Audit-side recomputes indexed by the assertion they exercise

Chain and tool matrix for Audit-side recomputes indexed by the assertion they exercise: each column is a tool this guide covers, most used first, each row is one chain, and a brighter cell means more of that row's chains run through the tool. Hover a cell to list them.

Index current as of 2026-11-06 · every entry links a node, workflow, or tool page live on this site · each line paraphrases that page's own scope statement

For anyone assembling audit-side evidence on figures they already hold: this page indexes the OpenChainGraph nodes, workflows, and the offline receipt verifier in the Audit & Assurance domain by the financial-reporting assertion each one recomputes something about, so the tool that answers a question is findable by the question rather than by its mechanism's name.

🔒 All inputs are processed locally in your browser. No data is transmitted. Do not enter real personal data — use synthetic or anonymised inputs only.

Every figure these tools touch is a declared input: a hash, a row count, a control total, a confirmation response, a ledger balance, a journal-entry extract. What a recompute demonstrates is that the arithmetic ran and the selection replays. The truth of the inputs themselves sits outside every tool on this page, and each row below states where that boundary falls for that tool.

Assertions and the tools that recompute them

One table row per assertion. A tool can serve an assertion only as far as its own scope statement allows, so each entry carries that bound inline. The standards mapping behind these assertion names lives in the existing guides: AI evidence auditors can review, the recomputation evidence pack, and closing out a client-trust review period.

Assertion	What recomputes something about it
Existence or occurrence	IPE Integrity Verifier (`verify_ipe_integrity`, art-460): reconciles a declared source-extract hash, row count, and control total against the same facts for the report built from that extract, so the extract you tested is demonstrably the extract the source produced. It reconciles declared values into a confirmed or exception verdict; it never re-derives the extract and cannot see the source system itself. Bank/AR Confirmation Matcher (`match_confirmations`, art-464): joins declared confirmation responses against declared ledger balances and reports exact matches, tolerance matches, mismatches, and the orphans on each side. Both sides are pasted inputs; the node has no reach into a live ledger or a counterparty's records. Report-to-GL Reconciliation (`reconcile_report_to_general_ledger`, art-526): ties each declared account's reported figure to its declared general-ledger figure, the node in this reconciliation set that is checked against an independent book of record rather than a second declared value. It ties figures by account; it is not a transaction matcher and cannot see the GL system.
Completeness	IPE Integrity Verifier (art-460): the row-count and control-total comparisons above are the completeness face of the same check, an extract that dropped rows or totals shows as a count or total discrepancy list. The bound is the same: declared values in, arithmetic out, no re-derivation of the source. Attribute Sampling Plan Generator (`plan_attribute_sample`, art-458): computes the sampling plan from declared confidence and deviation rates and selects items by fixed interval over the declared population hash, so any reviewer can replay the identical sample. The plan is arithmetic over declared parameters; the population is whatever hash and size the caller declares, and the node cannot inventory what the caller left out. Control-Test Evidence Composer (`compose_control_test_evidence`, art-461): reconciles the declared sample against per-item results into a conclusion record: sample coverage, exception count against a declared tolerable-deviation threshold, and the test conclusion. An exception count above zero flags a deficiency candidate; severity classification belongs to a separate reviewer record, never computed here.
Accuracy and valuation	Audit Recalculation Suite (`run_audit_recalc_suite`, art-463): independently recalculates depreciation (straight-line, double-declining-balance, units-of-production), interest accrual, basic and diluted earnings per share, intangible amortization, and prepaid-expense roll-forwards, and diffs each recomputed figure against the stated figure under a declared tolerance gate with an echoed default of flagging any nonzero variance. Diluted EPS uses the simplified form; it does not model if-converted or treasury-stock adjustments for specific convertible instruments. Audit Sampling + MUS (`compute_isa530_audit_sampling_mus`, art-693): monetary-unit-sample sizing, misstatement projection by the tainting method over caller-supplied sampled items, and a Benford first-digit screen with a chi-square deviation flag. It computes the arithmetic of the method; it never selects sample items and never offers a view on sufficiency. The method's source standard and its citation live on the node's own page. Report-to-GL Reconciliation (art-526): nets a declared designed plug before the residual, so an intentional adjustment is never counted as a break, and reports a per-account breaking-account list. Where the requested reporting cadence is finer than the GL schedule's own cadence it refuses to run, and a GL not yet declared closed is a distinct did-not-run outcome rather than a tie-out break.
Cut-off	Journal-Entry Ruleset Screen (`screen_je_ruleset`, art-462): the post-close rule compares each entry's posting date with the declared period-close date, with weekend/holiday-posting flags riding on the same screen. It is a screen over a declared extract; passing it does not close a period. Dates parse loosely, and an entry whose date does not parse is skipped by the date-dependent rules rather than flagged by them.
Classification	Journal-Entry Ruleset Screen (art-462): suspense-account and manual-account postings and unusual user/account pairings surface entries sitting in odd accounts. Which rules are active, what counts as a round number, the suspense list, and the authorized pairing list are all declared policy inputs, and the ruleset version is echoed verbatim so the artifact records which policy vintage produced the flags. Report-to-GL Reconciliation (art-526): the by-account tie-out means a figure parked on the wrong account line surfaces as a per-account break rather than netting away inside a total. The account structure it checks is the declared one; appendix and supplemental-schedule semantics are versioned policy inputs.

How the items get picked

Two nodes produce the sample itself, and both are built so the selection replays instead of being re-argued:

art-458 uses deterministic interval sampling over the declared population hash: no randomness, so the same declared inputs yield the same items for anyone, at any later date.
art-583 seeds an HMAC-DRBG selection from a caller-pasted public randomness beacon pulse combined with an item-manifest hash committed before the pulse round, and emits the full per-draw transcript for offline re-derivation. Committing the manifest first is what makes the selection cherry-pick-proof. The pulse, its round, and its randomness are caller-declared: the node performs hash math only, never fetches a pulse and never verifies a beacon signature. Its own page directs new runs to the version-2 sampler, art-774.

The workflows and what each runs

Chains (presented in the product as workflows) assemble these nodes into one hash-linked run with a composite receipt:

Audit & Assurance

Substantive Procedure Cycle

The journal-entry rule screen (art-462), the recalculation suite (art-463), and the confirmation matcher (art-464) run alongside each other, then the workpaper bundle composer (art-465) assembles the per-area evidence bundle: procedure id, declared population hash, the upstream execution-hash artifacts, the exception list with disposition inputs, and preparer, reviewer, and partner sign-off roles. Estimates, going concern, and materiality stay out of scope by design. The bundle records that a disposition was declared for an exception; whether that disposition resolves it is a reviewer's call, recorded as an approval record with a reason code, never a silent close.

ICFR

ICFR Control-Test Cycle

A single stage through the control-test evidence composer (art-461): population, sample, and per-item results reconcile into a test-conclusion artifact that carries coverage, the exception count against the declared threshold, and the conclusion, with reviewer sign-off recorded as required on every conclusion. A quarter-end bundle produced this way aligns to the sub-certification cascade; it is never a filed submission.

Audit & Assurance

GL Tie-Out Recompute

A single stage through the report-to-GL reconciliation (art-526): every declared account's reported figure ties to its declared GL figure, net of any declared plug, with a per-account verdict and a breaking-account count. The chain refuses to run rather than emit a misleading verdict when the cadence is finer than the GL's own schedule or the GL is not yet declared closed, and the same node doubles as a legacy-to-new core conversion diff when the caller supplies the two exports and labels the sides.

Audit & Assurance

Beacon-Sampled Audit Testing

The sampling ceremony as a hash-anchored composite: commit the item-manifest hash before the pulse round exists (art-583 derives the selection and its transcript), then recompute the sampled items through the recompute node configured for the regime under test, with the shipped configurations running certified-payroll and garnishment recomputes. A sample that did not derive stops the automated path. The pulse and its round are caller-declared throughout: no step fetches a pulse or checks a beacon signature.

Starting points

A first node to open, in task words, for a few questions that come up on almost every engagement:

"Is this report extract the one the source system actually produced?" Start with the IPE Integrity Verifier: declare the source extract's hash, row count, and control total, and the report's, and read the confirmed or exception verdict with its discrepancy list.
"I need a sample plan I can re-run later and get the same items." Start with the Attribute Sampling Plan Generator: the same declared parameters and population hash select the same items for anyone who replays them.
"Does every reported account figure tie to the general ledger before this bundle goes out?" Start with the Report-to-GL Reconciliation, or run it inside the GL Tie-Out Recompute workflow for a composite receipt.
"Which journal entries landed after the period closed?" Start with the Journal-Entry Ruleset Screen over the declared extract and declared close date, and read the post-close flags alongside the rest of the screen.

Packaging and verifying the evidence

The recompute results become reviewable records through the packaging nodes and the offline verifier:

Workpaper Bundle Composer (art-465): assembles the terminal bundle for the substantive-procedure workflow. It mints no new judgment: it does not re-run the upstream kernels, does not recompute the population hash, and records that a release occurred with review required, not that a countersignature happened.
AI-Tool-Usage Workpaper Record (build_ai_workpaper_record, art-380): composes a documentation record from an existing receipt: tool identity, execution hash, declared determinism class, limitations, and a reviewer sign-off statement. It is an evidence format; it claims no endorsement and no sufficiency verdict.
OCG Receipt Verifier (tool 568): paste a receipt and its inclusion proof or signed checkpoint, and the page recomputes the execution_hash, checks the eddsa-jcs-2022 signature, and verifies the Merkle inclusion entirely offline in the browser. A pass verdict speaks to the receipt's integrity, and its scope fence spells out what that does and does not cover.
Verify: recompute the execution_hash of any OpenChainGraph artifact client-side, without uploading it anywhere.

What this page is

An index. Each linked node, workflow, and tool page is the authority for its own scope, and the row lines here paraphrase those pages rather than extending them. Nothing on this page is an audit opinion, and nothing here is legal advice. Assertions are a way of organizing the questions; the tools answer arithmetic and replay questions over declared inputs, and the judgment layered on top stays with the people running the engagement.