Changelog

What changed, when, and why.

Every fix to the question bank and the product, in the open. Including the embarrassing ones, because a correction you cannot see is not a correction.

Tags: FIX something was wrong. ADD new content or feature. CUT something removed or retired. MEASURED we ran the numbers and published them.

September 19, 2026
  • FixAccrued liabilities question whose table did not match its answer. FAR-46001 keyed $35,700, which is salaries plus interest plus a $7,200 warranty obligation, but the warranty row and the $9,000 January dividend row were missing from the table and a utilities row that belonged to a different question had taken their place. Under the table as shown, the right answer was not among the options. The two rows are restored, the stray row removed, and every option now reconciles to the explanation. Found in routine maintenance.
  • FixSweep of every table question for the same defect: 19 more found and rebuilt. After FAR-46001 we checked all 698 MCQs that carry a data table, plus every question where the stem's figures barely reappear in the answer. 291 candidates were read in full. 19 had a table that did not support the keyed answer: 8 had a whole table pasted in from a different question (BAR-52072, 58016, 62013, 74071, 74175, 76041, 50055, 52107), 6 had rows whose amounts or labels disagreed with the explanation (BAR-50006, 52043, 76022, FAR-32054, 32079, 76013), and 5 were missing one input the answer depends on, such as a useful life or an allowance balance (BAR-28096, 34037, 50003, 76026, FAR-44035). Each stem was rebuilt so that the keyed answer and all three distractors reconcile to the table. Keys and explanations were already correct and did not change. 19 of 17,658 is 0.11 percent of the bank.
  • FixQuick ratio distractor did not foot. BAR-28310's explanation said total current assets were $280,000; the table sums to $290,000. The distractor is now 3.22 to 1 and the explanation uses $290,000. The keyed 1.78 was always right.
  • MeasuredAudited every explanation for letters pointing at the wrong option: 18 fixed out of 17,658. The keyed answer was right in all 18; only a pointer inside the explanation was stale, left over from evening out answer positions. The 27 phrases now name the thing instead of the letter, so a reshuffle cannot break them again.
September 17, 2026
  • AddPer-question timing and answer changes in your progress export. Each MCQ record now carries the seconds you spent on it before committing (explanation reading is not counted), the first option you touched, and how many times you changed your pick. This makes "accuracy under time pressure" and "right answers changed to wrong" readable from the file you email yourself. Nothing is sent anywhere you do not send it.
  • FixIndependence simulation had eight people on the roster and graded seven. A review of the sample noticed the eighth staff member was named in the exhibit but had no answer fields. Two fields added (not a covered member, no impairment, ET 0.400.12 and 0.400.15).
  • FixWrong citation on a constructive receipt question. The explanation cited Treas. Reg. §1.451-1 for the definition of constructive receipt. The definition is in §1.451-2. Key unchanged.
  • FixEncumbrance question did not say who approved the purchase orders. Under GASB 54 the fund balance classification of an outstanding encumbrance depends on the level of authority behind it, and the stem left that out, so committed and assigned were both defensible. The stem now names the finance director. Key unchanged, explanation rewritten to say why that matters.
September 13, 2026
  • AddSeven questions on the OBBBA business provisions. The September 4 sweep stopped at SALT and the estate exclusions, so nothing in the bank asked whether bonus depreciation is permanent or still phasing down. The seven cover that, the January 19 2025 acquisition-date trap, the 2026 Section 179 limits ($2,560,000, phasing out from $4,090,000), Section 174A domestic expensing, and Section 199A, which OBBBA made permanent with a wider $75,000 phase-in range ($150,000 joint) and a new $400 minimum deduction where aggregate QBI from all active qualified businesses is at least $1,000. Blind-solved twice, key stripped.
  • FixFive questions re-anchored to the law they describe. One taught the TCJA bonus phase-down as live law; two used pre-OBBBA Section 179 limits with no year attached. Each now says which law it describes and carries the 2026 figures. No keys changed.
  • FixSix approach hints the September 4 sweep missed. It updated stems and explanations but skipped the hint field, so six still taught the $10,000 cap. One read "TEMPORARY, sunsets after Dec 31, 2025" on a question whose key said the opposite.
September 4, 2026
  • FixTCP-58029 re-keyed. The key prorated a state tax refund between deductible and capped amounts; Rev. Rul. 2019-11 does not prorate, it recomputes the prior-year deduction as if the correct tax had been paid. Re-keyed to the recomputation answer, year anchored to 2024. Found in our own sweep below, so no bounty was paid; a subscriber finding it first would have cost us $50, which is the arrangement working as designed.
  • FixLaw-currency sweep of both tax sections after the One Big Beautiful Bill Act. The 2026 SALT cap is $40,400, not $10,000, and the TCJA sunsets many tax questions were written against are now permanent law. We flagged and adjudicated 106 questions across REG and TCP and changed 73: 4 rebuilt on 2026 facts where the law had outgrown the keyed amounts (TCP-38017, TCP-38018, TCP-50043, TCP-72051; their keys were right when written), 3 explanations corrected under unchanged keys (TCP-58027, TCP-45035, and REG-48492, which still taught the estate exclusion sunset as pending), 4 premises modernized, and 61 current-law notes added to explanations that remain correct. 33 flags cleared with no change needed, including one false positive where the flagged statement was simply correct law. The 77-row fix log is retained and ties to these counts.
September 2, 2026
  • AddThree simulations, found by tearing down a competitor. While pulling apart another course's EPS simulation (and it was a mess) we checked our own bank and found the gap ran both ways: one simulation on public company reporting topics, zero on earnings per share, zero on contracts. Three commissioned through the same pipeline as the 600. One model writes, two models from different companies blind-solve field by field with the key stripped, and a CPA reads it before it ships. The three: weighted-average shares and basic EPS, diluted EPS with antidilutive-security screening, and contract formation with statute of frauds and agent authority. The bank is now 603 simulations, 8,435 graded fields; the blueprint page and the published build log reflect it.
August 30, 2026
  • MeasuredWe ran our own Tell Index test against our own bank, and failed it. The test: is the correct answer at least 1.5x as long as every wrong answer, and long enough to notice? In AUD it was true of 10.0% of Advanced questions and 10.8% of Core; ISC measured 9.5%. Worst case: a 76-word correct answer beside a 27-word field. A candidate who knew nothing could beat random guessing on those sections by picking the longest option without reading it.
  • Fix1,915 questions rewritten across all six sections. The tell is now 0.1% of the bank: 24 questions out of 17,651, each one a defined phrase or required list that cannot shrink without breaking. The excess was almost always a clause explaining why the answer is right, which was moved to the explanation where it belongs; average answer options are now several words shorter bank-wide. Every rewritten question was re-solved by a reviewer given only the new wording, no answer key and no explanation, and told to flag anything ambiguous. That check caught three rewrites that had cut something load-bearing; all three were restored, re-checked, and are named in the fix log. It also enforced the reason none of this was scripted: in dozens of questions a wrong answer reaches the same conclusion as the right one for a different reason, so the reasoning is the entire question and a find-and-replace would have destroyed them. Keyed answers and option order are unchanged throughout. Also fixed along the way: 54 questions that printed the same opening phrase in all four answers (moved into the question), one question whose stem and answers described two different fact patterns (rebuilt), and 16 explanations that referred to the wrong answer letters.
  • MeasuredThe blind check found a keyed answer that was wrong, and it was ours. FAR-44043 says "Under current U.S. GAAP" but was keyed to the IAS 1 contractual-right-to-defer test; current U.S. GAAP is ASC 470-10-45-14, intent and ability to refinance, and the explanation argued against the correct answer. FAR-52223 is missing the table figure its answer depends on. Both are pulled pending rewrite. The first was caught because a reviewer who could not see the answer key disagreed with it, which is the entire reason that step exists.
August 21, 2026
  • FixA wrong keyed answer, found in CPA re-review. REG-59006 asked how much of a warehouse gain was unrecaptured Section 1250 gain when the taxpayer had $15,000 of nonrecaptured Section 1231 losses. The key preserved the full $25,000 depreciation amount; under Notice 97-59 and Reg. §1.453-12 the Section 1231(c) recharacterization reduces the 25% group first, so the correct answer is $10,000, which was not among the options. Both blind solvers and the judge had agreed on the wrong ordering. The question was rewritten with the correct key and options. A sweep of all 28 questions pairing the lookback with unrecaptured Section 1250 gain found no other instance.
  • FixThe keyed answer sat in the first dropdown position in 45% of simulation fields. Measured across all 3,238 dropdown fields in the library: 45.3% first-position against a uniform expectation near 25%, a writer-model artifact the build pipeline never shuffled. Every non-ordered dropdown was reshuffled deterministically; ordered lists such as numeric ranges kept their order. The first-position rate is now 22%. The trial simulations and the homepage demo received the same fix.
  • FixThe citation audit's own correction was wrong, and we shipped it. The August 13 audit moved the $100 trust exemption citation from IRC §642(b)(2)(A) to (B) in several simulations, both review models agreeing. The statute says the opposite: (A) is the $100 general rule, (B) is the $300 rule for trusts required to distribute all income currently. Caught during item selection for an accuracy benchmark; six authority fields across five simulations corrected back to (A), with the simple-trust $300 citations to (B) confirmed correct and left alone. Unanimity is not verification, including when the unanimous parties are the auditors. The citation ledger will be reissued with these rows and its hash updated.
  • FixDropdown options that restated an earlier field's answer. Two simulations' conclusion menus quoted computed results from fields above them, letting a reader back-fill graded fields from the option text (AUD-S2SAM-mzo83j, REG-S2CAP-2jflew). Both were rewritten without the numerals. A sweep flagged 75 more simulations for review under the same pattern; confirmed fixes will land here.
August 18, 2026
  • FixNine simulations rendered their intro paragraph as raw HTML tags. One batch of stems came out of the writer double-escaped, so the opening scenario showed literal <p> markup instead of formatted text: one AUD, one FAR, four REG, three BAR (AUD-S2SUB-2yufyz and family). Display defect only — the tasks, answers and grading were unaffected. Found while screen-recording the product, which is not the gate that should have caught it; an escaped-markup check is being added to the sim validator.
August 13, 2026
  • MeasuredThe citation audit announced yesterday is finished. Two independent models (OpenAI and Google, both at maximum reasoning) each read all 5,840 authority citations across the 442 FAR, BAR, REG, TCP and ISC simulations. 4,008 citations passed both reviewers. Every disagreement went to human adjudication: nothing changed unless the models agreed, or a confirmed error pattern plus an authority lookup said it should.
  • Fix347 citations corrected. 119 were unanimous two-model corrections; 184 were dual-flagged rows resolved by adjudication and primary-source lookups; 44 came from sweeping confirmed error families through the single-model flags. The systematic finds: COSO Principle 11 (technology general controls) cited for application controls across five ISC simulations where Principle 10 governs; the $100 trust exemption cited to IRC 642(b)(2)(A) instead of (B) in four simulations; the treasury-stock paragraph family from yesterday's entry, now swept bank-wide. Three of the unanimous corrections were themselves wrong: both models agreed derecognition lives at ASC 842-30-25-1(a), but (a) is the net investment — the requirement sits in the paragraph's lead-in sentence. Unanimity cannot catch a shared error; that is what the human layer is for, and it is why the August 7 Section 179 rule exists. Every change, including those, is in a downloadable before/after ledger (1,808 rows: the 349 corrections, each labeled with how it was caught, plus every restyled AU-C and AT-C citation), SHA-256 published on the methodology page. It can be reverted by script.
  • Fix274 AT-C citations restyled to section and topic, extending the AU-C granularity policy on the methodology page: the review models could not agree on the current AT-C paragraph map (the SOC report-type definitions are .08 to one model and .09 to another), so those citations now sit at the finest level that can actually be verified. 177 other flags were dismissed as noise — GAAP defines no formula for a ratio, so a nearest-topic anchor on a ratio field is a style choice, not an error. 19 flags remain open pending manual codification checks and will land here when resolved.
  • MeasuredTwo simulations received correct citation fixes but are queued for scenario review on content grounds: both are premised on the TCJA rate sunset (a 39.6% reversion path and a reverted AMT exemption), which OBBBA made moot. The citations now match what the simulations say; the question is whether the simulations should still say it.
August 12, 2026
  • Add600 task-based simulations. Every one written by an AI model, blind-solved field-by-field by two models from different companies with the answer key stripped server-side, and shipped only when both reproduced the key and a licensed CPA reviewed it. Each simulation grades per field with its own explanation and authority citation, then diagnoses missed fields by the underlying concept — "you missed five, but four of them are one idea" — and teaches the method for attacking that kind of simulation. 8,395 graded fields across the library. Pipeline and numbers on the methodology page.
  • CutThe entire old simulation library. An earlier version of our own llms.txt called it the weakest part of the product because it never went through the pipeline the questions did. It is retired outright rather than patched — including the 18 simulations we audited and fixed on August 4, which are retired with it. Nothing from the old library survives in the product.
  • Measured944 build attempts produced the 600. 649 attempts saw the blind solvers reproduce the entire key, seven of them on a single completed solver; 30 passing attempts were discarded on review, 8 were duplicate builds of the same simulation, and 11 more fell at final validation and the pre-ship read — early-prompt builds, reused entity names, calendar-dated tax facts — 4 failed final mechanical validation, and 1 was destroyed by a spreadsheet cell-size cap before it could be saved. Total cost $427, or 71 cents per shipped simulation. The full log is published with a SHA-256, every reject and reviewer kill included.
  • CutA simulation that was live on our public sample pages. The TCP showcase simulation carried an entity name from before our naming rules and was killed in CPA review; it has been replaced on sims and sample-questions with a current-generation simulation whose verification record is published beside it. Two more were killed for computing against a named tax year rather than stated figures.
  • FixThe mastery grid's simulation checkmark now requires actually engaging with the simulation — every field attempted, or at least half right. A blank submission no longer ticks the box. The accuracy gate also moved from 8-of-your-last-10 to 7-of-10: at a true 75% skill level, the old bar failed a genuinely passing candidate about half the time on a 10-question window, which made the grid discouraging rather than honest. And a bookkeeping bug that recorded each simulation twice in study history is fixed.
  • AddThe blueprint coverage map now shows simulations per Blueprint topic alongside question counts, down to the individual topic rows.
  • FixOne wrong authority citation in a REG simulation, caught in CPA review. A stock sale correctly classified as short-term capital gain cited IRC §1222(3), which defines long-term gain; the right subsection is §1222(1). The answer, amount, and holding-period analysis were all correct — only the citation was off. A sweep of all 600 simulations for the same class of error (character vs. cited §1222 subsection) found no other instance. Fixed in the bank and on the sample pages.
  • CutParagraph-level AU-C citations, replaced with section and topic — 1,234 fields across 116 simulations, e.g. "AU-C 315.28" is now "AU-C 315 (identifying and assessing risks of material misstatement)." The reason is disclosed rather than dressed up: SAS 143 and SAS 145 rebuilt the paragraph numbering of the estimates and risk-assessment standards, and when we ran three frontier AI models over our paragraph-level AU-C cites they could not agree on the current map — including confidently "correcting" citations that were right. Section-plus-topic is accurate, stable across recodifications, and checkable by anyone. IRC and ASC citations keep full subsection precision because their numbering is stable; those are getting a two-model independent audit instead. Every one of the 1,234 changes is logged with its before and after.
  • FixThirteen more wrong citations across four FAR equity simulations, same review, same day. Treasury-stock reissuances below cost cited ASC 505-30-30-9 (retirement of shares) instead of 505-30-30-10 (sales of treasury stock) — twelve fields across four simulations built from the same archetype, the same wrong paragraph replicated by the same writer prompt. One stock-split memo-entry field cited the stock-dividend threshold paragraph instead of ASC 505-20-30-6. Answers and explanations were correct throughout; the blind-solve gate verifies answers and never reads citations, which is exactly why the human read-through exists. A model-assisted audit of all 7,087 professional citations in the simulation library is now underway; findings will be posted here.
  • MeasuredThe CPA read-through reached 100% of the bank. The post-build read-through that started at the March build has now covered every one of the 17,651 questions. The methodology page is updated to say so — and to keep saying precisely what it means: read and signed off by the CPA; the independent model re-check was the two reviewers' job at build time. The daily reading continues as re-review, so corrections keep landing here.
August 7, 2026
  • Add1,007 new questions. The bank goes from 16,644 to 17,651. These fill the two holes the August 5 re-tag exposed: cost recovery went from zero questions to 125, and section 1231/1245/1250 recapture from 3 to 102. Estates and trusts went from 8 to 75. Every question was commissioned against a single Blueprint point, so no two were written to the same brief.
  • Measured1,150 generated, 142 deleted at the gate, a 12.3% discard rate against 37.2% on the March build. Same gate, better generator: the writing prompt was rewritten after reading what the first build's 9,741 rejections had in common. The full August log is published, all 1,150 attempts with both gate verdicts.
  • CutA question we had already shipped. A Section 179 item computed its phase-out on the pre-July-2025 thresholds. Both blind solvers agreed on the wrong answer and the judge passed it, because all three models had learned the same superseded figure. Unanimity proves nothing when the error is correlated, and that is the one failure this gate cannot catch by design. Pulled the day it was found. Any question naming a real tax year whose answer turns on a statutory dollar threshold now goes to human review before it ships.
  • FixA corrupted question, found by reading rather than by the gate. BAR-52064 asked for unassigned fund balance. Its stem listed four revenue amounts while its explanation classified a prepaid, a statutory restriction, a council resolution, encumbrances and a mayoral set-aside. The options, the keyed answer and the explanation all cohered with each other; only the stem's table did not, so this was data corruption from the table repair logged on August 4, not a badly written question. The table was restored from the components the explanation specifies and the question is back in the bank. Every gate we run checks whether a question is answerable and whether the solvers agree; none of them checks whether the explanation is describing the same fact pattern as the stem. That check is being added, and the other questions touched by that repair pass are being re-read.
  • FixReconciled the topic map to the AICPA's own wording. Merged 108 questions into Managerial and cost accounting, renamed 122 to the Blueprint's Investment alternatives using financial valuation decision models, and fixed 219 more that differed only by case or punctuation and would have split questions across phantom topics. No question changed content area; the map now matches the source.
  • FixSwapped two questions into the free REG sample so it stops being easier than the bank it advertises. Sample difficulty moves from 43.6 to 46.4, against a REG bank average of 46.3. One replacement is a post-OBBBA SALT cap question. Also removed an orphaned tell tag the swap left behind, the same defect we cleaned up on August 6.
August 6, 2026
  • CutRemoved a projection from the free sample's results screen that told students they "typically reach 75 in about N days of focused practice with ChatCPA." The day counts were hardcoded, not measured. Removed a "90% pass at 2,000 MCQs" milestone from the same screen for the same reason. An outside reviewer found the first one.
  • FixReplaced six AUD questions in the free sample that were materially easier than the paid bank. Sample average difficulty moved from 40.5 to 44.5, against a bank average of 46.3. Four more are still weaker than they should be; they are being rewritten rather than swapped, because every existing question on those topics had the same problem.
  • AddPublished the build methodology: the full pipeline, the 37.2% discard rate, what each rejection code means, what the build cost, and a downloadable log of all 26,162 generation attempts.
August 5, 2026
  • FixRe-tagged all 158 REG property-transaction questions, which had been sitting under one generic label, into five proper Blueprint topics.
  • MeasuredThat re-tag exposed a real hole: zero questions on cost recovery (MACRS, section 179, bonus depreciation, depletion) and only three on section 1231/1245/1250 recapture. Both are being written.
  • AddThe blueprint coverage map now expands: click any content area to see every topic inside it and how many questions each holds.
  • FixRebuilt the free 20-question sample. It was averaging 30% easier than the paid bank because we had been selecting short questions, and short questions are easy questions.
  • CutRemoved 279 questions that shipped in the sample file but were never shown to anyone.
August 4, 2026
  • FixAudited all 18 practice simulations and fixed 12 of them. Real errors: self-employment tax computed without the wage base cap, a refund figure off by $11,652, a WACC built from two incompatible rate chains, a contradictory revenue-recognition conclusion, and a proof of cash that did not reconcile to its own scenario.
  • FixTwelve simulation answer fields could never accept their own correct answers. Grading now tolerates dollar signs, commas and case.
  • FixSwept all 16,644 questions for explanations that cited the wrong option letter after answer shuffling. Found and fixed 8. Every answer key was correct; only the prose was stale.
  • CutRemoved an unsupported pass-rate milestone ("83% pass at 1,000 questions") from the results screen. We have no study behind that number, so it should not have been there.
  • FixReplaced the predicted-score model in the free sample. It used to show a number after a handful of answers; it now shows an honest range and stays hidden until 100 questions are in.
  • FixRe-tagged 87 questions that were filed under the wrong Blueprint area, and merged four duplicate topic labels that were splitting 78 questions across phantom topics.
  • FixMobile layout overflow and an unreadable dark-mode heading in the practice app.

Older changes were not logged. This page starts the day we decided that publishing our own corrections was worth more than looking flawless. Spotted something wrong? Tell us and it will show up here.