Domain 2 – Output Evaluation and Validation (21%) – Largest domain

• Check accuracy and completeness
• Detect hallucinations, inconsistencies, and bias
• Fact-checking techniques
• Decide when human review or additional verification is required
• Edit, refine, and choose appropriate output formats (artifacts, structured data, etc.)

CCAO-F Domain 2 · Task 2.1

What You Need to Know

Treat every Claude output as a strong first draft from a brilliant colleague who can be confidently wrong. Fluency and confidence are properties of the writing, not evidence about the facts.

Two Review Questions

1. Accuracy — are the specifics right? (numbers, names, dates, titles, quoted figures). Order by consequence.
2. Completeness — is anything missing? Walk the source’s own structure; confirm each element is in the output or knowingly excluded.

Verify Against the Input

When an output is built from attached material, the first check is the attachment. Ask Claude where in the source each figure or claim came from, then spot-check. Asking Claude whether it is sure is no check — the same model that made an error will repeat it with the same confidence.

Key Concept: Ownership settles who does the checking. The professional who sends Claude’s work owns it.

Exam Traps

• Sending the output because it reads well or because Claude expressed high confidence
• Asking Claude to confirm its own work and treating a confident yes as review
• Verifying one named figure when part of the source may have been dropped

Practice Scenario

Scout group treasurer has Claude draft the annual financial summary for the AGM from the accounts spreadsheet. Draft reads perfectly; she must be sure it is right and leaves nothing out, without taking Claude’s word. What should she do before circulating it?

Build Exercise: Run the Two-Pass Review (25 min)

CCAO-F Domain 2 · Task 2.2

What You Need to Know

Three failure types, each with a different response:

Failure	Shape	Response
Hallucination	Fluent, specific, fabricated content not grounded in fact	Verify against an authoritative source
Inconsistency	Output disagrees with itself (prose vs table, recommendation vs findings)	Re-verify the whole output against its source
Bias	Skewed framing, assumptions or language (pattern across descriptions of people)	Rework the pattern, not a single word

Hallucination Risk Concentrates Where Grounding Is Thinnest

• Private internal systems/documents Claude was never given
• Very recent events
• Precise figures not in the source material
• Highest-risk shape: confident numbered instructions for an internal tool Claude has never seen

Claude can also hallucinate its own capabilities (claiming to have sent an email or produced an external file).

Bias

Most consequential in outputs about people. Read across descriptions rather than sentence by sentence. Apply the read-aloud test.

Exam Traps

• Trusting confident, step-by-step instructions for a private internal system
• Believing Claude’s claim to have sent an email or produced an external file
• Fixing the one contradiction you spotted and sending the rest
• Treating bias as one bad word to delete

Practice Scenario

Skin clinic practice manager asks Claude how to produce the monthly patient-recall report in the clinic’s management software. Claude has had no access to this software, yet replies with confident numbered steps naming specific menus. What should the manager do with these steps?

Build Exercise: Run a Three-Failure Sweep (30 min)

CCAO-F Domain 2 · Task 2.3

What You Need to Know

Validation effort is ordered cheapest to heaviest, with a judgement layer for proportionality.

Cheapest First: The Source You Gave It

For output built from attached material, the first check is the attachment. Reach for anything heavier only when the given source cannot settle it.

Independent, Authoritative, Primary

External claims need external checking. The standard is the primary document (official publication, regulator’s text, funder’s published criteria, organisation’s records). A second AI summary, a forum post, or Claude’s own earlier output is not independent validation.

Make the Output Traceable

Ask Claude to quote relevant parts of the source / show where each claim came from. Spot-check the weighty ones. When a single figure drives the decision, triangulate it in two independent places.

Proportionality

Validation effort follows consequence. Heavy checking on low-stakes work is also wrong. Self-check is a cheap first pass for arithmetic slips and contradictions; it is not independent validation.

Key Concept: When a validation item offers several checking plans, test each against: (1) Does independent information enter the check? (2) Does the depth match what rides on each claim?

Exam Traps

• Asking Claude to confirm or re-check its claims and calling that verification
• Checking a claim against another AI summary or Claude’s own earlier output
• Verifying every sentence to the same maximal depth regardless of stakes

Practice Scenario

Heritage railway volunteer coordinator has Claude draft a grant application stating annual visitor numbers, a percentage increase in school visits, and that the railway meets the funder’s published eligibility criteria. One evening only; nothing may rest on Claude’s word alone. Right validation plan?

Build Exercise: Build a Validation Plan for One Real Document (35 min)

CCAO-F Domain 2 · Task 2.4

What You Need to Know

Escalation passes a Claude output to a qualified person to review before use. Triage decides who reviews, never whether anyone does.

The Triggers

• External audience (customers, public, press, funders)
• Legal, financial or contractual consequence
• Decisions about people (hiring, appraisal, allocation)
• Safety or health implications
• Irreversibility once sent

One trigger is enough. Confidence and polish are not on the list.

Qualified, Not Senior

Anthropic Usage Policy High-risk AI Recommendations require a qualified person (training or experience in that field) with authority to change the output, before it is delivered as advice or used to make a decision. The person concerned must also be told AI was used. Seniority and availability are neither necessary nor sufficient.

Both Failure Directions

• Under-escalation — sending externally-bound or consequential content because the draft reads authoritatively
• Over-escalation — committee for a rota note; every routine draft through a professional reviewer. Buries reviewers and erodes the channel high-stakes outputs need

Key Concept: Triage asks two questions in order. What does this output carry? If anything, who is qualified in that subject to catch the errors that matter?

Exam Traps

• Choosing the reviewer by seniority or availability
• Waving output through because it is only internal
• Escalating everything to be safe

Practice Scenario

Opticians’ practice manager has three Claude drafts: advertising copy claiming specific visual outcomes; reworded internal rota-swap policy; reply to a patient complaint that touches on what may have gone wrong clinically. Manager has no clinical training. Right triage?

Build Exercise: Triage Your Last Five Outputs (25 min)

CCAO-F Domain 2 · Task 2.5

What You Need to Know

Claude drafts; you edit. Earlier objectives asked whether an output is right; this one asks whether it is right for its reader.

You Are the Editor

Two remedies: edit the output yourself, or direct a revision that names what must change. A directed revision gives follow-up instructions, not a bare “try again”.

Right for Its Reader

A technically correct output can be wrong for its reader (jargon for a lay audience, detail-heavy analysis for a time-poor executive, internal register in a customer-facing note). Adaptation changes vocabulary, sentence density, and order so the urgent item leads. Substance must survive: findings, recommendations and costs. After adapting, check the new version against the original for anything lost or softened.

Criteria Before Reading

When choosing between versions, fix the criteria before reading (accuracy preserved, audience fit, tone, completeness, length). Style is the one contest fluent output always wins — a fluent but thinner version will beat a correct one if criteria are not fixed first.

Generator, Not Judge

Claude produces variants; you apply the criteria and pick. Asking Claude which of its versions is better hands over the judgement this objective certifies.

Exam Traps

• Sending an accurate but reader-hostile output because the content is correct
• Adapting for a lay audience by cutting the difficult content
• Asking Claude to pick its own best version and sending its choice
• Choosing between two versions by which reads better, with no criteria fixed beforehand

Practice Scenario

Building surveyor’s assistant has a survey summary that is accurate but dense with structural terminology; the urgent roof defect sits mid-document. It must go to first-time buyers with no construction background. Surveyor insists buyers can follow every finding and that the assistant, not Claude, settles the final wording. Right way to prepare it?

Build Exercise: Adapt and Compare for a Real Audience (35 min)

CCAO-F Domain 2 · Task 2.6

What You Need to Know

A long, productive conversation is not a deliverable. Curation extracts what the deliverable needs. Format selection gives it the container its reader will use.

Three Format Homes

Format	Fits
Inline chat response	Quick, one-off answers consumed in the moment
Structured table	Anything the reader compares or scans (options against criteria, schedules)
Artifact	Significant, self-contained content you are likely to edit, iterate on or reuse outside the conversation

Choice follows the reader’s use: ask-and-move-on → inline; compare-and-scan → table; reuse-and-revise → artifact.

From Working Material to Work Product

Curation starts from the deliverable, not the conversation: who reads it, what they do with it, whether it will be reused. Most of a working conversation is scaffolding; the deliverable owes its reader the decisions alone. Forwarding the whole conversation delegates curation to the reader — who will not do it.

What Each Mismatch Costs

• Reusable runbook left as loose chat prose → gets lost and re-created
• Comparison delivered as paragraphs → every reader rebuilds the table in their head
• Everything in a table → flattens narrative judgement
• Everything in an artifact → adds ceremony to throwaway answers

Key Concept: For every output, ask: what does the reader do with this? Read it once, compare across it, or come back to it? The answer is the format.

Exam Traps

• Answering a format question with Projects, Research mode or a model choice (Domain 3)
• Defaulting every output to the same container
• Sharing the whole conversation instead of curating the deliverable

Practice Scenario

Swimming club gala organiser has weeks of Claude conversations. Three outputs needed: agreed start time for tonight’s committee question; comparison of three candidate dates against pool availability, volunteers and fixture clashes; gala-day runbook that will be revised and reused every year. Which formats fit?

Build Exercise: Curate One Deliverable Out of a Working Chat (30 min)