A rubric is built.
Primwright Enterprise Demonstration 6 of 7. "Rate actionability 1–5" — score three synthetic responses against the unanchored instruction, then against the anchored version with worked examples. Nothing here is a customer result.
"Rate actionability 1–5" — score three synthetic responses against that instruction alone, then against the anchored version with worked examples. Feel the difference an anchor makes, then see the versioning discipline that keeps a rubric honest over time.
01 Score, anchor, re-score — "Build the dimension with us."
Step 1 — The dimension, unanchored
"Score these three responses."
Actionability
Rate how actionable the response is, from 1 (not actionable) to 5 (very actionable).
That is the entire instruction — the way most scoring rubrics actually ship. Score each response 1–5. There is no score for you; the point is to feel the ambiguity.
R1 · My invoice shows a charge I don't recognize.
"I've opened a billing review for invoice INV-2041. A specialist will investigate and email you the findings within 2 business days; if the charge is an error, the reversal is automatic and you'll see it on your next statement."
R2 · Can you change the email on my account?
"Sure — I can update that for you. Which email address would you like to use going forward?"
R3 · The app keeps logging me out on my phone.
"Sorry you're running into that. Our team is aware of login issues and working on improvements to the mobile experience."
0 of 3 scored.
Step 2 — The dimension, anchored
"Now score them again — with anchors."
The anchored dimension, as revealed in the interactive walkthrough above:
Actionability — anchored
Actionability: does the response tell the customer exactly what to do next, or what will happen next, with enough specificity that no follow-up question is required to act on it?
- 5 — Fully actionable: States the next step AND the timing AND what the customer must do (if anything) — no follow-up needed.
- 3 — Partially actionable: States a next step but omits timing, the responsible party, or a required customer action — a follow-up question would be needed.
- 1 — Not actionable: No next step stated, or the next step is vague reassurance ('we'll look into it') with no specifics.
Worked examples:
- Scores 5 — "A billing specialist will reverse the duplicate charge within 3 business days and confirm by email. You don't need to do anything." Why: Next step + timing + customer action (none) — nothing left to ask.
- Scores 3 — "I've flagged this for the billing team and they'll take care of it." Why: A next step is named, but no timing and no confirmation path — the customer would have to ask 'when?'
- Scores 1 — "Sorry about that. We'll look into the billing issue and get back to you." Why: Vague reassurance only — no step, no timing, no action.
Re-score the same three responses under these anchors:
The simulated panel's anchored spread — Simulated judges, anchored [simulated — not human]:
- R1: 5, 5, 4; spread 1 — Anchor reading: 5 — Next step, timing, customer action, and the conditional outcome are all stated. [planted in the build module]
- R2: 3, 3, 4; spread 1 — Anchor reading: 3 — A clear next step, but the customer must supply information first — one follow-up is inherent to the task. [planted in the build module]
- R3: 1, 2, 1; spread 1 — Anchor reading: 1 — Reassurance only — no troubleshooting step, no timing, no action for the customer. [planted in the build module]
The spreads narrowed. The mechanism is visible: anchored criteria constrain interpretation. Honest note — this is measured across simulated judges on synthetic responses. It demonstrates the mechanism, not a human calibration result. A real calibration qualifies human judges against blind gold.
Step 3 — The versioning discipline
"What happens when calibration finds a gap."
G1.0 → G1.1
Trigger: Calibration surfaced a recurring question: how to score a response where the customer's own reply is the blocking step (R2 pattern)?
Change: Added one clarification: when the immediate next step is information the customer must supply, score 3 — the response is actionable once the customer replies. One worked example added. Prospective only; no re-annotation of prior batches.
One clarification, one example, prospective only. The guideline is versioned; every delivered unit traces to the version that governed it. That is the discipline the amendment exists to protect.
Method, stated honestly — "What this demonstrates — and what it doesn't."
It demonstrates:
- Why rubrics are built, not just written: an unanchored dimension admits multiple defensible readings, and the spread is where the work is.
- The mechanism by which anchors constrain interpretation — shown as spread narrowing across simulated judges.
- The versioning discipline: calibration findings become prospective guideline amendments, never silent edits.
It does not claim:
- That the spread narrowing is a human calibration result — the judges are simulated personas.
- That anchored rubrics eliminate disagreement — they constrain it, and the residual spread is reported.
- Any accuracy or quality figure for real evaluation work.
Next step — "Rubrics are agreed before scoring begins."
In a real pilot, the rubric is approved by you before a single output is scored. Start with a scoping conversation.
Start with a scoping conversation →
Nothing here is a customer result. Validated revenue from Enterprise services: $0.