top of page

Use a Decision Rubric Without Hiding Its Uncertainty

Writer: Branden Bell
Branden Bell
4 days ago
5 min read
Show what’s missing. A cream score slip has two olive rules, an empty field and a terracotta tab.

One proposal scores eight. Another has seven verified points and a missing sample. If that sample earns three, the second proposal can reach ten. Would you select the first one today, or wait for evidence that could change the result?


If you use AI to review creator proposals, keep unknown evidence visible. Calculate which results are still possible, see whether they could change the order, and keep mandatory checks separate. A score helps a person decide where to look next. It doesn’t approve a proposal.


Start with a rubric the team already uses


In a public post on X, Brian D. Evans says he built a pitch-deck analyzer as one of two shareable bots he made during early access. The public preview shows his description, not an independently measured result. The independent GrokBot.dev case describes a screening workflow based on the reviewer’s own criteria. Its fuller prompt is the curator’s reconstruction. The pitch-deck review guide covers tracing a claim back to a slide. Here the question comes later: what should the reviewer do when the missing evidence could change the decision?


Try it with one comparison that comes up often, such as selecting a workshop for a creator event. Define the rubric before viewing submissions and say which requirements can’t be traded for points. Use synthetic submissions or materials the team is permitted to process. The example below is fictional.


Eight points versus a range


A small creator conference has one workshop slot. The organizer first requires an identified caption provider and confirmed coverage for the full workshop time. That’s an accessibility check, separate from the score.


Its review uses three criteria: audience fit (up to four), delivery preparation (up to three) and evidence of a previous exercise (up to three). The evidence criterion is the one at issue. Assign zero if the reviewer confirms there is no prior exercise, one for a sample that doesn’t show what attendees do, two for an agenda and handout that explain the exercise, and three for an approved recording excerpt that demonstrates it alongside the handout. If the file is missing or can’t be read, mark it unknown until the reviewer can tell what it supports.


Proposal A: eight points. The organizer verifies audience fit (4), a clear agenda and handout (2), and a run plan that names the facilitator but doesn’t give a backup (2). A names an approved live caption provider, and the organizer confirms coverage for the full time with the event team. A is eligible for human selection.


Proposal B: seven known points, plus an unknown. The organizer verifies audience fit (4) and a full run plan with a backup and the handoff steps (3). The previous-exercise sample link is missing. B’s possible total is seven to ten under the rubric, depending on what the reviewer can establish. That’s a range of possible scores, without any claim about how likely they are. B’s caption plan is also unconfirmed, so the organizer holds it regardless of the possible score.


Here’s what to ask next. If the caption plan can be confirmed, one point for the sample ties A, and two or three put B ahead. If the caption plan can’t be confirmed, a great sample doesn’t make B eligible. Ask the owner for those two pieces of evidence and keep the candidate on hold while they’re missing.


Show what would change the decision


  • Current decision: A is eligible at 8; B stays on hold. No invitation or booking has been approved.

  • What could change the order: If B clears the caption requirement and its sample earns at least 2, its score exceeds A’s.

  • Source and owner: The event coordinator checks caption availability and retrieves a rights-cleared sample. Record the source, date, criterion and reviewer.

  • Decision deadline: If the two checks are unresolved at the deadline, the event lead may select A. Record the deadline and the unresolved facts. Preserve the original score and why the lead acted.

  • Approval: The event lead approves the exact invitation, recipient and any spend separately. Keep the AI output in the internal review queue.


This card works in plain text and makes sense even if read one field at a time. If the same criteria keep failing to separate candidates, review the rubric for the next round. Preserve the version used for this decision. If someone approves an exception, record who approved what, why and which source they relied on. Don’t waive a hard requirement by writing a positive summary.


Keep the submissions in the right place


Start without inbox, account or outbound access. Keep private contacts and unpublished financials out of the pilot. Record which sample was approved for use, and describe the relevant evidence rather than copying the submission into a shared note. If a draft score or recipient is wrong, hold the invitation, correct the internal record and preserve the reason.


The provider’s approvals and privacy guidance and acceptable use policy are useful checks before connecting a tool. Read the X automation rules if the workflow ever interacts with X. The GrokBot.dev content license covers its repository’s content folder; a deck or recording needs separate permission. Any public release, sending, account change or payment stays under human approval.


Questions about scores with missing evidence


Should a missing fact get zero?


Only if the rubric defines a verified absence as zero and the reviewer has established that absence. An inaccessible file is still unknown. Mark it so, then show whether resolving it could change the decision.


Does the score range measure confidence?


No. It shows possible totals under the written rubric. For B, seven to ten means adding zero through three to the verified seven. It says nothing about which score is probable.


What if two proposals tie?


Give the reviewer a tie rule in advance, such as favoring a lower implementation burden after all mandatory checks pass. Record the reason for the decision. Don’t make small unsupported score changes just to avoid a tie.


Should the assistant email the shortlisted creator?


Keep its draft internal. Verify the candidate, the exact claims and the terms, then give the responsible person permission to release only that reviewed message. A shortlist score alone authorizes nothing.


Run the test before the next call for proposals


Take two synthetic proposals and remove one score-bearing fact from one. Ask the reviewer whether that fact could change the order, which check can still block selection and what they would do if the deadline arrives first. If that answer is hard to find, change the review card. If the same decision keeps landing in different inboxes, bring one anonymized review process to Bell Consulting Solutions, or start with the AI process optimization guide.

 
 
 

Comments


bottom of page