---
name: "standard-prioritisation"
type: standard
license: CC BY-NC-SA 4.0
description: The fleet's quality bar for ranking work — severity tiers for audit and release-gating work, Impact × Effort × Confidence scoring for roadmap and intervention work, cost of delay ÷ size (WSJF) for the queue waiting on the Owner, and the known/assumed/unknown confidence overlay that routes high-severity unknowns to a research-first track. Also carries the reserved-WIP-slot rule for splitting one delivery capacity between a committed goal and everything else. Read before scoring or sequencing any findings, opportunities, interventions or Owner decisions (skill-prioritise, skill-discovery-audit, task-cycle-planning, skill-time-management, skill-ops-retro) so priority is derived the same way and a number can be defended. Owned by Sonar for Methods A and B; Method C and the capacity rule are read across the estate. Surfaced by the a1.Muse engagement; grown as it is reused.
---

# Prioritisation standard

How Sonar decides what matters most, and in what order. This is a *standard*, not a convention: the axes are derived deliberately and a score can pass or fail review, rather than being an arbitrary-but-shared choice. It explains **how each axis is derived**, so two people scoring the same finding land in the same place. Owned by Sonar for Methods A and B; Method C and the capacity-allocation rule below are read across the estate — by `task-cycle-planning`, by Aide (`skill-time-management`) and by the Pulse drain — so the artefact is the fleet's ranking standard, not one agent's. Read by every skill that ranks or gates work.

> *This is the canonical copy of this standard. Any human-readable A1 client version is a downstream rendering — edit here.*
>
> **Note on type.** This is the first `standard-` artefact. `core-operating-model` defines *standard* as a Layer-5 atomic rule (a quality bar you pass or fail), distinct from a *convention* (an arbitrary-but-shared choice). `convention-artefact-format`'s type enum and `build-manifest.ts` do not yet list `standard`; adding it is a small follow-up so the agents site classifies this correctly.

## Pick the method by context

Three methods, chosen by the work in front of you — never two at once on the same item:

* **Method A — Severity tiers.** For audit findings and release-gating: "how bad is this, and does it block?"
* **Method B — Impact × Effort × Confidence.** For the roadmap and interventions: "what's the best thing to do next?"
* **Method C — cost of delay ÷ size (WSJF).** For the Owner's own queue: "of the things waiting on Aled, which one first?"

A finding is triaged with A; the intervention that answers it is scored with B.

## Method A — Severity tiers

Three tiers, for the design-audit and release-gating framing.

* **Critical** — blocks the core job, breaks a release gate, or causes real harm (accessibility failure, data loss, broken purchase). Must be fixed before ship.
* **High** — materially degrades the experience or conversion but does not block. Fix soon; schedule into the next cycle.
* **Low** — a real but minor issue. Fix opportunistically.

**Accumulation rule.** Severity is not only per-item. Where many High findings cluster at one point in the experience, the *point* escalates: as a guide, more than three High findings at a single step escalate that step to Critical, because their combined effect blocks the job even if no single one does. Accumulation is assessed per step, not across the whole audit.

## Method B — Impact × Effort × Confidence

For sequencing interventions. Each axis is High / Medium / Low, derived as follows — the rubric is the point of the method, so derive, don't guess:

* **Impact** — how much this moves the metric that matters (the core-job conversion or the named KPI).
  * *High* — moves the primary metric materially, or unblocks a Critical.
  * *Medium* — moves a secondary metric, or a modest move on the primary.
  * *Low* — marginal or indirect effect.
* **Effort** — cost to deliver, in time and complexity (inverted — low effort is good).
  * *Low* — days; one owner; no new dependency.
  * *Medium* — a cycle; some coordination or a new component.
  * *High* — multiple cycles; cross-team or a structural change.
* **Confidence** — how sure we are the impact and effort estimates hold (see overlay below).
  * *High* — grounded in evidence (data, direct observation).
  * *Medium* — reasoned from analogy or partial evidence.
  * *Low* — assumption or unknown.

Sequence by Impact first, break ties on Effort (prefer low), and treat low Confidence as a flag to de-risk before committing, not a reason to drop. Plot the set on an Impact × Effort 2×2 to make the quick wins (high impact, low effort) and the big bets (high impact, high effort) visible.

## Method C — cost of delay ÷ size (WSJF)

For ranking what waits on the Owner. Rank by **cost of delay ÷ size**, highest ratio first.

The denominator is the point. Urgency × importance carries no size term, so it cannot separate two items that sat at identical priority and opposite merit — a thirty-minute call releasing seven blocked notes and the recurring CI-watch and egress classes, against a fifteen-minute call releasing nothing (APP-965 against APP-633; both Urgent, at 24 and 49 days). A method without a size term cannot tell those apart, and for two months nothing did (APP-1170).

**Size** is the ticket's native Linear estimate on the Exponential scale, written by Aide at sweep time (`skill-attention-sweep`): 1 = 15m · 2 = 30m · 4 = 1h · 8 = 2h · 16 = over two hours and needing a split. Nothing new is instrumented. **A missing estimate is not a small item** — on an Owner-lane ticket it is an Aide defect (`convention-linear`), so an unestimated item is reported as unranked rather than defaulted to 1, and the gap shows as a gap.

**Cost of delay** is derived from three axes, never guessed — the rubric is the point of the method, as it is for Method B:

* **Release value** — how much work this unblocks. Countable: open `blockedBy` relations pointing at the item, plus dependents named in prose.
* **Recurrence cost** — whether not doing it reproduces an error. A fix that stops a daily failure compounds; a one-off does not.
* **Time criticality** — whether the value decays if it happens later, or is flat.

**The pull rule.** When capacity frees — a hold window opening, an item clearing — pull the highest-ratio item, never the oldest and never the loudest.

**Switching cost is a real cost the ratio does not carry.** Where several items share a subject, working them contiguously is cheaper than the ratio implies — so batch within tier and let the batch inherit its strongest member's rank. The ordering rule this implies lives in `task-owner-standup` (*Step 3 — Write the agenda*), which is where a queue is actually sequenced for a sitting; here it is only noted that the ratio does not price it (APP-1237).

**Age breaks ties; it does not set rank.** Two items at the same ratio are ordered by age; an older item at a lower ratio does not overtake a newer one. This is the size-aware form of the consequence-before-age rule in `skill-operating-loop`, *Escalating an Owner-blocked stack*.

**Who computes it.** The ranking of the Pulse queue's decisions-needed set is the **refine pass's**, computed where that set is written (`skill-ops-retro`). It is not the attention sweep's: `skill-attention-sweep` is scoped to the committed cycle and explicitly excludes `PUL-5E` notes (APP-686), and a `PUL-5E` note is never in a cycle by construction (`convention-linear`, *Never in a delivery cycle*), so the sweep structurally cannot see the set it would be ranking. Aide writes the size term and nothing else.

**The method is inert until the estimates exist.** Aide writes the size term, so while Aide's scheduled task is not running the Owner-lane tickets carry no estimate. A pass that ranks under that condition states it and lists the set as unranked; it does not rank on a default.

Worked case, and it is the case that produced the method:

| | Cost of delay | Size | Ratio |
| -- | -- | -- | -- |
| APP-965 | releases seven blocked notes, plus the CI-watch and egress classes recurring on every delivery run | ~30 min (estimate 2) | very high |
| APP-633 | releases nothing; nothing waits on it | ~15 min (estimate 1) | near zero |

Both were Urgent. Only the denominator separates them. (APP-1170.)

## The confidence overlay — known / assumed / unknown

Every scored item also carries a confidence tag, independent of the score:

* **Known** — backed by evidence.
* **Assumed** — a reasonable belief, not yet verified.
* **Unknown** — a gap; we don't yet know.

**The routing rule:** a high-severity or high-impact item tagged **unknown** does not go on the build track — it goes on a **research-first track**, to be turned into a known before it is committed. This stops the plan betting heavily on guesses. Low-stakes unknowns can ride along; high-stakes unknowns are research tasks first.

## Capacity allocation — reserved slots, not a percentage

When a single delivery capacity must split between a committed goal and everything else, express the split as **reserved WIP slots, not a percentage**. A percentage (50/50, 60/40, 70/30) is folklore — there is no defensible evidence for any particular ratio, and a percentage cannot be enforced until the period is already over. A reserved-slot policy is enforceable in one sentence at the standup — "this is not goal work, it does not get pulled" — where a percentage is not.

The policy holds only when it is **enforced structurally and then measured after the fact**:

* **Reserve a fixed count of the in-flight WIP slots for goal work** (with, where used, replenishment rules that keep those slots for the goal). The reservation is the structural enforcement, not a target on paper.
* **Measure the achieved ratio** — the delivered work by type once the period closes — and adjust the reserved count for the next period. The split is corrected by what actually delivered, never set by an a-priori percentage.
* **Stretch work never takes a reserved slot** until the required set is done — the reservation protects the commitment from the wish-list.

This is the mechanism `task-cycle-planning` uses to size a cycle toward its one goal, and that `assistant-aide` / `skill-time-management` use to weight the Owner's holds toward goal work (APP-815).

## Quick checklist before publishing a priority

- [ ] Right method for the context (severity for audit/gate, IEC for roadmap)?
- [ ] Each axis *derived* from the rubric, not asserted?
- [ ] Accumulation checked — clustered Highs escalated where they block a step?
- [ ] Confidence tagged, and high-stakes unknowns routed to research-first?
- [ ] Result legible as a ranked table and, for interventions, an Impact × Effort 2×2?
- [ ] Capacity split, where one applies, expressed as reserved WIP slots and measured after the fact — never an a-priori percentage?
- [ ] Tone clean per `cos.tov`?
