Reconciliation
Agents
Six agents watch this module. Each one explains, rates or routes — and none of them can change a difference, certify a balance, remove a control signal or decide that an evidence gap does not exist. Those are all decided by rule before the agent is called.
Replayed proposals
Agents in this module
7
2 model tier(s)
Calls this run
7
Taken from the audit trail
Mode
Scripted
Fixture-backed and deterministic
Figures an agent originated
0
Enforced by the numeric guardrail
The roster and its boundaries
What each agent does, what it cannot do, and the code that makes that true
| Agent | Does | Never does | Bounded by | Calls |
|---|---|---|---|---|
claude-sonnet-5 (replayed) | Proposes the likely cause of a break, the action that resolves it, and who it routes to. | Certifies a reconciliation or changes a difference. | reconciliation/accounts.ts computes ledger against support | 1 0 ms |
claude-opus-5 (replayed) | Gives an opinion on whether the support behind each reconciliation can be relied on, and what would have to change for it to be. | Decides whether a gap exists. Signatures, dates, sources and certification history are tested by rule before it is called. | reconciliation/evidence.ts — gaps are facts about the record | 1 0 ms |
claude-sonnet-5 (replayed) | Rates the risk in each accrual — rolled forward, estimated rather than measured, or not reversing — and recommends what to do about it. | Changes an accrual, its basis or the schedule total that must tie to the ledger. | reconciliation/accruals.ts raises every flag before the agent triages | 1 0 ms |
claude-sonnet-5 (replayed) | Explains what each cut-off error does to the reported period and restates the correcting entry in words. | Decides whether a transaction fell in the right period, or changes the direction of the effect. | reconciliation/accruals.ts — testCutoff() computes the entry and the profit effect | 1 0 ms |
claude-sonnet-5 (replayed) | Judges how serious a flagged entry is in the round and recommends an action. | Removes a control signal. The detectors are deterministic rules over the population. | reconciliation/journals.ts raises every signal before triage | 1 0 ms |
claude-opus-5 (replayed) | Briefs on the shape of the journal population as a whole — repeated account pairings, preparer concentration and digit distribution. | Claims a statistical result the population cannot support. Where a test is not meaningful it must say so. | reconciliation/anomalies.ts — the Benford test carries an is_meaningful flag | 1 0 ms |
claude-opus-5 (replayed) | Briefs on the reconciliation position across the group — what is certified, where differences are ageing, and whether the inventory still ties to the statements. | Certifies anything, or reports the inventory as tying when it does not. The tie is an arithmetic check over the entity ledgers. | reconciliation/accounts.ts — rollUp by account proves the entity ledgers sum to the group trial balance | 1 0 ms |
Workhorse
The default. Structured proposals with a rationale.
claude-sonnet-5R1R3R4R5
Deep reasoning
Judgement-heavy work where the cost of a wrong call is high.
claude-opus-5R2R6R7
What each agent did this run
Newest last, straight from the audit trail
- #1R1 Reconciliation Break Agent9:14:26 amTriaged 2 reconciliation break(s)
claude-sonnet-5 (replayed)0 ms - #2R2 Evidence Sufficiency Agent9:14:26 amOpined on the evidence behind 10 reconciliation(s)
claude-opus-5 (replayed)0 ms - #3R3 Accrual Quality Agent9:14:26 amAssessed 4 flagged accrual(s)
claude-sonnet-5 (replayed)0 ms - #4R4 Cut-off Impact Agent9:14:26 amExplained 2 cut-off error(s)
claude-sonnet-5 (replayed)0 ms - #5R5 Journal Review Agent9:14:26 amTriaged 6 flagged journal entries
claude-sonnet-5 (replayed)0 ms - #6R6 Population Anomaly Agent9:14:26 amBriefed on the shape of the journal population
claude-opus-5 (replayed)0 ms - #7R7 Reconciliation Control Agent9:14:26 amBriefed on the reconciliation position across the group
claude-opus-5 (replayed)0 ms