OP_CLSI · GUIDEclose intelligence · user guide

Run your first intelligent close.

This guide walks the full loop on live data: run the monitor, teach it with a disposition, run the close, and read commentary where every number traces to its source. Budget twenty minutes the first time; minutes a month after that.

G_00·Before you start

Before you start

Where it lives
Close Intel is a tab suite inside the Forecasting module: CI_01 Monitor · CI_02 Agent memory · CI_03 Expectations · CI_04 Close run · CI_05 Board page.
What it reads
GL actuals at cost-center grain, the active forecast cycle versions, the instrument tape, and the HR roster - all seeded in the demo environment.
What it never does
Compute a number with an LLM. Every figure in every narrative is string-match verified against the deterministic engines.
G_01·The analyst month

Who does what, when.

The guide's spine: three owners, four phases, and the two handoffs that matter — the queue passing to the analyst, and the analyst's teaching passing back to the agent. Dashed cell tops mark ownership transfers.

THE MONTH FLOWS →
Owner ▸ phase
Intra-month
weeks 1–4 · scheduled runs
Teach
minutes per finding
Close run
close week
Board review
day 3–5
Agent · deterministic
CI_01Runs the monitor: 21,585 cells baselined, scored, thresholded; suppression rules applied with attribution.
CI_02Drafts a suppression rule from each noise / known-recurring verdict; files the disposition into memory.
CI_04Computes every bridge: rate/volume/mix, headcount×rate, level + residual. Penny-tie enforced.
Analyst · you
CI_01 · HANDOFFReviews the findings queue — score factors visible; adds comments on GL line × cost center × period.
CI_01–02Dispositions each finding — real / noise / known-recurring + note; activates the drafted rule in Agent memory.
CI_03Keeps expectations current: efficiency path, NIM band, growth ceilings, unit seasonality notes.
CI_05 · HANDOFFReads the board page; clicks citation chips → bridge waterfall → the underlying GL transactions.
Narrator · guarded LLM
CI_01Briefs top findings: classification + criticality, citing specific transactions and prior dispositions.
CI_04Writes per-line, per-LOB, and company commentary from provided figures; guard string-matches every number.
CI_05Every paragraph carries its citations: bridges, findings, dispositions, comments, expectations.
G_02·Step by step

Step by step

01

Run the monitor

CI_01 → Run monitor. About a minute later the findings queue lands, sorted by score. Expand a row: actual vs baseline vs forecast, the score factors, and the GL transactions behind the cell.

02

Judge the top findings

Judge batch briefs the highest-scoring findings: a classification (timing, posting error, real change, known-recurring), a criticality, and a one-paragraph brief citing specific transactions.

03

Teach with a disposition

Pick a finding you recognize - say the annual FDIC true-up. Disposition it known-recurring with a note. A suppression rule drafts itself, scoped to the account, entity, and month.

04

Activate the rule

CI_02 Agent memory shows the drafted rule with your note attached. Toggle it active. Re-run the monitor: the item is suppressed, attributed to your rule - suppressed count 0 → 1.

05

Run the close

CI_04 → Run close for the period. The bridges compute (rate/volume/mix where the tape supports it, headcount×rate for comp, level plus a residual that ties to the penny), then the narrator writes.

06

Read the board page

CI_05 renders the P&L with expandable commentary. Every paragraph carries citation chips - bridge, finding, disposition, comment, expectation. Click through: chip → bridge waterfall → GL transactions.

G_03·The teaching moment

One verdict, three futures.

The disposition is the only decision the product asks of you — and each answer changes what the system does next month.

Is this finding real?
A note is required either way — your reasoning becomes part of the agent's retrieval memory.
real
Stays flagged. Cited by the narrator at close — “as flagged and dispositioned on the 3rd” — and carried as context for related lines.
noise
Drafts a one-time suppression rule anchored to this period. You activate it; the cell won't resurface for this month's pattern.
known-recurring
Drafts an annual rule — “FDIC true-up, every June.” Next year's June run suppresses it automatically, with the rule attributed in the run log.
G_04·Reading the citations

Reading the citations

BRIDGE
A deterministic variance component - click for the waterfall and the calc inputs (balances, rates, source: tape / gl_implied / roster).
FINDING
A monitor finding for the period, with its score decomposition and underlying transactions.
DISPOSITION
Your verdict and note, timestamped - the narrator quotes it: "as flagged and dispositioned on the 3rd."
COMMENT
An intra-month analyst comment on GL line × cost center × period, carried into close context.
EXPECTATION
The band the variance was judged against - target, tolerance, and verdict (within / above / below).
G_05·Good to know

Good to know

Scope the close
A whole-book close runs long; scope to a division or a few entities for demos - the board page renders whatever the run covered.
Dry run
Both the monitor and the close accept a dry-run flag: full deterministic math, no LLM, nothing persisted (monitor) - ideal for a first look.
Guard failures
If a narrative fails the numbers guard twice, a deterministic fallback ships and the row is marked - you will never read an unverified figure.
Demo data
Instrument tape and expectation seeds are synthetic and labeled as such in data (seed_batch, fdic_seed) - the GL history is the environment's own.

Twenty minutes to a smarter close.

Open Close Intel →Back to product