The 5 Ws
Checks whether the narrative establishes who, what, when, where, and why—with enough clarity to act on.
Coverage
Early prototype · v0.2
Complete is not the same as useful.
SAR Chasm reviews draft SAR narratives for the gaps that survive a checklist—because “the activity was suspicious” is not an explanation.
Built for people who write, review, and rely on SAR narratives.
Needs attention
The narrative identifies the activity, but the basis for suspicion is not fully supported.
Who, what, when, where, and why are present. The method used to move funds should be stated more directly.
“Structuring” is asserted, but the narrative does not connect the transaction pattern to the relevant threshold.
The summary reports $48,750. Itemized transactions total $46,750. Verify the discrepancy before filing.
The narrative is concise and organized chronologically. No material repetition detected.
Illustrative output · not an actual case
Inside the prototype
The narrative and the review, in one working view.
No decorative dashboard. The interface keeps the draft, safety boundary, and scorecard in the same line of sight.
Current prototype · v0.2
The application runs locally with fabricated narratives only. v0.2 adds dark mode, pass/flag totals, copyable reviews, clearer input guidance, a keyboard shortcut, and an explicit 20,000-character boundary.
What it reviews
Five lenses. One narrative that holds together.
Not proofreading. Not a filing decision. A repeatable rubric for the distinctions an experienced reviewer already makes.
Typology language
Tests whether labels such as structuring, funnel activity, or money laundering are used accurately and supported by the facts.
Precision
Specificity & support
Finds vague claims, missing amounts, weak behavioral detail, and profile comparisons that never establish the customer’s baseline.
Substantiation
Internal consistency
Cross-checks chronology, arithmetic, parties, transaction totals, and statements that conflict elsewhere in the narrative.
Integrity
Length & density
Flags bloated passages and repetition without treating brevity as the goal. The aim is a narrative that is efficient, complete, and usable.
Signal
From the test set
Different problem. Different flag.
Useful QA separates a material weakness from everything the narrative already does well.
The transaction evidence is strong. The broader customer-profile conclusion is not. SAR Chasm flags the unsupported leap without manufacturing problems elsewhere.
The narrative is detailed and internally consistent. Its CTR-avoidance theory is still wrong for the electronic activity described, so the typology language gets the flag.
Fabricated demonstration narratives only. No real customer, account, investigation, or SAR information.
20numbered scenarios
3Test 20 variants
The final v0.2 regression round passed every scenario, including crypto flows, unsupported attribution, arithmetic mismatches, clean controls, and adversarial text embedded inside customer-provided material.
How it works
A second set of eyes, before final review.
- 01
Paste a draft
Start with the narrative as written. No case-system integration required for the prototype.
- 02
Run the review
SAR Chasm evaluates the text across the five QA categories and tests factual relationships within it.
- 03
Review the findings
See passes, items needing attention, and specific flags—with a reason attached to each one.
- 04
Apply judgment
See what rule fired, disagree with it when warranted, and change the draft—or leave well enough alone.
Technically complete
The facts are in there.
Required elements present
Activity described
Conclusion stated
The
chasm
chasm
Operationally useful
The logic is clear.
Facts easy to locate
Typology supported
Amounts and chronology reconcile
Why it exists
The filing is not the finish line.
A narrative can contain every required element and still leave the next person doing unnecessary reconstruction. The dates may be there but hard to follow. A typology may be named but not demonstrated. The total may be stated but not reconcile.
SAR Chasm is being built to find that gap while there is still time to close it—and to recognize when the narrative is fine as written.
So what?
The prototype is local. The point is institutional.
Today, SAR Chasm is deliberately limited to fabricated narratives on localhost. That proves the review method without pretending the hard deployment questions have already been solved.
The practical destination is different: a quality-review capability deployed inside a financial institution’s approved environment, working with its own EDP, access controls, retention rules, and model-governance decisions. Not a public website where investigators paste confidential casework.
01 / Draft
Investigator prepares the narrative
Inside the institution
→
02 / Review
SAR Chasm applies the QA rubric
Inside the approved environment
→
03 / Judgment
A human resolves the findings
Before final review or filing
The value is not “AI writes the SAR.”
It is making experienced QA judgment more consistent, explainable, and difficult to skip—without moving the filing decision away from the people accountable for it.
Built with boundaries
Assistance, not adjudication.
It is
A structured QA pass for draft narratives, designed to make omissions and inconsistencies easier to see.
It is not
A filing decision, legal conclusion, compliance certification, or an excuse to upload confidential SAR data because the interface looks nice.
Right now
A localhost-only demonstration for fabricated narratives. No accounts, uploads, saved history, filing integration, or production support.
Read the work
View on GitHub ↗
The rubric, privacy model, tests, and limitations are public.
Private beta
Help test the space between complete and useful.
SAR Chasm is being tested against fabricated narratives—some obvious, some subtle, and some clean enough to test whether it can leave well enough alone. Practitioner beta access will open in a limited group.
Follow development ↗
Beta access details are coming soon.