another relies on, which is the problem AI-assisted research makes pressing. They do not claim to describe good teamwork in general: planning, communication and training within a team are conditions of teamwork that the principles presuppose rather than govern.

The principles come from software engineering and security research. They are set out in a working paper by Amirhossein (Miros) Zohrehvand and Aldona Kapačinskaitė; Where TRACE comes from says how they were derived and what that derivation can and cannot show. They state what a team needs to do, whatever tools it uses, and TRACE-OS is one implementation of them: each section below gives what a principle requires, then what TRACE-OS keeps for it. Underneath all five is one assumption: a researcher remains responsible for the outcome.

1. T

   Track records

   Record what happened, in what order, and why, and keep the work’s current state visible to everyone who uses it.

   Shared version history, and a trace record for every agent session that commits work, linking the captured instructions to the agent’s choices, with their reasons, and to the changes they produced.

2. R

   Review against sources

   Compare an output with an explicit criterion or another source before others rely on it. Sources include the data and the logic of a claim, not only citations.

   Checking records that name the candidate, criterion, sources, method, outcome and limits; model objections must quote the text, and the quotes are checked.

3. A

   Accept independently

   Give each unit of work an owner, and separate producing the work from the decision to accept, return or refuse it, taken by an actor other than its producer. A researcher who authorizes an agent or a process to accept stays accountable for what it accepts.

   Decisions by a person, or by an agent or process under a named researcher’s authority, recorded against the version they cover and apart from model critique.

4. C

   Control transitions

   Check and authorize outside material and AI output before they enter the manuscript, the project’s data or its standing instructions.

   Intake records of who admitted, returned or refused outside material, which waits apart until then, and transition records for drafts moved into the shared manuscript and instructions adopted as defaults.

5. E

   Evaluate corrections

   Take what review or use finds back into the work, and keep a detected error open until the team has checked that the correction worked.

   A defect ledger linking findings, claimed fixes, recurrence and targeted rechecks of the original problem, one for each output a finding reaches.

T: Track records

What it requires. A workflow should connect each unit of work to the instructions, inputs, alternatives and decisions that shaped it, in a record the team shares. The artifact alone may not tell a collaborator enough to assess the work.

What TRACE-OS keeps. Each agent session that commits work leaves a trace record under ai/records/, a record of human direction and machine execution: your words in evidence.json, what the agent chose in summary.md, and what it checked, with its limits, in checks.jsonl. A commit you type yourself owes none. A session that made a consequential choice adds a decision account, decisions.jsonl: the question, who proposed and who selected the choice, why it was taken over the alternatives, what it used, and under what authority. The agent records each choice when it makes it, not at the end of the session. Before each commit, the agent brings the record up to date in one command, which drafts the commit body from it. Each commit names the human author, and a commit an agent executed ends with an execution marker, the Execution-agent: line.

Say to your agent

Show me the draft record before you commit: my words, what you chose, and what you did not check.

The agent follows the trace-state skill.

R: Review against sources

What it requires. A workflow should specify how a candidate is checked against the intended requirement before others rely on it. A passing check supports only what it examines, so the record must keep enough evidence to judge whether the check was adequate.

What TRACE-OS keeps. Every check names the candidate, the criterion, the method, the outcome, who checked and what the check did not cover, and can name the sources it compared the candidate with. A model review that diagnoses a defect in a text must quote the text, and the quote is checked word for word against the source. Model review is advisory: it is never acceptance on its own. A review run keeps its report and findings in ai/reviews/, so a coauthor on another clone can judge what was examined and found.

Say to your agent

Review my coauthor’s last commit against the task and the sources it rests on, and tell me which of your findings quote the commit word for word.

The agent follows the trace-review skill. For a single claim, claim-evidence-audit rates the claim against the evidence the project holds and says what it did not open.

A: Accept independently

What it requires. A workflow should authorize an actor independent of production to accept, return or refuse a specified unit of work. That actor may be a person, or an agent or process that a named researcher has authorized; the researcher remains accountable for the delegation and for relying on the accepted work. Checking produces evidence about a candidate, while acceptance records an authorized decision about it, so a favorable model critique is not an acceptance.

What TRACE-OS keeps. Decisions go into ai/records/acceptances.jsonl. When you state a decision to your agent, it records the decision in your name, with your words verbatim and its own model id as the recorder. An agent never decides for you, and an actor that executed the work, or selected a choice the record logs, is not independent of it: when the execution markers of the covered commits or the record’s decisions name that actor, the tool records its attempt as self-review. The tool cannot see production those records do not show, such as who proposed the work. Each decision names the version it covers, the commits the brief showed you, so accepting one unit of work does not accept later changes, and a commit that lands after you read the brief stays waiting.

Say to your agent

What is waiting for my decision?

Say to your agent

I accept number 1, the yearly shares. Record my decision.

The agent follows the trace-acceptance skill. See Accept, return or refuse work.

C: Control transitions

What it requires. A workflow should identify the checks and the authority required when material enters the project or advances into work that others rely on. There are three such transitions: admitting an input, moving a draft into the shared manuscript, and adopting an instruction as a default for later work. Checking whether material is suitable and authorizing its use remain distinct acts.

What TRACE-OS keeps. Outside material (a dataset, a source, a reference) waits in incoming/ inside the shared folder until someone decides on it. The decision gets an intake record in sources/intake.jsonl: where the material came from, its version, who decided, whether it was admitted, returned or refused, and what was checked first. The material moves into the project’s data only once it is admitted. A draft moved from workspace/sandbox/ into paper/, and an instruction adopted as the new default, each get a transition record in sources/transitions.jsonl that names the commit and the words that authorized it.

Say to your agent

Promote my results paragraph into paper/, then commit and push.

The agent follows the trace-ctransition skill. A transition record records the decision; the authority is the person’s words it names.

E: Evaluate corrections

What it requires. A workflow should connect a detected failure to an assigned response and to a recheck of whether that response worked. Recording an edit, or a later review not repeating the complaint, does not establish that the correction worked. When a standing instruction contributed to the failure, the correction must reach that instruction, and its effect on later runs must be checked.

What TRACE-OS keeps. The defect ledger, ai/defects.jsonl, gives each finding a stable id and keeps three names apart: who raised it, who repaired it, and who rechecked the repair. A finding reads verified only after a targeted recheck against the original criterion. When a finding reaches several outputs, such as every table built by the same code, the row lists them, and it verifies only when each has its own passing recheck. When a faulty default caused it, it verifies only after a later run under the revised instruction has been examined. The commit that fixes a finding names it after Fixes, so the report can join the correction to its recheck.

Say to your agent

Open a defect in my name: the overall share averages the yearly percentages instead of dividing total cases by total patents.

The agent follows the trace-evaluate skill.

Closed is not verified

closed means a later review did not raise the finding again. verified means a targeted recheck of the corrected work was recorded, with its criterion, method, outcome, checker and limits. The absence of a complaint is not evidence of a correction, so TRACE-OS never turns one into the other automatically. A row someone closes by hand is recorded as a repair claim, and waits for a recheck like a fixed one.

Reading the account back

The trace-check skill runs the project’s mechanical checks and the governing-actions report. For every commit since the project was set up, and for each of the five principles, the report says whether the action is recorded, missing, not applicable or unknown. For a recorded review, it also says whether the checks name their sources, only a criterion, or neither. Missing is a reported state, not a refusal: the records make gaps visible without establishing that every required action occurred.

Say to your agent

Run the trace stats and open the report.