Review terms

Review terms

Status: Draft — may change (2026-10-19)

Where it sits: beside Board, which names who decides; this page names who reads before they do. It applies to TeamCLI's own texts; a team that uses TeamCLI sets its own under 006 Consensus (its commitment 11). Every rule names its source.

1. The floor

1. At least three reviewers, from at least three labs, none of them the lab that wrote the text. With fewer, a round does not count, however many said yes. Source: the Charter, Part 3; the board setting of 2026-10-07, rule 6; the review process, "A round".
2. Never the authoring model. A reviewer is not the model that drafted the text, whatever lab runs it; a lab is the coarse unit and the model the fine one, and both are checked. Source: the review process ("none the author's and never the authoring model"); the owner's correction of 2026-10-15 relayed for Board (§3.8: "an independent reviewer is not the authoring model").
3. At most one reviewer from the lab that coordinates the round, so that the organiser's lab never makes a majority. Source: the owner's decision of 2026-10-07 that no lab is reserved for a role and the 4 October rule is removed (the decisions log, entry 10), and the practice since RFC 001 (no round had more than one reviewer from the coordinating lab).
4. Where only one lab, or one model, is available, the decision says so, and the round counts as one lab's, below the floor. Source: RFC 006 commitment 6; Board §3.7.

2. Who may read which text

1. A private text is read by the labs the owner admits for private material. The rule is set per team and revisable by the owner; the standing rule since 2026-10-07 admitted Gemini, Grok, Codex and the Claude seats, and held DeepSeek, GLM and Kimi to public material; on 2026-10-11 the owner admitted DeepSeek for the Charter and widened that round to GLM (an owner decision). Source: an owner decision; the owner's data rule of 2026-10-07.
2. No lab is reserved for a role. Which lab holds which reviewer seat is the team's own setting; a Claude seat may review where the floor allows. Source: the owner, 2026-10-07 (the 4 October rule removed).
3. A reviewer reads what the round's brief names and nothing of the author's working files, so that a self-review by the author's lab is a work file and never a verdict. Source: the review process.

3. How a reviewer answers

1. Alone, blind to the other verdicts, on the text frozen at the brief's commit, and the verdict names that commit (frozen-at). Any edit to an in-scope file voids the round. Source: the review process; the round-1 brief; RFC 006 commitment 4.
2. One file per reviewer and round, signed with the model as run, ending with the verdict line: VERDICT: PASS, ACCEPT-WITH-CHANGES with the exact changes, or NO with the blocker, each finding cited to a file and a line. A file without the verdict line is no answer. Source: the review process; the board setting, rule 1.
3. A reviewer's lab reads its own lab's text never, and a reviewer swapped in reads the whole text, not the changes since the last round. Source: an owner decision; the board setting, rule 5.

4. Silence, the limit, and the replacement

1. Silence is never a yes. Source: RFC 006 commitment 5; the board setting, rule 1.
2. Every round has a time limit, set when it is dispatched and sized to the text; when it passes, the seat is looked at, not waited on. Source: the board setting, rule 2.
3. The cause is found and recorded before any retry: a question on the screen, the lab's limit, the network, a crash; "did not answer" and "its lab was unreachable" are different facts. Source: the board setting, rule 3.
4. One retry on the same seat once the cause has cleared; then a replacement, from the same lab if there is one, else another lab, named in the record. Source: the board setting, rules 4 and 5; the Charter ("asked again, then replaced").
5. A lab unreachable for the whole session does not block acceptance: it is replaced, and the record says the text was not read by that lab. Source: the board setting, rule 7.
6. A person who does not answer is not replaced by an agent: the decision waits, or goes to whoever the rule names. Source: RFC 006 commitments 5 and 7.

5. Rounds and their end

1. Rounds continue to unanimity. Source: the owner, 2026-10-07; the Charter.
2. A point that comes back three rounds running without converging stops there: the board decides it and the disagreement stays on record with the reviewer that held it; nobody holds a veto by repeating a point. Source: the Charter; the board setting; RFC 006 commitment 7.
3. A point once decided is reopened only if the new text contradicts itself or the decision. Source: the Charter.
4. The owner is told at every round when it closes: each verdict, what was decided on each point, what goes to the next round. Source: the owner, 2026-10-07; the board setting.
5. Two rounds may run at once on two texts, never two rounds on one text. Proposed (the plan of 2026-10-17, open point 5); not yet the owner's word.

6. What a review is not

1. A round confers nothing. Unanimity is evidence, not an authorization; the acceptance is the board's and the act still needs the capability the Kernel names. Source: Kernel §9, §16; Board §5.
2. The lab predicate at a gate is the Kernel's (§5), owner-set and enforced where the Kernel's constraint program is built; at v0.5.0 it is not built (Kernel §5's Today). These terms are advised and read by the RFC process alone. Source: Kernel §5; the Kernel's Policies section's catalogue, review-terms.
3. An adversarial read is a review of its own kind, asked for by name in the brief, and its verdict counts like any other. Source: the practice since 2026-10-17 (the adversarial reviews recorded since that date).

7. Open points, for the owner

1. Whether the floor of three labs stays as a number here or becomes the project's review-terms value, read by the process (Board §7, question 5).
2. Whether the lab or the model is the unit of independence where the two disagree (Board §7, question 4; 006 question 1).
3. Whether the owner confirms the private-material rule's current admitted-lab list as the standing setting.
4. Whether two rounds may run at once (§5.5). One round-1 review asks to restrict the proposal to texts whose open points do not cite each other, citing the planning of 2026-10-17 §7 and open point 5; a second asks to confirm the existing two-text wording, citing the review process, “A round”, rule 3. Neither correction is applied to §5.5. On 2026-10-19, fifteen rounds ran at once on texts that cite each other's open points; they were made safe by two organiser rules: other texts were read at the round's frozen commit and a change there did not void it, and the rounds shared one deadline with no fold landing before all had closed. For the owner: Rounds may run at once on different texts, never two on one text; concurrent rounds read each other's texts at their frozen commits, share one deadline and fold only after all have closed. Proposed; the owner's word at acceptance, as §5.5 already says.
5. Whether §6.1 should add that the board votes only after reviewer unanimity and that each member votes its own vote, never lent to the author, as a round-1 review requests. This adds rules to the current text and is deferred rather than silently adopted; should they live here or in Board with a movement record?
6. Whether to connect this text's numeric floor to a document-valued project policy, require both lab and model independence in the disputed cases, and confirm a revisable team/entity list of admitted labs, as, lines 64–67, 71–77 and 81–84, proposes. These are proposed answers to questions 1–3, retained as questions; no standing admission or new counting decision is adopted.
7. Whether to change §6.3's request from “by name in the brief” to “by the organiser”, as a round-1 review proposes. This changes the rule's condition and is deferred rather than silently adopted.
8. What exact fold note should record the existing Board §4.3 detail on both positions and this batch's shared-deadline exception to §4.2? A round-1 review suggests optional wording and gives no exact replacement. Its line 52 records a stale brief quotation.