```html
Signal Anatomy
A Critical Audit Matter is best read as a written confession. Under the PCAOB's critical-audit-matter standard, the auditor must disclose every matter involving especially challenging, subjective, or complex judgment. That mandate makes CAM text a direct readout of where the audit nearly broke, and it means the signal lives in the topic, not the tally: a CAM on revenue recognition points at the account class most associated with misstatement, while a CAM on pension accounting mostly points at an old actuarial report. Counting paragraphs without classifying them discards the only predictive part.
Tenure arrives through a different pipe. Form AP is filed with the PCAOB within 35 days of audit-report completion, naming the lead engagement partner and slotting them into a categorical tenure band. The field is categorical and machine-readable, so tenure enters the screen as a covariate you condition on, never a score you escalate on. The folk flag — a decade-plus partner implies complacency, therefore cheap early warning — is the least reliable input of the three and reverses sign in the most-cited archival work. Treat it as background conditioning, nothing more.
Going concern runs on two parallel assessments: management's liquidity evaluation covering the twelve months after the filing date, and the auditor's separate evaluation under the professional going-concern standards. Three disclosure states result:
| Disclosure state | Who concluded | What it reveals | Screen treatment |
|---|---|---|---|
| Silent | Neither party found substantial doubt | Baseline; no liquidity stress surfaced | No trigger |
| Substantial doubt disclosed, then alleviated | Auditor raised doubt; management's plans resolved it | A documented management–auditor disagreement settled in management's favor | Highest-information rung; escalate and rank |
| Full going-concern opinion | Auditor concluded doubt was not alleviated | Explicit adverse conclusion on liquidity | Immediate escalation |
The middle rung carries the most information precisely because it is a compromise: the auditor judged conditions severe enough to put substantial doubt on the record, then accepted management's mitigating plans anyway. Judgment difficulty and liquidity stress appear in a single paragraph.
Base rates force the topic weighting. According to the PCAOB's October 2020 Staff Spotlight, large-filer audits average 1.9 CAMs, with revenue recognition appearing in roughly six in ten audits and goodwill impairment in about four in ten. Two consequences follow. First, a three-CAM threshold clears the mean by a wide margin, so crossing it means something. Second, without topic filters the threshold mostly measures industry mix — estimate-heavy sectors cross it routinely, and the screen ends up ranking business models instead of risk.
The pipeline that makes this operable nightly is easy to specify: pull new 10-Ks from EDGAR; extract the Critical Audit Matters block; embed each sentence and classify it against a fixed topic taxonomy; capture going-concern phrases such as "substantial doubt" and "within one year" with regex rather than embeddings, because the trigger leg demands precision over recall; join the filing to its Form AP tenure bucket; emit a composite weekly risk score per filer. Regex on the trigger, embeddings on the ranker — that division of labor keeps false escalations survivable.
For the 2026 filing cohort, everything downstream tests one falsifiable mechanism claim: restatement risk peaks where auditor-judgment difficulty and liquidity stress coincide — CAM intensity concentrated in revenue recognition and fair-value or impairment estimates alongside going-concern language — because both are symptoms of the same underlying estimation uncertainty. Tenure moderates, weakly. That coincidence structure is why the screen fires on going concern, ranks on weighted CAMs, and never lets tenure pull the trigger.
The Empirical Record
US restatement counts climbed steeply across the years covered by the GAO's baseline census. In the years since, according to Audit Analytics' ongoing Financial Restatements series, restatements have continued to arrive at a heavy clip among US filers. Read those two records together and the screening problem comes into focus: restatement risk is not a crisis-era artifact but a permanent, high-volume feature of the disclosure environment. Any monitor built for the current filing cycle is ranking against a target stream that never shuts off.
The going-concern record points the opposite direction on frequency and the same direction on usefulness. According to Audit Analytics' going-concern studies, only several hundred such opinions land each year among thousands of large accelerated filers — a vanishingly rare signal. Yet the bankruptcy-classification research of Hopwood, McKeown, and Mutchler achieves roughly 85 to 90 percent accuracy using going-concern-opinion variables. High accuracy on a rare event is exactly the profile you want in a tripwire: when the opinion fires, the posterior odds move further than any continuous covariate can. That asymmetry — rare but decisive versus common but weak — is the empirical case for a going-concern-triggered architecture rather than three interchangeably weighted inputs.
Why build the screen now rather than trust incumbent review processes? According to the PCAOB's SPOTLIGHT on its 2023 inspections, nearly half of audits at the six largest US firms drew Part I.A deficiencies, a sharp deterioration from two years earlier. Deteriorating inspection outcomes raise the prior probability that dense CAM text reflects genuine estimation trouble rather than defensive boilerplate — precisely the environmental condition under which a disclosure-based monitor earns its keep.
Tenure, by contrast, fails in both directions. The stale-auditor folk theorem holds that a long-tenured engagement partner grows complacent, making seniority the cheapest early-warning flag on the market. According to Johnson, Khurana, and Reynolds in The Accounting Review (2002), the archival record runs the other way: shorter auditor tenure associates with more discretionary accruals and less-persistent cash flows. The most-cited tenure result carries the opposite sign from the folk theorem — the strongest possible reason tenure cannot anchor the screen, and the reason the canonical rule never escalates on tenure alone.
Last, the behavioral precondition that makes timing feasible. According to Backof, Bowlin, and Goodson in Contemporary Accounting Research (2021), investors read CAM disclosure as adverse news about the client. If outsiders systematically price CAM text negatively, then a monitor that ranks filers by CAM composition can position ahead of the price-moving restatement announcement instead of behind it.
| Evidence stream | Named source | Recorded figure | Role in the screen |
|---|---|---|---|
| Restatement growth | GAO baseline census | Steep climb in US restatement counts across the census period | Defines the event class worth ranking |
| Ongoing restatement flow | Audit Analytics, Financial Restatements series | Restatements arriving year after year among US filers | Persistent target rate; the screen never idles |
| Going-concern rarity | Audit Analytics, going-concern studies | Several hundred opinions yearly among thousands of large accelerated filers | Rarity makes the trigger high-information |
| GC classification power | Hopwood, McKeown & Mutchler | Roughly 85–90% bankruptcy-classification accuracy | Licenses GC as the escalation tripwire |
| Inspection quality trend | PCAOB SPOTLIGHT, 2023 inspections | Nearly half of audits at the six largest US firms drew Part I.A deficiencies, up sharply from two years earlier | Raises the prior that heavy CAM text is real trouble |
| Tenure effect | Johnson, Khurana & Reynolds, The Accounting Review (2002) | Shorter tenure: more discretionary accruals, less-persistent cash flows | Sign reverses the folk theorem; stay tenure-blind |
| Investor perception | Backof, Bowlin & Goodson, Contemporary Accounting Research (2021) | CAM read as adverse news about the client | Permits front-running price-moving announcements |
The practical lesson of this record is architectural, not additive. For any filer on your watchlist, pull the latest 10-K and apply the hierarchy the evidence dictates: going-concern language escalates immediately, on its own; absent that, count CAMs weighted toward revenue recognition and fair-value or impairment estimates and require one corroborating distress marker before escalating at three or more; and strip tenure from the feature set entirely, because the best-cited coefficient in the literature runs backward.
Weighting the Three Inputs
Equal weights are how monitoring screens fail. The three candidate inputs — CAM count plus topic mix, going-concern language, Form AP tenure — do not deserve one vote apiece, because they are not three versions of the same signal. One is a trigger, one is a ranker, and one is metadata. The table below is the instrument that forces that assignment: run each input through five tests — coverage, refresh cadence, sensitivity to eventual restatement, specificity, monitoring cost — and let exactly one role survive per row.
| Input | Coverage across filers | Refresh cadence | Sensitivity to eventual restatement | Specificity / false-positive burden | Monitoring cost | Verdict |
|---|---|---|---|---|---|---|
| CAM count + topic mix | Essentially all large accelerated filers, every annual-report cycle since the CAM mandate took effect | Annual, with each 10-K | Rises with estimate-heavy audits; weak predictor alone | Moderate — raw counts admit boilerplate CAMs until topic-filtered | Free (EDGAR full-text search) | Ranker |
| Going-concern language | A sliver of filings in any given year | Annual, with each 10-K; surfaces in the notes or the audit opinion | Direct read on the 12-month liquidity assessment | Highest — rare event with near-term cash consequences | Free (EDGAR full-text search) | Trigger |
| Form AP tenure bucket | Virtually all filers, nearly a decade of engagement-partner records | Updates with each Form AP filing | No reliable direction; sign reverses across the most-cited archival work | None — adds noise, not signal | Free (PCAOB Form AP database) | Metadata |
Read the coverage column first, because it contains the trap. CAM disclosures exist for essentially all large accelerated filers, cycle after cycle; Form AP tenure buckets exist for virtually all filers, with partner names on file for nearly a decade. Going-concern language, by contrast, appears in only a sliver of filings. On coverage alone, tenure looks like the richest feed in the system — and that is exactly backwards. Coverage without discriminant power is inert: a field populated for every issuer conveys nothing if its values carry no information about the outcome.
The specificity column decides the winner. Going-concern language is the only input tied directly to a 12-month liquidity assessment with near-term cash consequences — management and the auditor are on record, in writing, that substantial doubt exists — which makes it the highest-precision distress marker available and the natural trigger. CAM mix cannot clear that bar, because unfiltered counts drag in routine disclosures; its job is ordering filers inside the escalated pool, not opening the gate. Tenure wins nothing. If long partnerships bred complacency, the tenure bucket would be the cheapest early-warning flag on the market; instead it is the least informative of the three inputs, and its estimated effect reverses sign in the most-cited archival work. Zero weight follows — not a small negative weight, because a coefficient that unstable is noise in either direction.
Translate the verdicts into asymmetric bars. The going-concern channel fires unconditionally: any appearance, in the notes or in the audit opinion, escalates the filer immediately. The CAM channel fires only in the tail — three or more CAMs, filtered to revenue recognition and fair-value/impairment topics, and only alongside one corroborating distress marker. The tenure channel carries zero weight in every state of the world. Asymmetry, not equal weighting, is the framework's core output: blend the three inputs evenly and the single high-specificity signal drowns inside the two low-information ones.
Price the workflow and the design locks itself in. All three inputs are free — EDGAR full-text search covers both the going-concern sweep and the CAM text, and the PCAOB's Form AP database carries the tenure fields nobody needs. The parser is a weekend build; weekly rescoring costs pennies. The binding constraint is false-positive triage capacity, not data acquisition, and the downside of a miss is larger than it looks: according to McCarthy Tetrault's 2026 analysis, a company can face a securities class action even when a public correction causes no price decline. That is why the high-specificity going-concern trigger anchors the design — it spends scarce analyst hours only where the base rate justifies them.
The concrete move this week: stand up one full-text query for going-concern phrasing across both disclosure locations, one counter keyed to estimate-heavy CAM wording, one weekly rescore — and delete the tenure column before anyone asks for it.
What the Data Doesn't Tell You
Every joint-signal estimate behind this screen is an in-sample estimate. Critical audit matters entered the opinion only recently (see Signal Anatomy above), which means the entire observable history of the CAM-plus-going-concern interaction spans a handful of annual filing cycles — none of them a full credit downturn layered onto a rate shock. Any precision attached to the joint premium comes with confidence bands wide enough to drive a truck through. The inputs are artifacts of disclosure law, not natural constants: according to Wikipedia's legislative summary, the Sarbanes-Oxley framework shaping today's opinion architecture was enacted as Public Law 107-204, 116 Stat. 745, and statutes get amended. When the disclosure standard moves, the feature definition moves with it, and a monitor trained on the old definition decays silently.
Three structural problems compound the short window. Going-concern language is a rare event and restatements are rarer still, so the joint cell driving the thesis holds a thin count of positive cases; one or two unusual filers can swing the estimate. Survivorship bites from the other direction — restaters get delisted or acquired, and samples conditioned on survivors understate the base rate. And the outcome label is noisy: "restatement" pools fraud-grade corrections with immaterial revisions, so the screen predicts a pooled category, not misconduct specifically.
Variance across cases is where the aggregates hide their heterogeneity. A pre-revenue biotech can carry going-concern language for consecutive years purely as a statement about cash runway, tripping the trigger annually without a single reporting failure; a regional retailer showing the same language beside revenue-recognition CAMs sits in a very different base-rate regime. Complex-but-healthy businesses — derivative-heavy energy marketers, banks digesting acquisitions — typically run three-plus valuation and impairment CAMs indefinitely. First-year engagements inflate CAM counts mechanically, because a new team documents more. Identical signatures, opposite meanings: the screen's precision is sector-dependent even though the rule is not.
The rule breaks in two directions, both worth naming precisely. On the false-negative side, it monitors disclosed signals only: a concealed misstatement behind a clean opinion, one boilerplate CAM, and no going-concern language passes straight through, surfacing later via whistleblower or SEC action rather than the screen. That is a design boundary, not a defect — the rule never promised to catch concealment. On the false-positive side, chronic going-concern filers generate escalation fatigue; the going-concern leg earns its premium only when the language is new or worsening against a liquidity marker, not when it is yet another consecutive appearance.
A final failure mode lives in the analyst, not the data. When false positives pile up, the reflexive repair is bolting a long-tenure flag back on as a tiebreaker — the old complacency story wearing a disguise. Resist it. As the weighting analysis above established, tenure is the least informative input and its sign reverses across the most-cited archival work; a tenure tiebreaker injects noise exactly where you are trying to recover precision. Tenure-blind is the design, not an oversight.
| Failure mode | Trigger pattern | What happens | Correct handling |
|---|---|---|---|
| Chronic GC filer | Going-concern language, repeated annually | Trigger fires every cycle; no restatement | Escalate on new or worsening language plus a liquidity marker |
| Concealed misstatement | Clean opinion, single boilerplate CAM, no GC | Screen stays silent; surfaces via SEC or whistleblower | Pair the screen with external event feeds |
| Complex-but-healthy filer | Three-plus valuation CAMs, no corroborating marker | No escalation under the rule | Hold at watch; require the marker first |
| First-year engagement | Elevated CAM count, mixed topics | Count inflates mechanically | Rank on topic mix, not raw count |
| Tenure tiebreaker | Analyst adds a long-partner flag | Adds unstable-sign noise | Keep the screen tenure-blind |
One discipline closes the gap between published estimates and your own coverage: log every escalation decision this filing season — trigger fired, corroborating marker yes or no, sector, and the outcome at the next non-reliance notice — so the 2027 season opens with an out-of-sample record nobody can argue with. And when someone hands you a backtest of this screen, ask one question first: how many restating filers actually sat in the joint going-concern-plus-CAM cell each year? If the answer is a handful, treat the joint premium as provisional — exactly as this section does.
What the Screen Cannot See
Gutierrez, Minutti-Meza, and Tautiva delivered the strongest counter-result this screen has to absorb. Writing in the Review of Accounting Studies (2023), they find that CAM disclosures add little information beyond what existing filings already contain and produce economically small market reactions. That single finding sets the screen's economics: a going-concern paragraph or a third revenue-recognition CAM is not novel intelligence but a leaf the market has largely priced, so the edge cannot be informational. It can come only from cross-sectional ranking — using the signals to sort thousands of 2026 filers into a review queue faster than sequential human reading ever could.
The base rates then cap how hard anyone can act. Restatement-relevant events run below 1% of filers in a year. Grant the screen performance almost no real screen achieves — detecting nine of every ten true restatements while correctly clearing nine of every ten clean filers — and Bayes' theorem still returns a positive predictive value almost too poor to act on: of twelve escalations, roughly eleven are false alarms. That ratio is a behavioral constraint, not a footnote: a queue this imprecise justifies pulling the file and checking the 8-K feed, and it forbids short positions, client calls, or anything carrying asymmetric downside.
Two mechanical defects corrupt raw counts. CAM topics cluster in a handful of recurring areas — revenue recognition, goodwill impairment, fair-value measurement — so a keyword-level screen does not isolate deteriorating issuers; it flags entire industries at once. And auditors add CAMs defensively after PCAOB inspection criticism, mechanically inflating counts with no change in underlying client risk. A count that rose because the audit firm took inspection heat is indistinguishable, to a naive counter, from a count that rose because the client is failing.
The confounders stack on top. Material weaknesses, recent auditor changes, and going-concern language all co-occur with elevated CAM counts. FilingRadar's finding that multi-year material weaknesses correlate strongly with eventual restatements makes persistent Item 9A disclosures a tempting screening variable — but folding them into one composite alongside CAM counts double-counts a single underlying distress factor and overstates apparent accuracy out of sample. The contamination is substantive, not merely statistical: a study published in the Journal of Business Finance & Accounting (DOI: 10.1111/JBFA.12363) finds that stock-price reactions to restatement announcements differ depending on whether a material-weakness disclosure is associated — the overlapping markers change how the market processes the event itself.
Tenure gets the bluntest verdict. The archival record's inverse-tenure results sit beside the congressional and GAO rotation studies, which found no reliable quality gain from capping tenure, and the PCAOB has never mandated rotation. Evidence points both ways and identifies cleanly in neither. The popular belief that a ten-years-plus partner flag is the cheapest early-warning system therefore inverts the record: tenure is the least informative of the three inputs, reverses sign in the most-cited archival work, and any threshold written into a 2026 screen is convention, not evidence. The defensible build is tenure-blind.
Finally, the clock bounds everything. CAMs and going-concern paragraphs exist only in the annual 10-K; interim 10-Qs carry neither input, so lead time is bounded to weeks — often just ahead of the restatement announcement itself. Driven Brands shows the worst case: on February 25, 2026, the company substituted a restatement 8-K for its expected 2025 annual report, and the stock lost 30.2% in a day, according to the Driven Brands restatement analysis. The screen's inputs never arrived; the Item 4.02 filing preceded the 10-K it was supposed to precede. That failure mode is why the 2026 architecture runs this screen as one module inside a broader SEC 8-K material-events tracker, and why some more sophisticated investors also track little-r revisions alongside big-R events, per the clawback-era analysis in The Vanishing Restatement. Intra-quarter liquidity telemetry is the layer a 10-K-based screen structurally cannot supply.
| Structural blind spot | Hard cap it imposes | Design response that wins |
|---|---|---|
| Disclosures already priced (Gutierrez et al., 2023) | Near-zero novel information; small market reactions | Cross-sectional ranking of filers — never trade the flag itself |
| Sub-1% event prevalence | Positive predictive value dominated by false alarms; roughly 11 of 12 escalations false | Escalation means open a file, not size a position |
| Boilerplate topic clustering | Keyword screens flag whole industries | Require the joint condition: three-plus estimate-heavy CAMs plus one corroborating distress marker |
| Co-occurring distress factors (Item 9A, auditor changes, going concern) | An uncontrolled composite double-counts one factor | Score persistent Item 9A separately from CAM count |
| Tenure carries no reliable signal | Any threshold is convention, not evidence | Omit Form AP tenure from scoring entirely |
| Inputs exist only in the annual 10-K | Lead time bounded to weeks; 10-Qs carry nothing | Bolt the screen to an 8-K events tracker with little-r revision monitoring |
Build to the right-hand column: rank, triage, separate the confounders, drop tenure, and wire the screen into the 8-K stream — the blind spots then become design constraints instead of silent failures.
Retroscreening Carillion
Signature after signature, year after year. Across Carillion's entire life as a listed company, KPMG signed every annual audit the firm filed — the maximum tenure bucket the screen recognizes. The closing full-year opinion disclosed exactly two key audit matters, construction-contract revenue recognition and pension-scheme valuation, and neither the full-year report nor the intervening interim reports carried a going-concern qualification or even an emphasis-of-matter paragraph. Those three inputs are the complete case file; everything else is outcome knowledge a contemporaneous reader did not have.
Run those inputs forward and the verdict writes itself: investigate, do not escalate. The KAM count lands on the low end of the scale, and the going-concern leg never fires.
```
Frequently Asked Questions
How quickly after an audit report is finished does the lead engagement partner's identity become available to screeners?
Form AP is filed with the PCAOB within 35 days of audit-report completion, naming the lead engagement partner and slotting them into a categorical tenure band.
When a company disclosed substantial doubt about its ability to continue but the doubt was later resolved, what does that disclosure state actually reveal?
It reveals a documented management–auditor disagreement settled in management's favor, making it the highest-information rung of the three disclosure states and warranting escalation and ranking.
How many critical audit matters does a typical large-filer audit contain, and why does a three-CAM cutoff carry weight?
According to the PCAOB's October 2020 Staff Spotlight, large-filer audits average 1.9 CAMs, so a three-CAM threshold clears the mean by a wide margin.
Why does the nightly pipeline use regex instead of embeddings to detect going-concern language?
Because the trigger leg demands precision over recall, phrases such as "substantial doubt" and "within one year" are captured with regex rather than embeddings.
Isn't a decade-plus engagement partner the cheapest early-warning flag for auditor complacency?
No—Johnson, Khurana, and Reynolds found that shorter auditor tenure associates with more discretionary accruals and less-persistent cash flows, so the archival record runs opposite the stale-auditor folk theorem.
If a watchlist filer shows three or more weighted CAMs but no going-concern language, what additional condition must be met before the screen escalates?
Absent going-concern language, the screen requires one corroborating distress marker before escalating at three or more CAMs weighted toward revenue recognition and fair-value or impairment estimates.
Quick answers
| What does the PCAOB's October 2020 Staff Spotlight report about average CAM counts and topic frequency in large-filer audits? | Large-filer audits average 1.9 CAMs, with revenue recognition appearing in roughly six in ten audits and goodwill impairment in about four in ten. |
| How should auditor tenure enter the restatement risk screen according to the article? | Tenure arrives via Form AP's categorical tenure band filed within 35 days of audit-report completion, so it enters as a covariate you condition on, never a score you escalate on. |
| What did Johnson, Khurana, and Reynolds find about auditor tenure in The Accounting Review (2002)? | Shorter auditor tenure associates with more discretionary accruals and less-persistent cash flows, carrying the opposite sign from the stale-auditor folk theorem. |
| Which going-concern disclosure state carries the most information and how is it treated? | Substantial doubt disclosed and then alleviated is the highest-information rung because it documents a management–auditor disagreement settled in management's favor, so it is escalated and ranked. |
| Why does the screen use regex rather than embeddings for going-concern language? | Because phrases like 'substantial doubt' and 'within one year' form the trigger leg, which demands precision over recall to keep false escalations survivable. |
Also worth reading: The Evolution of PCAOB's Integrated Audit Standards Key Changes and Impact Analysis 2024: Evolution of PCAOB's Integrated Audit · 7 Critical Changes in PCAOB's 2024 Guidelines for Internal Control Over Financial Reporting Assessment: 7 Critical Changes in PCAOB's · Latest Changes in Performance Materiality Thresholds Analysis of 2024 PCAOB Guidelines and Their Impact on Financial Statement Audits: Latest Changes in Performance Materiality