AN APPLIED AI LAB INSIDE SWARN
Evidence before an answer.
A withdrawal record, a campaign export and a vendor acknowledgement can tell different stories. This investigator helps a reviewer explain the records, preserve what remains unknown and choose the next diagnostic step.
This page plays back recorded fictional cases in your browser. It does not contact SWARN, a vendor or Claude, and it cannot change consent, send messages or complete a privacy request.
Recorded software control · Original synthetic sources · Scripted responses · Claude quality not measured
01 · Acknowledgement found — assessment ready for a reviewer
Cedar Retail, authored case clear-records, subject DEMO-017. A withdrawal is recorded, followed by a vendor acknowledgement; no later campaign inclusion appears in the available window. Variant 1 has its own source window and identifiers.
What the records show
The source records a withdrawal for DEMO-017 at 2026-10-15T10:04:00+00:00.
The vendor source reports an acknowledgement for DEMO-017 at 2026-10-15T10:08:00+00:00.
What remains unknown
No missing item within this bounded fixture window. Independent downstream propagation remains unverified.
A reviewer’s next step
Inspect the named delivery attempt and verify its subject, timing and acknowledgement with the reviewer.
Limits
- Authored synthetic sources and a bounded fixture window.
- No independent downstream propagation or message-delivery check; no business action or legal certification.
Compare the software paths
These are recorded fixture trials. A scripted result can verify the plumbing, but cannot tell us which Claude approach is more accurate. Zero fixture cost is not a model price benchmark.
Deterministic rules
assessment ready
1.56 seconds · $0 fixture cost
One prepared call
assessment ready
5.41 seconds · $0 fixture cost
Bounded MCP loop
assessment ready
1.48 seconds · $0 fixture cost
Inspect the recording’s provenance
SWARN source commit: 02248fe7f1946f48250aea3e9fb0f57bdb255d82
Original fixture revision: 85ca9abd4674d4353d936580bc0ad3ca531975b3f61d94547105d545361af7c8
Dataset content hash: 00d8282bf239dae88cd89db063493668d41634e2f23b6ac5d85879ae98866185
Software consistency checked with automated tests and Codex review. No independent human quality evaluation has been completed.
WHAT THE COMPARISON WILL NEED
The simplest useful approach should win.
The local project contains 12 development and 24 held-out synthetic cases from separate families. Rules, one prepared call and the bounded tool loop use the same source universe. The model comparison still needs funded live trials and independent human grading.
Pending grades stay in the denominator. Mechanical citation checks do not prove a claim is supported. Failure, abstention, reviewer corrections, elapsed time and unknown costs remain visible. The result may favor rules, a model approach or insufficient evidence.
Continue the SWARN story →