Sample report. The session behind it is invented, and it is not attributed to any of the candidates on the shortlist. The layout, the compliance block and the structured observations are what WorkProbe renders from a real session; the narrative was written for this demo.
← Back to the dashboard

What the full report looks like

Each row on the shortlist opens a short report of the facts about how the candidate worked. This is the long form a reviewer reads before an interview.

WorkProbe's reading

SkillLevelAEDT
Data filteringstrong positivestrong
Prompt qualitypositiveacceptable
Uncertainty handlingneutralacceptable
Noise handlingpositiveacceptable

What the candidate did

Gave the manager the right answer
The last conclusion they gave the manager matches the scenario's answer. Partly: the right direction without the deciding part.
partly
  • 48:00 — “Attributed the increase primarily to the rate component, with one lane left unexplained.”
Questioned the AI's mistake before using it
Before using the AI's planted wrong statement, they questioned or checked it: asked the AI again, compared a file or asked a colleague.
yes
  • 22:00 — “Checked the quoted surcharge against the rate card before using it.”
Kept the AI's mistake out of the answer
Their answer to the manager neither states nor relies on the AI's planted wrong statement.
yes
Checked the source file after the AI's mistake
After the AI gave the wrong statement, they opened the file it was about.
yes
  • 22:12 — “Reopened the rate card and located the quarterly band.”
Asked every colleague who held a key fact
Some colleagues know a fact the answer depends on. They asked each of them about that topic. Partly: they asked only some.
no
Asked two or more colleagues about the task
They sent a real question about the task to at least two colleagues. The manager and greetings do not count.
no
Shaped their own questions to the AI
They asked focused questions of their own, instead of pasting whole files or the manager's text and passing the answer on unchanged. Partly: a mix of both.
yes
Said what they did not know
They told the manager or the AI about a gap, an assumption or a figure they could not confirm.
yes
  • 47:00 — “The lane I could not close — I would have wanted to ask someone who was there that week.”
Answered the manager's direct question promptly
When the manager asked directly for an answer or recommendation, they gave one in their next message or the one after.
yes
Noticed when the AI only agreed with them
Sometimes the AI starts agreeing with whatever the candidate believes. They questioned that agreement. When pushed: only after the manager pushed back.
did not come upthe AI never started only agreeing with them
Stayed on task
Everything they wrote was about the problem. After a detour: they went off topic and came back.
yes
Time of first explanation
When they first offered an explanation of the problem.
11:00 of 60:00
Time used
Time until the final reflection, or until the end of the session.
48:00 of 60:00

Executive Summary

The candidate worked the cost question in three passes: a read of the material before touching the assistant, a modelling pass with it, and a verification pass against the source documents. First hypothesis was formed at 00:11 and revised once, at 00:34, after a figure failed a check against the rate card.

The pattern through the session is source-first. The assistant was used to structure and to compute, and its outputs were treated as claims to be checked rather than as answers — most visibly when a fuel-surcharge figure it supplied was traced back to the rate card before it entered the working. Where the brief was ambiguous, the ambiguity was carried forward openly rather than resolved silently.

One competency the scenario was designed to probe had no observable evidence: no situation arose in which the candidate had to defend a position against a stakeholder who disagreed, so nothing was seen either way.

Skills Summary

Demonstrated strengths

  • Deliberate source navigation. Opened the rate card and the invoice extract before the first prompt, and returned to both when a figure was in question. File selection narrowed as the session went on rather than widening.
  • Verification of consequential claims. The one figure that would have changed the answer was checked against its source. The check was initiated by the candidate, not prompted by a colleague.
  • Decision ownership. The conclusion sent to the operations manager was the candidate's own, with the assistant's contribution visible in the working rather than substituting for it.

Areas for development

  • Eliciting from people. The warehouse lead held a fact the documents did not contain and was never asked. The gap was noticed — it appears in the final message as an open question — but it was reported rather than closed.
  • Stating assumptions at the point of use. An ambiguity about which quarter applied was resolved internally and surfaced only at the end.

Chronological Analysis

00:00–00:06 — Orientation. Opened the invoice extract, then the rate card, then the Slack thread, in that order. No AI query in this window.

00:07 — First prompt. Asked the assistant to lay out a cost bridge from the invoice extract, supplying the columns rather than asking what they meant. Evaluator note: scoped request, structure not conclusion.

00:11 — First hypothesis. Stated to the assistant that the rate component looked like the driver. Weak-to-neutral signal at this point: plausible, not yet checked.

00:19 — Assistant supplies a surcharge figure. The response quoted a fuel surcharge that the candidate did not carry into the working.

00:22 — Return to source. Reopened the rate card and located the quarterly band. Strong signal: the check was unprompted, and it was applied to the figure that mattered rather than to an easier one.

00:34 — Revision. Re-ran the bridge on the corrected figure. The revision is visible in the working sent later.

00:41 — Boss pressure. The boss asked for a number. The candidate replied with the figure and the one part still unresolved, rather than dropping the caveat under time pressure.

00:48 — Final message to the operations manager. Conclusion, method, and the open question, in that order.

Skill-by-Skill Observation

Data filtering

  • Evidence trail: file opens at 00:02, 00:04, 00:22; narrowing selection across three queries.
  • Observation progression: evidence accumulated steadily from the first pass; no reversal.
  • Boss interventions: one, at 00:41; the response did not change the candidate's reading of the material.
  • Final behavioural pattern: consistently source-first, with the source consulted before rather than after the answer took shape.

Prompt quality

  • Evidence trail: three prompts, each narrower than the last; none asked for a conclusion.
  • Observation progression: consistent from the first prompt.
  • Final behavioural pattern: the assistant was given work to do, not a decision to make.

Uncertainty handling

  • Evidence trail: one ambiguity carried from 00:14 to the final message; one open question reported rather than closed.
  • Final behavioural pattern: uncertainty was reported honestly, and late. The candidate did not manufacture a number to fill it, and did not ask the person who could have removed it.

Outcome: conclusion vs. ground truth

The candidate's final message attributed the increase primarily to the rate component, with the volume and mix contributions quantified and one lane left unexplained. The scenario's expected outcome names the rate component as the driver.

Where it landed: reached, with one component unresolved. The decisive fact behind that component was held by the warehouse lead and was never elicited, so it is not established by the candidate-visible files alone — those files are insufficient by design.

This section describes where the answer landed. It does not change the observations above, which describe how the candidate worked.

Noise Handling & Hallucination Detection

The Slack thread carried a plausible but unsupported explanation. The candidate read it, did not cite it, and did not treat it as corroboration for the rate hypothesis. The assistant's hedged phrasing around the surcharge did not move the conclusion; the source did.

Hypothesis Timeline

First hypothesis at 00:11, held for 23 minutes, revised once at 00:34 on a source check rather than on a challenge. No further revisions. The revision narrowed the claim rather than replacing it.

Delegation Quality Analysis

Three prompts in fifty minutes. Volume is not the signal here and is reported only for completeness: the assistant was asked for structure and arithmetic, each request narrowed the file selection, and no prompt forwarded the task itself.

Where the candidate judged the assistant unnecessary — the rate-card lookup — they did the work directly. Abstention of that kind is observed competence, not disengagement.

Focus & Off-task Behaviour

No off-task activity observed.

Unverified-answer Delivery

None. The one figure that could have gone out unverified was checked at 00:22, before the answer was composed.

In the candidate's own words

"I was fairly confident on the rate piece. The lane I could not close — I would have wanted to ask someone who was there that week."

Reviewer context only. Self-report is not an assessment input and did not move any observation above.

How each skill was assessed

Data filtering strong positive

What was assessed: Which files they worked from: whether they used the files that hold the answer, and whether they treated a file with nothing useful in it as proof.

Consistently source-first: the invoice extract and the rate card were opened before the first prompt, and the rate card again at 00:22 when a figure was in question. File selection narrowed across the session rather than widening.

Back to the table

Prompt quality positive

What was assessed: How they asked the AI assistant: whether each question was their own, whether they chose what to give it, and whether they built on its answers and steered it back when it went wrong.

Three prompts, each narrower than the last, asking for structure and arithmetic. None asked the assistant for the conclusion; the columns were supplied rather than asked about.

Back to the table

Uncertainty handling neutral

What was assessed: How they handled what they did not know: whether what they stated firmly was backed by what they had, whether they marked guesses as guesses, and whether they said what would confirm them.

One ambiguity about which quarter applied was carried from 00:14 to the final message, and one open question was reported rather than closed. Reported honestly and late: no number was invented to fill the gap, and the warehouse lead, who could have removed it, was not asked.

Back to the table

Noise handling positive

What was assessed: What their final conclusion rested on: whether it would still stand if the misleading material in the scenario were taken away.

The Slack thread carried a plausible but unsupported explanation. It was read and not cited, and it was not treated as support for the rate hypothesis. The assistant's hedged surcharge figure did not move the conclusion; the rate card did.

Back to the table

Questions to discuss

The report says what happened in the session. These are what is worth asking about it — open questions about the candidate’s reasoning, not a re-marking of their answer.

  1. You went to the rate card before the assistant gave you a reason to. What were you looking for?
  2. Walk me through the moment you decided the surcharge figure needed checking — what made that one stand out?
  3. You left one lane unexplained rather than estimating it. How do you decide when to report a gap instead of closing it?
  4. The warehouse lead had something the documents did not. Looking back, what would have prompted you to ask?