OSO Audit

Capabilities · the full inventory

Everything it actually does.

An inventory, not a pitch. Every capability below is built, running, and demonstrable on a complete engagement today.

11,683
standards clauses across 19 frameworks
28
detection engines in one waved pipeline
8
gates every AI output passes, failing closed
0
conclusions delivered on the model's authority

01 · The wedge

Every conclusion resolves to a clause.

Not a citation typed into a memo — a stored link to the primary-source paragraph, with its verbatim text, effective date and amendment history one click away.

osoaudit.com/app/trace
Trace view: a finding linked through risk, assertion and procedure to the standards clause behind it, with clause chips
The trace, in the running system — one action, conclusion to clause.

Four frameworks carry 95% of the corpus

11,683 clauses across 19 frameworks, counted from the live corpus — not estimated.

ISA & ISQM IFRS PCAOB AS IESBA 15 more ISA and ISQM — 6,463 clauses, 55% of the corpus IFRS — 1,766 clauses, 15% of the corpus PCAOB Auditing Standards — 1,445 clauses, 12% of the corpus IESBA Code of Ethics — 1,424 clauses, 12% of the corpus Fifteen further frameworks — 585 clauses combined, 5% of the corpus 6,463 · 55% 1,766 1,445 1,424 585 11,098 clauses — 95% of everything the file can cite
View as table
FrameworkClausesShare
ISA & ISQM — IAASB6,46355.3%
IFRS1,76615.1%
PCAOB Auditing Standards1,44512.4%
IESBA Code of Ethics1,42412.2%
Fifteen further frameworks5855.0%
Total11,683100%
Also PCAOB practice alerts and rule-adopting releases, SEC Reg S-X, SOX, the EU AI Act and regional regimes. Full breakdown →

Citation-resolving search

Type ISA 315.31 and land on ISA 315.31 — not on the paragraphs that merely mention it.

Effective-date aware

Anchored to the standard as it stood, not as it reads now.

Standards watch

The corpus ages against the regulators, not your last release.

Custom clause packages, with two-person authoring

Firms add their own methodology as first-class clauses, held to the same two-person approval as the file itself.

02 · Depth

The parts that took the longest to build.

Three capabilities that exist because the ordinary version of them is not audit-grade. Each is running, and each is measured below rather than claimed.

A twin of the control, with adversarial cases fired at it

Testing whether a control operated is the easy half. This asks the harder question: is the control, as designed, capable of catching what it claims to?

The compiled control its actual logic, not its description Adversarial cases split invoices · aliased users · backdating Run in memory, quarantined no synthetic row ever reaches the audited file a trial never writes a finding Four verdicts, and only one of them is interesting ✓ Hit should catch — and caught ✕ Miss → a named design gap should catch — and let it through ✓ Clean should pass — and passed ! False positive should pass — but caught What a case should do is recorded independently of what the program actually does — so the twin can disagree with the control.
A miss comes back as a design gap in words — for example, that matching on a single document prefix hides a break in the sequence across prefixes.

A working twin of how the client actually operates

Reconstructed from the audited ledger itself — the same rows the file audits. No model anywhere in it; the figures are a pure function of the GL.

Settlement lag  p50 / p90 days Period-end load  % Hand-off intensity  per journal entry Rework share  % Throughput  entries / month

The process reality behind the balances: when payments actually settle, how much of the period lands in its last five days, how many hands touch an entry, how much gets redone. It goes to the client as a signed deliverable — and only signed figures are charted, so every mark on the page can be verified by verifying the run.

Uncertainty drawn honestly, or not drawn at all

Five scenario models, 500 samples each. The seed is the hash of the resolved assumptions, so the same assumptions replay to the same output — a divergence is a real change, never noise.

Outcomes are twenty quantile dots a frequency framing — never a bar with an error whisker lower outcome higher outcome — the long tail is visible, not averaged away Projections are bands with no centre line none is drawn, because none is computed p10–p90 p25–p75 month 1 month 12

Sensitivity is shown as tornado bounds at each driver's P10 and P90, ordered by impact — and the one-driver-at-a-time caveat travels attached to the data, not in a footnote somebody can crop.

The models: working-capital release · liquidity runway · settlement-tail compression · period-end smoothing · lease-versus-own NPV. Business quantities only — a scenario never computes an accounting consequence.

Also shipped, and less loud about it

Peer benchmarking against same-sector public filers Engagement memory, append-only Smart work allocation to staff Air-gapped mode refuses the cloud provider outright

03 · AI that is accountable

The model proposes. A named human disposes.

Every AI output enters as a proposal and leaves as a judgment attributed to a person. Not a setting — the only path.

The only path an AI output can take into the file

Eight gates screen it. A partner's typed attestation — never a gate — is the decision to deliver.

AI proposal basis · evidence · confidence G1G2 G3 eight gates · every one fails closed Partner attests typed · recorded · bound Judgment on file attributed to a person Gate cannot evaluate Refused. Nothing is delivered. The failure mode is silence — never an ungated answer.

The human-EP gate is a hard lock

No configuration, no override, no advanced mode removes the partner from the path to an opinion.

A reasoning graph behind every finding

Which signals combined, in what order, to reach the conclusion on screen.

Ask the engagement, in plain language Faithfulness evaluated as a release gate Model spend in cents per engagement

04 · Fieldwork depth

A real audit file, not a findings dashboard.

Planning through completion, each workpaper a real artifact with its own sign-off — not a tab that exports to Excel.

Twenty-eight detection engines, run as one waved pipeline

Findings arrive ranked, each carrying the engine that raised it and the clause behind it.

Journal Entry Testing Benford's Law Segregation of Duties Graph / Network Fraud ML Ensemble Analytical Procedures Duplicate Detection ICFR Control Testing ITGC Assessment Compound Risk Continuous Monitoring Predictive Risk Subledger Analysis Qualitative Analysis Consolidation Document Intelligence Evidence Matching Industry Rule Packs AI Governance Inspection Readiness Regulation Rules Confirmation Management TB Reconciliation +5 more

Materiality recomputes live, and cites itself

Change a benchmark and all three thresholds move on screen, each carrying the clause it derives from.

A group audit does not average

An unresolved component blocks the group conclusion rather than quietly averaging into it.

A control is compiled, then reperformed over the whole population

Not a sample of 25. The certificate replays, so the result can be re-executed rather than believed.

The control, in words the firm's own description Compiled to logic executable · versioned Reperformed over 100% of the period's transactions signed certificate · replays on demand AS 1105 ISA 500 the clauses it answers to — carried on the certificate, not typed in afterwards

And the rest of the file

Fraud risk & presumed risks Related parties Going concern Group audits Sample design ICFR ITGC Statement of cash flows Any GL, any column names Document-driven autofill PBC consolidation SEC EDGAR filings

05 · Defensibility

Built to be inspected, not just to be used.

An inspector asks you to prove what you did, when, and on what data. The file writes the answer as you work.

What the inspector asksWhat the file already holds
Who decided this?The person, the moment, and a fingerprint of the data behind it recorded
Has it been altered since sign-off?Sign-off is HMAC-bound to the file's state. Alter it and verification fails tamper-evident
Show me the basis for this conclusion.One button: the clause, the evidence and the judgments beneath it
What was open, and who cleared it?The judgments queue, lowest confidence first, plus manager review state
Was the team independent and competent?Credentials checked against the issuing registry, not self-asserted
Would this survive inspection?The inspection simulator, run against the file as it stands
Was it archived in time?Retention and archive windows the system enforces, not a remembered date

What the seal actually does

Sign-off is HMAC-bound to the state of the file at that moment. The check is arithmetic, not a promise.

The file as signed Seal verifies — unchanged since sign-off One field altered afterwards Verification fails — the change is evident Tamper-evident, not decorative. The same seal is what an archive check re-runs years later.

06 · Trust & deployment

It runs on your infrastructure. The data never leaves.

Self-hosted by design. There is no vendor cloud in the path of the audit file.

Isolation the database enforces, not the application

A dedicated database per tenant, with row-level security inside it. A cross-tenant read returns nothing.

Tenant A Dedicated database Row-level security inside it Tenant B Dedicated database Row-level security inside it Blocked 0 rows returned Enforced by the database — not by application code remembering to filter.
Multi-factor authentication SAML / Entra SSO Point-in-time recovery, drilled Metrics, traces and logs Cryptographically signed licences, verifiable offline

The restore is exercised, not assumed. Read the security posture →

07 · Adoption

The learning curve ships with the software.

An experienced auditor needs the route through a new file. A new joiner needs teaching. Two artifacts, not one manual.

28

guided walkthrough tasks

In-app, against a live engagement — one task at a time.

39

course chapters

A built-in course with a certification path, taught in the product.

10

sector risk packs

129 sector risks and 123 real enforcement cases — planning starts from what has actually gone wrong.

Built and running. Shown on a complete engagement.

Everything here can be shown on a complete engagement. We would rather you check that than take it on trust.

See it unstaged How it works