Lesson 11 · Intermediate · 15 min
Comparing two answers with a grid
Compare two Copilot Chat answers with a four-criteria grid: accuracy, relevance, format, sources. Method, example, exercise and quiz.
- Goal
- You will be able to obtain two answers from Copilot Chat, score them on four criteria starting with accuracy, and keep the better one for good reasons rather than for its fluency.
- Skills
- Check
Two answers are better than one when you know how to decide between them. The grid holds four criteria: accuracy, relevance, format, sources. You score each answer from 0 to 2 per criterion, starting with accuracy, then you keep the better one or merge the two. The most fluent answer is not the most accurate: the grid protects you against that reflex.
1Four criteria to tell two answers apart
Copilot Chat produces a different text every time, even with the same prompt. This variability is a resource. By comparing two outputs, you see what is stable, probably drawn from the source, and what varies, often filled in by the model. Work on self-consistency, a method where the model compares several of its own answers before deciding, shows a clear gain in accuracy; you apply the same idea by hand.
Three ways to get two answers. In a single prompt: "propose two numbered variants, then state what sets them apart". In two separate conversations: the same prompt run twice to observe stability.
You can also vary response modes: Copilot Chat offers Quick response and Think deeper. Quick response prioritises speed; Think deeper takes time to plan and check. Regeneration controls described in 2026 complement these options depending on your interface.
The grid is read in order of priority. Accuracy immediately eliminates any incorrect response: every fact, figure, or name must appear explicitly in the source or behind a verified link.
The three following criteria then rank the accurate options: - Relevance: does the answer address the question asked for the named recipient without digression? - Format: are requested structure, length constraints, and tone strictly respected? - Sources: are source passages quoted word-for-word and gaps clearly declared?
Criterion Question to ask 2 points if Accuracy Is every fact in the source? No unverifiable statement Relevance Does it answer the question, for this reader? Nothing off topic Format Structure, length, tone respected? All three instructions met Sources Passages cited, gaps declared? Every point is traceable Score from 0 to 2 per criterion, with one sentence of justification. The total is only a guide: an answer at 7 out of 8 with a wrong figure loses against an accurate answer at 5 out of 8. You can ask Copilot to fill in the grid on its two variants; this is useful for format, but accuracy remains for you to check, against the source.
Diagram "Compare two answers with a grid"Diagram generated by AI and reviewed 2Two committee summaries, only one is accurate
A project manager wants a summary of minutes for his steering committee.
Turn 1 (Generation and scoring of two competing variants).
Summarise the minutes below in five bullets for the steering committee. Propose two numbered variants. In each, one bullet = one decision or one action with owner and due date when they are given; if a topic was not settled, write "not settled". 80 words per variant. End with one sentence on what sets the two variants apart. Minutes: [pasted text]Observed response: Variant 1 writes "work package 3: not settled, sponsor's decision awaited". Variant 2 adds a deadline date absent from the source document. The evaluation rubric awards 2 for accuracy to the first and 0 to the second.
Turn 2 (Selective merging follow-up and operational calibration).
Retain the factual accuracy of variant 1 (work package 3 not settled) while adopting the direct action style of variant 2. Produce a final summary in three concrete bullets, 60 words maximum.Observed Copilot response:
- Lots 1 and 2: technical acceptance completed with zero blocking defects.
- Lot 3: budget arbitration pending, final sponsor decision expected.
- Next step: distribution of revised financial appendix Tuesday at 2 pm.
What changes. Two variants make hallucinations apparent by contrast. The "not settled" fallback gives the model an honest exit when uncertain. Turn 2 merges factual correctness with sharp executive formatting without reintroducing errors.
3Score two versions of an announcement email
Here is the fictional internal briefing note from "NovaTech SAS", ready to paste into your prompt context:
[Internal Note - Operations Reorganisation NovaTech SAS] - Context: Merging the Logistics team (12 people) and Customer Support team (8 people) into a unified "Customer Experience & Operations" department. - Employment: No redundancies, no office relocations, full retention of pay and benefits. - Timeline: Effective from 1 November 2026. Sophie Martin appointed head of the new department. - Discussion: Open Q&A session for all staff on Tuesday 22 September 2026 at 10:00 am in the main hall and via video conference.The "Load exercise" button in the composer pre-fills the objective with the weak starting prompt:
"Draft this internal announcement email in two distinct versions."
From this starting point, complete the prompt using the composer. Specify the audience for all company employees and set a 120-word limit per variant. Impose a four-part structure: announcement, rationale, job guarantees, and Q&A session. Add the accuracy constraint: "Make no promises absent from the provided facts." Enable the two-variant booster, score each version across the four criteria, and retain the best.
Self-assessment grid: (a) all four criteria are scored for each variant; (b) accuracy has been checked against the note's facts; (c) the choice is justified in one sentence.
4Twelve confident lines against one accurate line
An assistant asks for two variants of a reply regarding a contractual withdrawal period. Variant A runs twelve lines, cites three articles, and concludes with assurance that fourteen days apply. Variant B fits in three lines: "The attached terms provide ten days (article 7). The statutory fourteen-day period only applies to consumers; customer status is not specified."
The assistant intuitively picks A, assuming longer means better. This is a critical mistake: A invents clauses absent from the contract, whereas B surfaces a genuine legal ambiguity. Length and eloquence are never proxies for accuracy.
Correction: score accuracy first by opening the attachment for each cited clause before reading further. Any variant with an unverified claim scores zero and drops out of contention. Rule to remember: on the rubric, one accurate line beats twelve confident lines.
5Quiz
Three questions, instant feedback. Each option comes with an explanation.
6Proof of mastery
Paste the prompt, the two answers obtained, your completed grid (four criteria scored from 0 to 2 with one sentence of justification) and the version you kept.
Criteria
Going further
The grid prepares the hallucination hunt, which trains the eye on planted errors. The lesson on iteration shows how to correct the variant you kept rather than redoing it. The term hallucination is defined in the glossary.
Frequently asked questions
How do I get two answers?
Three ways: ask for two numbered variants in the same prompt; run the same prompt in two conversations; switch response mode between Quick response and Think deeper. The regenerate buttons described in 2026 for Copilot Chat are probably available without a licence, but Microsoft does not say so.
Do I always need all four criteria?
Accuracy is mandatory and is checked first. The other three are weighted according to the task: relevance matters more for a summary, format for a table, sources for a search.
Can I ask Copilot to score its own answers?
Yes, for format and relevance, where it is good at spotting an overrun in length or an off-topic passage. For accuracy, it cannot judge without a source; you check yourself against the document or the links.