Build Calibration Packets from Evidence, Not Adjectives — and Draft Promos You Can Defend

Workflow area: Hiring & Performance · Tier: Core · Hands-on time: ~35 minutes

Two failure modes dominate performance season. The first is the adjective packet: "consistently strong, great team player," which forces the calibration room to debate prose style instead of work. The second is the memory packet: built the weekend before, from whatever the manager can recall, which is mostly the last six weeks (recency) and the loudest wins (salience). AI fixes the assembly problem — mapping a cycle of evidence to a level rubric is exactly its kind of work — but only from evidence you curated, and only one report at a time. The moment a model compares your people to each other, you've delegated the one judgment that is unambiguously yours.

When to use — and when not to

• Use it every calibration cycle, starting three weeks outthe packet's value is arriving before the meeting with gaps flagged while there's still time to gather missing evidence.
• Use it for promo packetsthe same evidence-to-rubric mapping, extended with the claim-verification pass that promo committees quietly grade you on.
• Use it year-round, lightlya running evidence log per report (ten minutes a week) makes cycle-time assembly trivial and kills recency bias at the source.
• Skip any prompt that sees two reports at oncestack-ranking, "who's stronger," or cross-report comparison by AI is both an ethics line and, under emerging automated-employment-decision rules, a legal one. Comparisons happen in the calibration room, between humans.
• Skip it as a substitute for knowing the workif the packet surprises you, the packet isn't the problem; a leader who first learns what a report did from an AI summary has a different playbook to run (pb-02, 1:1s).

The recipe

1. 1.

   Curate the evidence yourself. For one report, gather the cycle's raw material: your 1:1 notes, design docs and PRs of note, peer feedback quotes, incident contributions, project outcomes. You choose what goes in — that curation is the judgment. Strip other people's names where they aren't needed, and never include another report's work as contrast.

2. 2.

   Map evidence to the rubric, one report per conversation. Enterprise tenant only:

   Copy

   You are helping me prepare ONE calibration packet. Do not
   compare this person to anyone, rate them, or recommend a
   rating — mapping and gap-finding only.
   
   Level rubric for [LEVEL]: [PASTE RUBRIC DIMENSIONS]
   Evidence (my curated notes for the full cycle): [PASTE]
   
   Produce:
   1. EVIDENCE MAP — per rubric dimension: the specific items
      that demonstrate it, each tagged with its source and
      date. Verbatim quotes where the evidence is a quote.
   2. GAPS — dimensions with thin or no evidence. Distinguish
      "no evidence gathered" from "evidence suggests a
      development area" — never convert absence into weakness.
   3. RECENCY SKEW CHECK — the timeline of my evidence; flag
      if >60% comes from the final third of the cycle.
   4. QUESTIONS — what I could still collect in the next two
      weeks to close each gap.
   

3. 3.

   Treat the gaps as your to-do list, not the report's. Good output looks like:

   Copy

   GAPS
   - "Technical leadership beyond own team": no evidence
     gathered. Note: the API-guild RFC (March) may qualify —
     you mentioned it in 1:1 notes but included no artifact.
   - "Operational excellence": evidence suggests development
     area — two incidents, both with slow initial diagnosis
     (Apr 12, Jun 3 notes).
   
   RECENCY SKEW: 7 of 11 items from the last 8 weeks. The
   Q1 migration work is underrepresented.
   

   The RFC line is the packet working: evidence that exists but wasn't captured. Spend the pre-meeting weeks retrieving artifacts and asking peers, not wordsmithing.

4. 4.

   Write the narrative yourself; bring the map to the room. The packet you present is a page in your voice: your assessment, backed by the evidence map as an appendix. In calibration, when challenged, you answer from evidence ("two design docs and the guild RFC — here's the map") instead of adjectives. That's the entire competitive advantage of the method, and it accrues to your report.

5. 5.

   For promos, add the claim-verification pass. Draft the promo packet from the evidence map, then run the discipline in reverse:

   Copy

   Here is a promo packet draft. List every factual claim it
   makes — impact numbers, scope statements, "led X" phrases —
   as a checklist with the evidence item that supports each.
   Mark any claim with no supporting item as [UNSUPPORTED].
   

   Every [UNSUPPORTED] claim gets evidence or gets cut — committees are pattern-matched to inflated packets, and one unverifiable "led" can sink four true ones. Anything the model phrased more impressively than the evidence supports, dial back yourself. The packet goes forward under your name; every sentence is a claim you're personally making.

Make it yours

• Small team or startup: no formal calibration? Run the evidence map anyway before comp decisions — the gap and recency checks are exactly the biases that creep into founder-gut compensation, and the packet is your paper trail when the company grows into needing one.
• Large org or GCC: confirm with HR which tools are approved for performance data — this is enterprise-tenant-only work, and some orgs restrict performance content to HRIS systems entirely; if so, the curation and gap-check discipline still applies with you as the mapper. Align packet format across your leads so calibration compares evidence against one rubric, not five formats.
• Managers of managers: run the packet process for your managers, and audit their packets for adjective density — a skip-level whose packets are all evidence maps has taught you something about who's actually managing.

Quality check

• One report per conversation, no exceptionscheck your history; a single session containing two reports' data is the line you committed not to cross.
• Every mapped item traces to a real artifactspot-check three: the model occasionally "improves" a paraphrase into a claim the source doesn't make.
• Gaps triggered collection, not prosethe measure of the packet is what you gathered in the two weeks after running it.
• The narrative is yoursread it aloud; calibration rooms detect model prose and discount it, which taxes your report, not you.
• Promo claims all cleared verificationzero [UNSUPPORTED] items ship; your credibility with the committee is a multi-year asset spent one packet at a time.