Skip to content

Support Ops Demo

This demo shows a realistic support-agent trace moving through Lerim's compiler so the next support agent starts with the durable constraints instead of re-deriving them. The trace and the extracted records below are real: the output was captured by importing the checked-in example through Lerim's support profile.

1. The completed run

A support agent handled a customer refund request for a failed annual renewal. The interesting signal is not the refund itself; it is the standing policy the run surfaced: a renewal where the gateway captured the charge but activation failed must never be retried, because retrying creates a triple charge.

The trace lives at docs/examples/traces/support-agent-run.jsonl.

2. Import it

lerim trace import docs/examples/traces/support-agent-run.jsonl \
  --source-name support-agent \
  --source-profile support \
  --scope-type domain \
  --scope support-ops

3. What Lerim kept

Five records were extracted. Three are durable constraints that govern future support work, one is a reusable fact about a known bug, and the specific ticket itself is archived as an episode (kept as history, not promoted to standing context).

Constraints (active — these govern future runs):

  • No retry for captured-but-inactive renewals — Renewals where the payment gateway captures the charge but activation fails must not be retried. Only manual activation or customer refund are authorized follow-up actions. Evidence: lines 4, 7.
  • EUR 500 manager approval threshold for refunds — Refunds exceeding EUR 500 require manager approval and cannot be auto-approved. Evidence: line 2.
  • EUR 500+ refund escalation diagnostics requirements — Escalations for EUR 500+ refunds must include gateway timeout log, duplicate capture evidence, and plan-inactive status confirmation. Evidence: line 8.

Fact (active):

  • Known gateway-capture-without-activation bug — Payment gateway times out on activation but captures the charge; retry causes triple charge. Resolution is manual activation or refund only. Evidence: line 4.

Episode (archived — history, not standing context):

  • Support escalation for the specific EUR 640 refund (kept as an episode record, not promoted to a durable constraint).

Inspect them yourself:

lerim context records --source-profile support

4. What the next agent now knows

Before this trace was compiled, a fresh support agent handling the next captured-but-inactive renewal would most likely retry the charge (causing a triple charge) or escalate without the required diagnostics. After compilation, the standing constraints above are available before work begins.

5. The improvement loop

Lerim compiles cited context and eval-ready signal; it does not train models. Model specialization is a separate business (Nablo, https://nablo.io) and is independent of Lerim.

Notes

  • Replace the checked-in example with your own cleaned support-agent source session for real evaluation.
  • Do not put private customer datasets or converter outputs under public docs. Use customer-owned storage for raw traces and commit only small sanitized examples when a public example is useful.
  • If the trace contains customer PII or secrets, run a cleaner or the defense-in-depth redaction helper before import. Lerim extraction is selective, but it is not a privacy firewall.