OP_CLSIaceanalytics.dev / close-intel

The close that
writes itself.

A continuous monitoring agent watches the GL between closes - baselined by account, cost center, and month-of-year. You teach it what is real in minutes. At month-end, a guarded narrator turns penny-tied variance bridges into board-grade commentary where every number is computed and every judgment cites its source.

21,585
GL cells scored in ~50s
592
penny-tied variance bridges
0
LLM-computed numbers
close-intel.ace / MAY 2026 · CLOSE
● MONITOR · LIVE MAYJUN
NOTICE
RetailFDIC assessmentz +4.892
OpsSoftware renewalz +3.981
TreasuryDeposit betaz -3.174
MortgageFee incomez +2.869
candidates 21,585findings 100
TEACH
KNOWN-RECURRINGRetail · FDIC assessment · score 92
annual FDIC true-up - every June
RULE · annual · month 06 · active
AGENT MEMORYsuppressed 0suppressed 1
NARRATE

Deposits ran $35.9MM favorable - as flagged and dispositioned on the 3rd. Efficiency ratio 57.9%, within the 54-63% band.

[BRIDGE][FINDING][DISPOSITION][EXPECTATION]
GUARD ✓ 10/10 narratives
NOTICETEACHNARRATE
01

The agent notices

Robust month-of-year baselines per GL account × cost center, plus a leg against the active forecast version. Score = statistical surprise × dollar materiality × relevance weight.

02

You teach it

Disposition each finding - real, noise, known-recurring - with a note. Rules draft themselves; you activate. Agent memory stays visible, and month N+1 provably behaves differently.

03

The narrator delivers

Rate/volume/mix, headcount×rate, and level bridges that tie to the penny feed cited commentary - per line, per LOB, and a company summary judged against expectation bands.

04

The guard enforces

Every numeric token in every narrative string-matches a deterministically computed figure, or a deterministic fallback ships instead. The discipline is code, not a prompt.

CI_D1

Notice. Teach. Narrate. Repeat, smarter.

One loop per month. The agent notices intra-month, the analyst teaches it what's real, and at close the narrator writes commentary where every number is computed and every judgment is grounded. The teaching returns to the top of next month's loop - that's the compounding.

CI_01 · MONITOR
Agent notices
Baselines per GL account × cost center × month-of-year, and vs the active forecast version. Score = surprise × dollar materiality × relevance.
deterministic
CI_01 · JUDGE
Findings briefed
Each surviving finding classified - timing, posting error, real change, known-recurring - citing the GL transactions. Numbers are provided, never computed.
guarded LLM
CI_02 · TEACH
Analyst dispositions
Real, noise, or known-recurring - with a note. Noise drafts a suppression rule; every verdict becomes retrieval context. Agent memory is visible.
analyst
CI_04-05 · NARRATE
Close writes itself
Penny-tied variance bridges + expectations bands feed cited, board-grade commentary - “as flagged and dispositioned on the 3rd.”
guarded LLM
rules suppress · dispositions become retrieval context → month N+1 starts smarterMAY → JUN → JUL · suppressed 0 → 1 → …

Blue chips mark the deterministic layer - the math the narrator is never allowed to do for itself. The dashed blue return edge is the product: a disposition in month N provably changes month N+1 behavior.

CI_D2

Zero LLM-computed numbers.

The claim the whole product stands on, drawn as the pipeline that enforces it.

Deterministic engines compute
Baselines, scores, 592 penny-tied bridge components, expectation verdicts - SQL and Python only.
rate/volume/mix · headcount×rate · residual ties to $0.00
figures provided to
The narrator writes
Sonnet drafts per-line, per-LOB, and company commentary - quoting only the exact formatted figures it was handed.
"USE THESE EXACT FIGURES" · one guarded retry
string-matched by
The guard verifies
Every numeric token in every narrative must match a computed figure - or the deterministic fallback ships instead.
close 6693b249 · 10/10 narratives guard_pass = true
OP_CLSI_03·Flow

Notice. Teach. Narrate.

01

Run the monitor

Scheduled or on demand: 21,585 cells baselined and scored; suppression rules applied with attribution; the top findings briefed by a guarded model.

02

Disposition in minutes

Real, noise, or known-recurring - with a note. Drafted rules wait for your activation; every verdict becomes retrieval memory.

03

Run the close

Bridges compute, expectations judge, the narrator writes, the guard verifies. The board page renders commentary with citation chips down to the GL transaction.

Notice everything.
Explain every dollar.

Close Intel gives FP&A a monitor that never sleeps, a memory that compounds, and commentary the board can trust to the penny.