The most expensive thing about a reconciliation break isn't the break itself. It's the half-hour your most senior ops analyst spends understanding what happened, when the answer is almost always the same as the last twelve times this pattern showed up.

This post walks through a real IBOR/ABOR break we saw last month, the agent's reasoning trace, and how it resolved in under two minutes. Names and identifiers have been changed; everything else is as it happened.

The setup

Mid-sized multi-strategy fund. Investment Book of Record (IBOR) is computed continuously through the trading day. Accounting Book of Record (ABOR) is struck overnight by the fund administrator. They should reconcile to within a tolerance band of 0.5 basis points on NAV.

On the morning in question, the delta on Fund VI came in at 4.2 bps. Three to four hundred basis points over tolerance. Big enough that the ops team wouldn't sign off until it was understood.

The break

At 06:14:22 UTC, MatchStrat surfaced the exception:

{
  "break_id":   "brk_2026_05_14_VI_0142",
  "type":       "ibor_abor_nav_delta",
  "fund":       "Fund VI",
  "ibor_nav":   1_482_910_443.18,
  "abor_nav":   1_482_287_904.62,
  "delta_bps":  4.20,
  "tolerance":  0.50,
  "first_seen": "2026-05-14T06:14:22Z"
}

The break itself is just a number. Useful as a tripwire, useless for resolution. The interesting work happens next.

The agent's first pass

Within four seconds, the agent had decomposed the NAV delta into its constituent layers. It looks something like a waterfall:

positions       : 0.04 bps
fx_translation  : 0.11 bps
cash            : 0.08 bps
fees_accruals   : 0.13 bps
income_receivable: 3.81 bps   ← flagged
other           : 0.03 bps
─────────────────────────
total           : 4.20 bps

One component, income receivable, accounts for nearly all of the variance. Useful. The agent has narrowed the search space from "anywhere in a $1.5bn book" to "approximately $620k of expected income that the two systems disagree about."

Bar chart of the 4.20 basis point NAV delta by component. Income receivable accounts for 3.81 basis points; every other component is under 0.15 basis points.
The 4.20 bps delta, decomposed. Income receivable dominates — everything else is noise inside tolerance.

Memory lookup

The next step is the part generic reconciliation tools don't do. The agent queries the fund's institutional memory: have we seen breaks of this shape before?

"This shape" means: same fund, same break type, same dominant component, and a delta in the same order of magnitude. The query returned eleven candidates from the prior nine months. Of those, eight had been resolved with the same root cause: a corporate action announcement processed by the administrator with a different ex-date than the one the IBOR was using.

Diagram of the memory lookup: the new break is matched by structural similarity against eleven prior breaks, eight of which share the same root cause.
Matched by structural similarity, not exact match: 8 of 11 priors share one root cause.

The memory isn't a static rulebook. It's a graph of breaks, root causes, evidence trails, and the human who closed each one, keyed by structural similarity, not by exact match.

The candidate root cause

The agent surfaced this hypothesis with confidence 0.84 and supporting evidence:

Why this matters A junior analyst would have spent 20–45 minutes piecing this together by hand: opening the corporate actions feed, cross-referencing with admin reports, doing the arithmetic. The agent did it in 12 seconds because it had seen the pattern eight times before. And remembered.

Resolution

The agent's recommendation appeared at 06:15:48 UTC:

  1. Confirm the timing discrepancy on C/A-2026-2841 with the administrator
  2. Apply a one-day reclassification entry on the income accrual side
  3. Expect break to close on tomorrow's run when admin's accrual catches up
  4. Flag the corporate actions feed as needing a timing sync review

The on-call ops analyst, alerted via Slack at 06:16, confirmed the recommendation, accepted the reclassification, and routed the timing question to the admin liaison. Total elapsed time: 86 seconds from detection.

Timeline from detection at 06:14:22 through decomposition, memory lookup, recommendation at 06:15:48, and analyst confirmation at 06:16 — 86 seconds end to end.
Detection to confirmation in 86 seconds. The human stays in the loop — at the end, not the start.

What this is and isn't

This isn't magic, and it isn't replacing humans. The analyst still made the call. The agent didn't "discover" anything new, every fact it used was already in the fund's own systems. What changed is the time-to-context: from 30 minutes of investigation to 90 seconds of confirmation.

Manual investigation by a senior analyst takes about 30 minutes per break.
Before: manual investigation, per break.
With the agent, time to context drops to about 90 seconds of confirmation per break.
After: the analyst confirms, per break.

Multiply that by the 40–80 breaks the team sees in a typical week, and you start to see why we built this the way we did.

If you'd like to see this on your own data, email demo@forestrat.ai. We'll book a 45-minute working session.