Practice data

What learners get right and wrong

Name: Cred Farmer practice-exam accuracy by exam domain
Creator: Cred Farmer
License: https://credfarmer.com/terms

This is how people score on Cred Farmer’s own practice questions under a timed clock. It is not the real exam, it is not a pass rate, and where the numbers cannot tell two things apart we say so.

1,561Timed practice papers sat
Drills and study runs are not among them
829Learners who sat one
292 of 829 left their first paper unfinished
31,899Answers in the measured set
16,084 unanswered and 2,273 repeat encounters removed
31.3%Papers left less than half answered
489 of 1,561 papers
Read on 3 October 2026. The third figure counts only answers that survive every filter below; the first two count everything, filtered or not. Each bar is drawn against the count printed above it, and none of them is an accuracy.

Read this first

What this page is not

It is not the real exam. Every figure here comes from our practice bank: questions we wrote, marked against our own answer key. The certification exams are written and delivered by other people, and nothing measured here transfers to them.
It is not a pass rate. We have verified outcomes for almost nobody, and we already published the reason we would not quote one even if we did, on our scoring methodology page. Nothing here names a score threshold, and nothing here should be divided by one.
It is not a claim about candidates. Everyone in this data chose a preparation site and then chose to sit a timed paper on it. That is a self-selected group and it is not a sample of anything wider.
It is not an ordering. The tables are sorted so they can be read, not to rank anything. Almost every gap between two domains here is smaller than the uncertainty around it, and we say which ones are not.

CCDV-F

Claude Certified Developer – Foundations, domain by domain

223 learners, 386 timed papers and 16,028 answers that survive every filter below.

Accuracy by exam domain

Pooled across every domain, 68.4% of answers were correct, 67.69% to 69.14% counting items, and 66.0% to 70.84% once the answers are grouped by the learner who gave them. Where two bands overlap, this data cannot tell those domains apart.

Every pairwise contrast

Every pairwise contrast. 28 contrasts between the 8 published rows. 16 of them separates after correction; the other 12 sit inside the noise, which means the measurement cannot tell those rows apart.

Claude Code with Security and Safety: this contrast separates after correction.
Tools and MCPs with Security and Safety: this contrast separates after correction.
Prompt and Context Engineering with Security and Safety: this contrast separates after correction.
Eval, Testing, and Debugging with Security and Safety: this contrast separates after correction.
Applications and Integration with Security and Safety: this contrast separates after correction.
Applications and Integration with Claude Code: this contrast separates after correction.
Applications and Integration with Tools and MCPs: this contrast separates after correction.
Applications and Integration with Prompt and Context Engineering: this contrast separates after correction.
Agents and Workflows with Security and Safety: this contrast separates after correction.
Agents and Workflows with Claude Code: this contrast separates after correction.
Agents and Workflows with Tools and MCPs: this contrast separates after correction.
Agents and Workflows with Prompt and Context Engineering: this contrast separates after correction.
Model Selection and Optimization with Security and Safety: this contrast separates after correction.
Model Selection and Optimization with Claude Code: this contrast separates after correction.
Model Selection and Optimization with Tools and MCPs: this contrast separates after correction.
Model Selection and Optimization with Prompt and Context Engineering: this contrast separates after correction.

One cell per pair. Filled survives correction for testing them all; open is inside the noise.

Of the 28 possible comparisons between these domains, 16 survive once the intervals are grouped by learner and corrected for testing them all: learners answer Applications and Integration about 18 points more accurately than Security and Safety; learners answer Model Selection and Optimization about 20 points more accurately than Security and Safety; learners answer Agents and Workflows about 19 points more accurately than Security and Safety; learners answer Agents and Workflows about 10 points more accurately than Tools and MCPs; learners answer Model Selection and Optimization about 10 points more accurately than Tools and MCPs; learners answer Agents and Workflows about 9 points more accurately than Prompt and Context Engineering; learners answer Applications and Integration about 8 points more accurately than Tools and MCPs; learners answer Model Selection and Optimization about 9 points more accurately than Prompt and Context Engineering; learners answer Prompt and Context Engineering about 11 points more accurately than Security and Safety; learners answer Tools and MCPs about 10 points more accurately than Security and Safety; learners answer Applications and Integration about 7 points more accurately than Prompt and Context Engineering; learners answer Eval, Testing, and Debugging about 15 points more accurately than Security and Safety; learners answer Model Selection and Optimization about 10 points more accurately than Claude Code; learners answer Agents and Workflows about 10 points more accurately than Claude Code; learners answer Applications and Integration about 8 points more accurately than Claude Code; learners answer Claude Code about 9 points more accurately than Security and Safety. Every other gap in this table is inside the noise.

The two rows furthest apart, and why both intervals are printed

The two rows furthest apart, and why both intervals are printed
Row	Accuracy	Per-item interval	Learner-clustered interval	Correct of answered	Learners
Security and Safety	53.3%	50.45% to 56.15%	49.14% to 57.49%	628 of 1,178	196
Model Selection and Optimization	73.0%	71.39% to 74.64%	70.51% to 75.59%	2,106 of 2,883	223

Both intervals agree that the gap between these two is real.

Claude Certified Developer – Foundations accuracy by exam domain, on timed practice papers taken on the current question bank.
Domain	Answered correctly	95% interval per item	95% interval clustered by learner	Correct of answered	Learners
7Security and Safety	53.3%	50.45% to 56.15%	49.14% to 57.49%	628 of 1,178	196
3Claude Code	62.8%	58.79% to 66.65%	58.29% to 67.32%	363 of 578	185
8Tools and MCPs	62.9%	60.54% to 65.14%	59.3% to 66.44%	1,065 of 1,694	199
6Prompt and Context Engineering	63.9%	61.63% to 66.17%	60.22% to 67.64%	1,099 of 1,719	209
4Eval, Testing, and Debugging	68.3%	63.63% to 72.61%	62.87% to 73.72%	280 of 410	187
2Applications and Integration	71.3%	70.04% to 72.49%	68.97% to 73.59%	3,743 of 5,251	215
1Agents and Workflows	72.7%	70.8% to 74.44%	69.88% to 75.43%	1,682 of 2,315	200
5Model Selection and Optimization	73.0%	71.39% to 74.64%	70.51% to 75.59%	2,106 of 2,883	223

CCDV-F detail by objective: 24 shown, 1 withheldAn objective appears here once at least 20 learners have answered at least 100 of its questions. The rest are not shown, rather than shown small.

Claude Certified Developer – Foundations accuracy by exam objective, on timed practice papers taken on the current question bank.
Objective	Answered correctly	95% interval per item	Correct of answered	Learners
O01Agent Architecture	78.7%	75.4% to 81.57%	534 of 679	188
O02Agent Construction with Claude	64.6%	61.15% to 67.89%	498 of 771	198
O03Agent Patterns and Frameworks	75.1%	72.15% to 77.91%	650 of 865	200
O04Understanding Requirements	45.8%	41.62% to 49.98%	249 of 544	189
O05Systems Life Cycle	57.1%	52.21% to 61.87%	229 of 401	160
O06Claude API Mechanics	82.5%	80.19% to 84.65%	926 of 1,122	214
O07Software Engineering Foundations	81.5%	79.22% to 83.55%	1,012 of 1,242	214
O08Claude Application Design	70.5%	67.92% to 72.99%	878 of 1,245	205
O09Configuration Management	64.4%	60.79% to 67.89%	449 of 697	194
O10Claude Code Operation	62.8%	58.79% to 66.65%	363 of 578	185
O11Debugging and Error Handling	68.3%	63.63% to 72.61%	280 of 410	187
O12LLM Fundamentals	79.3%	76.61% to 81.79%	748 of 943	221
O13Technical Fundamentals	75.9%	73.29% to 78.36%	829 of 1,092	213
O14Model Selection and Tradeoffs	53.3%	48.84% to 57.67%	260 of 488	206
O15Cost and Token Management	74.7%	69.98% to 78.94%	269 of 360	202
O16Context Engineering	70.2%	66.44% to 73.77%	420 of 598	202
O17Prompt Engineering	62.5%	58.68% to 66.14%	403 of 645	207
O18Output Handling	58.0%	53.5% to 62.34%	276 of 476	200
O19AI Application Security	56.9%	52.52% to 61.1%	290 of 510	184
O20Guardrails and Safe Deployment	41.8%	36.28% to 47.43%	124 of 297	185
O21Claude Hooks	52.9%	47.34% to 58.32%	166 of 314	159
O23Tool Implementation	60.3%	56.53% to 64.02%	394 of 653	196
O24MCP Server Development	64.8%	60.07% to 69.27%	267 of 412	193
O25Agentic Customization	64.2%	60.4% to 67.88%	404 of 629	198

Intervals here are per item only. These cells are far smaller than the domain cells above, and this page makes no claim that any objective differs from any other.

CCAO-F

Claude Certified Associate – Foundations, domain by domain

197 learners, 297 timed papers and 10,384 answers that survive every filter below.

Accuracy by exam domain

Pooled across every domain, 60.4% of answers were correct, 59.44% to 61.33% counting items, and 57.67% to 63.11% once the answers are grouped by the learner who gave them. Where two bands overlap, this data cannot tell those domains apart.

Every pairwise contrast

Every pairwise contrast. 21 contrasts between the 7 published rows. 9 of them separates after correction; the other 12 sit inside the noise, which means the measurement cannot tell those rows apart.

Product and Model Selection with Prompting and Task Execution: this contrast separates after correction.
Governance, Risk, and Responsible Use with Prompting and Task Execution: this contrast separates after correction.
Troubleshooting and Optimization with Prompting and Task Execution: this contrast separates after correction.
Troubleshooting and Optimization with Output Evaluation and Validation: this contrast separates after correction.
Workflow Integration and Solution Design with Prompting and Task Execution: this contrast separates after correction.
Workflow Integration and Solution Design with Output Evaluation and Validation: this contrast separates after correction.
Workflow Integration and Solution Design with Configuration and Knowledge Management: this contrast separates after correction.
Workflow Integration and Solution Design with Product and Model Selection: this contrast separates after correction.
Workflow Integration and Solution Design with Governance, Risk, and Responsible Use: this contrast separates after correction.

One cell per pair. Filled survives correction for testing them all; open is inside the noise.

Of the 21 possible comparisons between these domains, 9 survive once the intervals are grouped by learner and corrected for testing them all: learners answer Workflow Integration and Solution Design about 14 points more accurately than Prompting and Task Execution; learners answer Workflow Integration and Solution Design about 11 points more accurately than Output Evaluation and Validation; learners answer Workflow Integration and Solution Design about 9 points more accurately than Configuration and Knowledge Management; learners answer Workflow Integration and Solution Design about 8 points more accurately than Product and Model Selection; learners answer Troubleshooting and Optimization about 9 points more accurately than Prompting and Task Execution; learners answer Governance, Risk, and Responsible Use about 8 points more accurately than Prompting and Task Execution; learners answer Workflow Integration and Solution Design about 7 points more accurately than Governance, Risk, and Responsible Use; learners answer Product and Model Selection about 6 points more accurately than Prompting and Task Execution; learners answer Troubleshooting and Optimization about 6 points more accurately than Output Evaluation and Validation. Every other gap in this table is inside the noise.

The two rows furthest apart, and why both intervals are printed

The two rows furthest apart, and why both intervals are printed
Row	Accuracy	Per-item interval	Learner-clustered interval	Correct of answered	Learners
Prompting and Task Execution	54.4%	52.05% to 56.74%	51.04% to 57.77%	945 of 1,737	196
Workflow Integration and Solution Design	68.7%	66.36% to 70.99%	65.05% to 72.38%	1,061 of 1,544	127

Both intervals agree that the gap between these two is real.

Claude Certified Associate – Foundations accuracy by exam domain, on timed practice papers taken on the current question bank.
Domain	Answered correctly	95% interval per item	95% interval clustered by learner	Correct of answered	Learners
1Prompting and Task Execution	54.4%	52.05% to 56.74%	51.04% to 57.77%	945 of 1,737	196
2Output Evaluation and Validation	57.5%	55.44% to 59.43%	54.18% to 60.72%	1,361 of 2,369	161
5Configuration and Knowledge Management	59.4%	56.35% to 62.29%	54.94% to 63.77%	625 of 1,053	125
3Product and Model Selection	60.5%	57.77% to 63.23%	57.41% to 63.66%	747 of 1,234	130
6Governance, Risk, and Responsible Use	62.0%	59.49% to 64.52%	58.4% to 65.68%	889 of 1,433	119
7Troubleshooting and Optimization	63.4%	60.4% to 66.33%	59.6% to 67.23%	643 of 1,014	117
4Workflow Integration and Solution Design	68.7%	66.36% to 70.99%	65.05% to 72.38%	1,061 of 1,544	127

CCAO-F detail by objective: 28 shown, 0 withheldAn objective appears here once at least 20 learners have answered at least 100 of its questions. The rest are not shown, rather than shown small.

Claude Certified Associate – Foundations accuracy by exam objective, on timed practice papers taken on the current question bank.
Objective	Answered correctly	95% interval per item	Correct of answered	Learners
O01Create effective prompts for business and technical tasks	45.9%	41.2% to 50.64%	195 of 425	189
O02Apply task decomposition techniques to structure complex requests	60.0%	55.62% to 64.23%	297 of 495	176
O03Iterate prompts to improve output quality	52.3%	47.35% to 57.18%	206 of 394	177
O04Adapt prompting strategies based on task type (e.g., analysis, research, drafting, brainstorming)	58.4%	53.64% to 63.0%	247 of 423	157
O05Evaluate Claude-generated outputs for accuracy and completeness	50.6%	45.66% to 55.63%	194 of 383	158
O06Identify hallucinations, inconsistencies, and biases in responses	64.0%	59.0% to 68.62%	243 of 380	153
O07Apply fact-checking and validation techniques	61.2%	56.13% to 66.01%	227 of 371	149
O08Determine when human review or additional verification is required	66.0%	61.1% to 70.53%	254 of 385	143
O09Edit, adapt, refine, and compare the output generated by Claude model(s) for the intended audience	60.5%	54.72% to 65.99%	173 of 286	137
O10Organize and curate information for optimal outputs and select appropriate output formats (artifacts for code/documents, inline for conversational responses, structured formats for data)	47.9%	43.77% to 52.0%	270 of 564	133
O11Select appropriate Claude product features (e.g., projects, research mode, chat, artifacts)	53.5%	47.39% to 59.44%	139 of 260	122
O12Differentiate between Claude model types (e.g., Haiku, Sonnet, Opus)	46.0%	40.13% to 51.9%	125 of 272	126
O13Align model selection with task requirements (e.g., cost, speed, quality)	71.2%	66.29% to 75.57%	259 of 364	126
O14Understand, manage context limitations and memory considerations and when to restart, summarize, or persist	66.3%	61.07% to 71.11%	224 of 338	127
O15Apply Claude to analyze requirements and use cases	69.0%	63.8% to 73.76%	227 of 329	124
O16Leverage Claude for research, planning, and process optimization	75.8%	70.21% to 80.58%	197 of 260	121
O17Integrate Claude into existing workflows to augment or redesign workflows	63.5%	60.06% to 66.78%	499 of 786	123
O18Communicate Claude’s value and limitations to stakeholders	81.7%	75.14% to 86.77%	138 of 169	119
O19Configure Claude Projects with instructions and knowledge sources	55.7%	50.69% to 60.54%	216 of 388	123
O20Manage uploaded knowledge and connectors (e.g., Google Drive, Gmail)	61.9%	56.99% to 66.52%	245 of 396	120
O21Create effective system-level instructions	61.0%	55.02% to 66.61%	164 of 269	115
O22Identify appropriate and inappropriate use cases	71.0%	65.68% to 75.74%	220 of 310	116
O23Apply data, privacy sensitivity, regulatory policies, and privacy considerations	63.0%	58.95% to 66.78%	367 of 583	115
O24Follow organizational AI policies and governance standards	64.7%	59.19% to 69.85%	198 of 306	114
O25Understand ethical implications of AI usage	44.4%	38.21% to 50.85%	104 of 234	108
O26Identify, diagnose, and resolve issues with underperforming prompts and/or poor outputs	57.9%	52.32% to 63.24%	180 of 311	105
O27Adjust approach based on feedback and results	66.4%	62.04% to 70.48%	318 of 479	115
O28Optimize workflows for efficiency and effectiveness	64.7%	58.27% to 70.7%	145 of 224	109

Intervals here are per item only. These cells are far smaller than the domain cells above, and this page makes no claim that any objective differs from any other.

CCAR-F

Claude Certified Architect – Foundations, domain by domain

75 learners, 130 timed papers and 5,487 answers that survive every filter below.

Accuracy by exam domain

Pooled across every domain, 78.3% of answers were correct, 77.22% to 79.41% counting items, and 74.84% to 81.82% once the answers are grouped by the learner who gave them. Where two bands overlap, this data cannot tell those domains apart.

Every pairwise contrast

Every pairwise contrast. 10 contrasts between the 5 published rows. 6 of them separates after correction; the other 4 sit inside the noise, which means the measurement cannot tell those rows apart.

Agentic Architecture & Orchestration with Prompt Engineering & Structured Output: this contrast separates after correction.
Agentic Architecture & Orchestration with Tool Design & MCP Integration: this contrast separates after correction.
Agentic Architecture & Orchestration with Claude Code Configuration & Workflows: this contrast separates after correction.
Context Management & Reliability with Prompt Engineering & Structured Output: this contrast separates after correction.
Context Management & Reliability with Tool Design & MCP Integration: this contrast separates after correction.
Context Management & Reliability with Claude Code Configuration & Workflows: this contrast separates after correction.

One cell per pair. Filled survives correction for testing them all; open is inside the noise.

Of the 10 possible comparisons between these domains, 6 survive once the intervals are grouped by learner and corrected for testing them all: learners answer Agentic Architecture & Orchestration about 15 points more accurately than Tool Design & MCP Integration; learners answer Context Management & Reliability about 17 points more accurately than Tool Design & MCP Integration; learners answer Context Management & Reliability about 17 points more accurately than Prompt Engineering & Structured Output; learners answer Agentic Architecture & Orchestration about 15 points more accurately than Prompt Engineering & Structured Output; learners answer Context Management & Reliability about 14 points more accurately than Claude Code Configuration & Workflows; learners answer Agentic Architecture & Orchestration about 11 points more accurately than Claude Code Configuration & Workflows. Every other gap in this table is inside the noise.

The two rows furthest apart, and why both intervals are printed

The two rows furthest apart, and why both intervals are printed
Row	Accuracy	Per-item interval	Learner-clustered interval	Correct of answered	Learners
Prompt Engineering & Structured Output	70.6%	67.71% to 73.26%	65.22% to 75.9%	731 of 1,036	56
Context Management & Reliability	87.9%	85.58% to 89.96%	84.77% to 91.1%	751 of 854	69

Both intervals agree that the gap between these two is real.

Claude Certified Architect – Foundations accuracy by exam domain, on timed practice papers taken on the current question bank.
Domain	Answered correctly	95% interval per item	95% interval clustered by learner	Correct of answered	Learners
4Prompt Engineering & Structured Output	70.6%	67.71% to 73.26%	65.22% to 75.9%	731 of 1,036	56
2Tool Design & MCP Integration	71.2%	68.25% to 73.9%	66.58% to 75.72%	703 of 988	57
3Claude Code Configuration & Workflows	74.4%	71.78% to 76.92%	69.82% to 79.05%	824 of 1,107	74
1Agentic Architecture & Orchestration	85.8%	83.96% to 87.5%	82.56% to 89.08%	1,289 of 1,502	67
5Context Management & Reliability	87.9%	85.58% to 89.96%	84.77% to 91.1%	751 of 854	69

CCAR-F detail by objective: 29 shown, 1 withheldAn objective appears here once at least 20 learners have answered at least 100 of its questions. The rest are not shown, rather than shown small.

Claude Certified Architect – Foundations accuracy by exam objective, on timed practice papers taken on the current question bank.
Objective	Answered correctly	95% interval per item	Correct of answered	Learners
O01Design and implement agentic loops for autonomous task execution	75.0%	69.94% to 79.46%	237 of 316	66
O02Orchestrate multi-agent systems with coordinator-subagent patterns	85.0%	79.21% to 89.44%	159 of 187	54
O03Configure subagent invocation, context passing, and spawning	92.5%	88.14% to 95.33%	197 of 213	63
O04Implement multi-step workflows with enforcement and handoff patterns	93.5%	89.41% to 96.11%	202 of 216	56
O05Apply Agent SDK hooks for tool call interception and data normalization	88.8%	83.37% to 92.65%	159 of 179	59
O06Design task decomposition strategies for complex workflows	79.1%	73.33% to 83.92%	178 of 225	54
O07Manage session state, resumption, and forking	94.6%	90.01% to 97.13%	157 of 166	52
O08Design effective tool interfaces with clear descriptions and boundaries	65.5%	58.45% to 71.83%	125 of 191	54
O09Implement structured error responses for MCP tools	73.7%	67.19% to 79.38%	146 of 198	54
O10Distribute tools appropriately across agents and configure tool choice	67.4%	60.36% to 73.7%	126 of 187	54
O11Integrate MCP servers into Claude Code and agent workflows	66.0%	59.05% to 72.28%	128 of 194	53
O12Select and apply built-in tools (Read, Write, Edit, Bash, Grep, Glob) effectively	81.7%	75.98% to 86.23%	178 of 218	54
O13Configure CLAUDE.md files with appropriate hierarchy, scoping, and modular organization	70.7%	65.83% to 75.17%	256 of 362	70
O14Create and configure custom slash commands and skills	64.4%	57.54% to 70.64%	130 of 202	67
O15Apply path-specific rules for conditional convention loading	92.2%	85.41% to 96.02%	95 of 103	67
O16Determine when to use plan mode vs direct execution	84.1%	76.22% to 89.68%	95 of 113	69
O18Integrate Claude Code into CI/CD pipelines	73.7%	67.76% to 78.94%	174 of 236	52
O19Design prompts with explicit criteria to improve precision and reduce false positives	62.5%	56.31% to 68.4%	152 of 243	56
O20Apply few-shot prompting to improve output consistency and quality	74.8%	66.87% to 81.39%	101 of 135	55
O21Enforce structured output using tool use and JSON schemas	56.3%	48.5% to 63.71%	90 of 160	53
O22Implement validation, retry, and feedback loops for extraction quality	74.0%	66.57% to 80.31%	114 of 154	54
O23Design efficient batch processing strategies	76.2%	69.58% to 81.78%	141 of 185	54
O24Design multi-instance and multi-pass review architectures	83.7%	77.11% to 88.59%	133 of 159	51
O25Manage conversation context to preserve critical information across long interactions	90.4%	84.91% to 93.98%	150 of 166	63
O26Design effective escalation and ambiguity resolution patterns	86.4%	79.77% to 91.14%	121 of 140	57
O27Implement error propagation strategies across multi-agent systems	96.0%	90.9% to 98.27%	119 of 124	62
O28Manage context effectively in large codebase exploration	90.5%	84.2% to 94.52%	115 of 127	57
O29Design human review workflows and confidence calibration	76.5%	69.02% to 82.71%	111 of 145	56
O30Preserve information provenance and handle uncertainty in multi-source synthesis	88.8%	82.81% to 92.9%	135 of 152	59

Intervals here are per item only. These cells are far smaller than the domain cells above, and this page makes no claim that any objective differs from any other.

Not yet published

The other certifications

Same computation, same code path, not enough learners yet. A table of its own is earned at 30 distinct learners, and not before.

Learners per certification, against the floor

Learners per certification, against the floor
Certification	Learners	Floor	Tables published
CCDV-F: Claude Certified Developer – Foundations	223	30	yes
CCAO-F: Claude Certified Associate – Foundations	197	30	yes
CCAR-F: Claude Certified Architect – Foundations	75	30	yes
CCAR-P: Claude Certified Architect – Professional	23	30	no

Short of the line we publish no accuracy table and no objective detail: not a partial one, not a summary. The count is all this page says.

CCAR-PClaude Certified Architect – Professional
23 learners have sat a timed paper on the current question bank, across 40 papers. We publish accuracy by domain at 30 learners.

A certification with fewer than 5 learners is left off this page altogether, including its learner count. At that size a domain profile is close enough to one identifiable person’s results that publishing the shape of it, even as a card saying how few there are, is not something we are willing to do.

Hundreds of learners, not dozens

How people behave under the clock

Far more data than the accuracy tables, because these depend on what a learner did rather than on which question they were served.

Most papers that go wrong are never finished

31.3%489 of 1,561 timed papers, 29.07% to 33.68%

A paper counts as abandoned when fewer than half the questions answered. An unanswered question is marked wrong when a paper is scored, so this is not a low score; it is a paper that was left.

Every timed paper, and the median abandoned one

Every timed paper, and the median abandoned one. 489 of 1,561 timed papers were left less than half answered: 31.3%, 29.07% to 33.68%. 80 of those were left entirely blank. 292 of 829 learners did so on a first paper: 35.2%, 32.04% to 38.54%. The median abandoned paper had 4 of 60 questions answered, handed in at 16.6 minutes.

No outcome and no advice attached: this is a description of what happens, and the largest single pattern in everything we hold.

The flag button knows something

−32.1ppwithin-learner difference, −36.0pp to −28.12pp, 308 learners

Comparing each learner against themselves (items they flagged against items they did not), flagged items come back markedly less accurate. The difference spans every practice answer, not timed papers alone: the button is used far more often away from the clock.

The within-learner difference

The within-learner difference. flagged minus unflagged: −32.1pp, −36.0pp to −28.12pp, over 308 learners. Measured on 566 of 4,133 sittings, sittings where the flag button was used at all.

The two levels, on the current bank

The two levels, on the current bank
Row	Accuracy	Per-item interval	Learner-clustered interval	Correct of answered
Flagged for review	48.9%	not published	not published	940 of 1,922
Not flagged	75.4%	not published	not published	11,050 of 14,656

Absolute levels are published on the current bank only, where they mean something. No interval is published for either, so none is drawn.

Only 566 of 4,133 practice sittings carry a flag at all, so this describes people who use the button. Read it as evidence that the button works (that learners can tell which of their own answers are shaky) and not as evidence that flagging changes anything.

People are less right than they feel when they are certain

Accuracy by what the learner said about the answer

Sure of the answer and Fairly confident: these two separate once answers are clustered inside learners.
Fairly confident and Guessing: these two separate once answers are clustered inside learners.

A bracket says whether that step survives clustering the answers inside learners. A dashed one does not.

Accuracy by the learner's own confidence rating, recorded against each question before the paper was marked.
What the learner said	Answered correctly	95% interval per item	95% interval clustered by learner	Correct of answered	Learners
Sure of the answer	85.9%	84.5% to 87.28%	82.66% to 89.23%	2,079 of 2,419	179
Fairly confident	63.6%	61.73% to 65.46%	60.72% to 66.51%	1,633 of 2,567	246
Guessing	48.3%	45.58% to 50.96%	44.55% to 51.98%	640 of 1,326	234

answers tagged “Sure of the answer” come back more accurate than answers tagged “Fairly confident”; answers tagged “Fairly confident” come back more accurate than answers tagged “Guessing”.

Every self-rated answer on the current question bank, in any mode: a wider population than the tables above, and a small self-selected one: only learners who rate their own answers are in it. We publish it because our readiness model assumes far more of a “sure” than this measures.

Methodology

What we counted, and what we threw out

Seven filters, each applied to whatever the one before it left. Nothing on this page was typed in by a person.

Seven filters, in the order they are applied

The first three leave 938 timed papers on the current question bank

Our own accounts4 papers
Why
We test the product on the product. Those runs are not learners, so they go first.
Papers sat on the retired question bank668 papers
Why
The bank in use until the August release was far too easy to measure anything with. It is more than half of everything we hold, and it appears on this page only in the section that explains why it was dropped.
Drills and untimed study papers2,116 papers
Why
Drills deliberately re-serve questions the learner previously got wrong, so drill accuracy measures the queue rather than the learner. Study mode has no clock, which makes it an open-book instrument. Three instruments reported as one number is not a measurement.
Papers whose question bank moved underneath them219 papers
Why
For every remaining paper we rebuild, from today’s bank, exactly which domains and objectives it presented, and require an exact match with what was recorded at the time. Where it does not match, the question at a given position is no longer the question that was served, and the paper is dropped rather than guessed at.
Questions that were never answered16,084 answers
Why
An unanswered question is marked wrong when a paper is scored, which is right for a score and wrong for a measurement. Every figure above counts only questions a learner actually completed.
Second and later encounters with the same question2,273 answers
Why
Only a learner’s first answer to each question counts. Repeat sittings score several points above first sittings, which is memory rather than learning, and pooling them would flatter every figure on this page.
Questions already seen answered in a drill1,331 answers
Why
Drills use the same questions as the timed papers, and they show the answer: when a learner checks one, and for every question on the review afterwards. A question the learner had already seen answered in a drill before submitting the paper tests memory rather than a first encounter, so that answer is dropped, exactly as the readiness score drops it.

The last four leave 31,899 answers

Bars are scaled within their unit (papers against papers, answers against answers), never across the two. A step that removed nothing shows no bar.

Why there are two intervals

Answers are not independent of one another. They arrive in papers, and papers belong to learners, so one learner having a bad afternoon moves many answers at once. The per-item interval a reader expects, and which every published figure here carries, ignores that, and is measurably too narrow as a result. Grouping the answers by learner gives an interval several times wider, and that wider one is the honest measure of how much this sample can settle. Both are printed at the same weight, in the same row, so nobody has to take our word for which is which.

Interval bounds are printed at the precision they were computed to, and rounded outward rather than to the nearest tenth. Rounding a bound inward would publish a narrower claim than the data supports.

A learner is an account, not a person

Every count on this page counts accounts. Signing in with two providers makes two of them, and we do not link accounts across providers, so every learner figure here is an upper bound on how many distinct people are behind it. That is one of two reasons the threshold for publishing a table sits as high as 30.

The measured answers cover a short window

The banks were edited in place without a version change, so an answer recorded before that edit can no longer be matched to the question that was actually served. The fourth filter above drops those, and it drops a great many of them, which means the accuracy tables describe a narrow and recent slice of history rather than everything we hold. That is a real limit, and we would rather print it than leave it to be inferred.

Answers on the earlier question banks were given on longer questions, and some carry an answer-key cue

Until the question bank labelled 2026-10-02 went live, the questions were longer than the official sample questions. Outside CCAR-F the typical question ran up to about twice their length, and as many as two in five questions on a paper asked for more than one answer. That bank rewrote every paper in the official form: shorter questions and options, four options and one answer on all but a few questions, and “select two” on at most one question in ten.

Until the bank labelled 2026-09-30 went live, the correct option on a single-answer question also tended to be longer than the wrong ones, the longest option on a multiple-answer question was almost always a wrong one, and on CCAR-F the correct option was usually option B. That bank rewrote the options so the key no longer stands out by length, and spread the keys evenly across the letters.

Where this page says the current question bank, it means every bank it counts, and those include the earlier ones. So the accuracy in its tables includes answers given on the longer questions and, on the earliest banks, the cue as well, and a gap between two domains measured on them partly reflects how the items were written rather than only what learners knew.

The retired question bank

Why we threw away more than half the data

The only section where figures from the retired bank appear at all, and the clearest evidence we have that a question bank, and not a learner, produces most of what looks like a finding.

Take one domain, “Prompting and Task Execution”, and put all 7 domains of that certification in order of accuracy. On the retired bank it sat in sixth place counting up from the lowest. On the current bank it sits in first place, 38 points lower than it was.

On the retired bank: 92.3%90.3% to 93.86% · 811 of 879
On the current bank: 54.4%52.05% to 56.74% · 945 of 1,737

Same platform, same domain, same kind of learner, opposite conclusion. Nothing about the people changed between those two figures. The questions did, and that is the entire argument for excluding the 2026-07-23 bank, and the 404 timed papers sat on it, from every accuracy figure on this page.

What these two figures do not license

Neither of those two orderings is one we would publish on its own. The gaps inside each of them are smaller than the uncertainty around them, and the certification this domain belongs to may not have enough learners to earn a table anywhere on this page. The movement between the two banks is the part that is not in doubt: the two intervals above do not come close to meeting.

A second symptom of the same problem. On the retired bank, of 209 papers with no unanswered questions, 209 cleared our own scoring threshold, the lowest of them scoring 745 on a thousand-point scale. An instrument that separates nobody is not measuring anything.

The same release also renumbered domains for CCAR-F, so papers sat before it carry domain numbers that no longer exist. There is no archive of the pre-release tree, which is why those papers cannot simply be remapped.

An earlier internal draft of this page read the first of those figures as a fact about people and drew the opposite conclusion to the one above. It was a property of the questions. This section is here so that mistake stays on the record instead of in the archive.

The negative results

What we measured and would not publish

Four things we computed, looked at, and left out, because a reader cannot otherwise tell a claim we cannot support from one we never thought to make.

Whether studying first improves your first paper: The differences do not survive on the current bank, and the striking version comes from the retired one.
Why
On the current bank the group with no study sections behind them and the group with several are not distinguishable, and the ordering is not even consistent. The striking version of this (a large gap in first-paper scores) comes almost entirely from the retired bank and from unfinished papers, and quoting a score against a threshold would be a back door to exactly the claim we refuse above.
A list of the concepts people get wrong: Structurally impossible rather than merely underpowered.
Why
Each question carries a concept label, and those labels are unique per question: one distinct value for every item in the bank. Any list built from them is a list of individual questions wearing a costume, so we do not build one.
How many learners pass: Off the table permanently, not withheld out of modesty.
Why
This is not a pass rate we are withholding out of modesty. It is off the table permanently, for the reason already published on our methodology page. Note the trap: the retired bank would have supported a spectacular-sounding version of it. That sentence would have been true, checkable, and a fact about a broken question bank.
Anything at all about the certification exams themselves: Our questions, our answer key, our self-selected learners.
Why
Every sentence on this page is scoped to our practice bank, and where you see a domain name it describes how our items for that domain were answered, and nothing further.