Patterns — the shapes this skill reuses

The recurring forms, each named once. A rule that instantiates a pattern cites it and stops — it does not restate the reasoning. A pattern named once and cited nine times is a form; nine paraphrases are prose, and a rule living in two files goes stale in one of them.

The meta-rule above all of these: form beats prose, and this was measured, not argued. An earlier regression gate found prose the mid tier ignored until it was re-formed into structure. When a rule does not hold, the repair is a form — a list, a required field, a gate — never a stronger sentence.

Adding a pattern: it enters when the same shape appears in three places under different names. Two is a coincidence; three is a form — the same evidence bar as making a skill (§8).

One deliberate non-pattern. Opsinist-style "nothing transitions itself" is not a law here: Multica's stage barrier — the parent waking when all stage-N sub-issues are done — is native, desirable, and exactly what the conveyor runs on. Where the platform moves things, the platform is trusted; what may not move by itself is the owner's surfaces (dates, priorities, acceptance criteria — §11).

Truth and derivation

1 · Store the atom, derive the rest. Measure once, at the smallest event; never store rollups as truth — they drift the moment anything is edited. On Multica the atoms live in the platform (issue runs, issue usage, comments); our rollups (_ops/analytics/, ECONOMICS.md, the status digest) are derived and re-derivable, marked as such. Applies to: the cost ledger · the waste slices · /multica-ops:status · utilization review.

2 · Generate the index, don't maintain it. A frozen list ages faster than anything around it — generate it or search for it. Applies to: scripts/issues.py over the board · the workspace fingerprint · awesome-{topic} instead of a frozen catalogue · /multica-ops:mops team from agent list.

3 · Check-date, not a tick. Every recorded fact that can change carries when it was verified; past its recheck it is unknown, not fine. Applies to: _ops/TOOLING.md rows · STACKS entries · sources/SOURCES.md · prices · platform caps quoted in these files.

4 · Provenance is mandatory. Origin · when · who · what version. Without it /multica-ops:upgrade cannot tell what it is updating and /multica-ops:audit cannot tell what is old. Applies to: imported skills · personas (grounding artifact, consent) · assets (_ops/assets.md) · imported issues (source_id / source_url).

5 · Evidence, not verdicts. A tool that pattern-matches produces findings; a clean report is not approval and a flag is not a rejection. Who decides what a finding means is named separately. Applies to: skill-import scanners · synthetic reactions · audit output · cost slices.

6 · The rung travels with the claim. Claims carry measured › cited › recalled › judgement call (or unknown), and a lower rung never borrows a higher one's authority — including across a handoff: quoting does not promote. Applies to: issue bodies · research findings · ICE scores · expert verdicts · the two audience pyramids (a live reaction › a twin's › a validated's › a proto's — never pooled with each other or with token counts).

Time and change

7 · Effect at the next boundary, never mid-flight. A change to the machinery applies to the next dispatch; work in flight finishes on the bytes it started with. Applies to: /multica-ops:mops autonomy (boundary-safe) · guide edits (batched at /multica-ops:mops sync) · skill versions (one version per feature) · /multica-ops:mops pace.

8 · Twice is the evidence bar. Once is a task, twice is a pattern, and "we might need it" is neither. Naming both occasions is part of the proposal. Applies to: making a skill (/multica-ops:skill create) · adding a company-specific docs guard · adding a pattern to this file.

9 · Promote what outgrew itself; don't nest deeper. When a part needs parts of its own, it stops being a part — and Multica's hierarchy is two levels by design, so the promotion is real, not cosmetic. Applies to: a sub-issue that needs sub-issues → its own issue · a quick job → a crew → a company · a workspace skill that proved itself → its own repo (/multica-ops:skill release) · a runbook procedure → a skill. And seeing it is not doing it. A transition noticed ends in a named offer, never in an open question — "this becomes X, carrying what it already produced — yes?" Measured in opsinist 2026-07-31: runs recognised every seam unprompted, named the recurrence bar and the crafts involved, and moved none of them, closing with "finish it as it stands, or step back?" — an answer that reads as competent status while the hour quietly becomes a roadmap. The destinations here are the platform's own, so the offer can name them exactly: the outgrown sub-issue loses its parent link and keeps its key; the friction seen twice becomes a tooling issue with both dates in it; the milestone that spans crafts becomes its own issue with the children's keys intact. The owner still chooses — but between named acts, not between moods.

10 · Ages like a request. Anything pending carries its age and surfaces past a threshold, because a wait nobody chased and a wait everybody forgot look identical from outside. Applies to: owner approvals · escalations · grants · ripe _ops/LATER.md items · a live participant's round.

11 · Nobody edits the bar they are measured against. Locked (acceptance criteria, rubrics, the budget cap, the guide's invariants) · editable (code, specs in flight) · append-only (DECISIONS.md, ledger, FIELD-NOTES.md, this skill's own definition inside a company) · human-only (spend, credentials). Applies to: review gates · dispute settlement · self-editing · trust proposals.

Interaction

12 · The bridge. A conversation with an open outcome: most exchanges end in an answer and leave no footprint; the ones that cross ("let's build it") seed work and leave the conversation as it was. No mode to switch. Applies to: /multica-ops:consult → a project · a status question → a fix · an audit finding → an issue.

13 · Earned, never reflexive. Unprompted advice names the consequence of what just happened, while the decision can still change for free — at most two options, one concrete sentence, one nudge each, never nagging. A recommendation repeated after an answer is not diligence. Applies to: one-step-ahead advice · ripe deferrals · settings friction · the crew→company offer.

14 · The shape of a choice is decided before the choice is put. Readable from the ground → read it, never ask. A default is defensible → a filled-in form needing a nod. No default is defensible → a real choice, at most two options with consequences. Applies to: day zero (read, don't ask) · the interview waves (defaults visible) · the two hard gates (real questions) · where the code lives (never guessed).

15 · One position on a ladder, not a row of switches. Where things are ordered by one property, the choice is where to cut — one value to state and change — not one boolean each. Applies to: the free-first tool ladder (STACKS) · control level (hands-on · checkpoints · hands-off) · /multica-ops:mops pace · model tiers.

Honesty

16 · Honest enforced_by. Every gate declares what actually holds it — request · validator · git-host · platform · prose-only — and the prose-only rules are listed by name. The four lenses are the compensating control for everything prose-only. Applies to: the gates table (PLAYBOOKS → Gates) · branch protection with no remote · start dates · the review-not-author rule.

17 · Radical transparency: the record says how, not only what. Anything that produced a result names the model that answered (not the one asked for — gateways fall back), the helpers spawned and their tiers, the rung of every claim, and what enforces each rule invoked. A stated gap reads as a fact; an omission reads as a negative. Applies to: the ledger · review verdicts · /multica-ops:mops validate output · status digests · DECISIONS entries.

18 · Three-way compare on drift. Theirs changed · ours changed · both — surfaced with options, never silently merged; hashes over content, never modification time. Applies to: the workspace fingerprint (attribute before asking) · the tracker bridge (direction of truth per field) · skill upgrades (diff against the version you screened) · adopted host design systems (preview the host's new version against the delta layer).