The useful result
Route unresolved records with enough evidence while keeping clear records unchanged.
- Route unresolved records with enough evidence while keeping clear records unchanged.
A mixed task often contains easy records and a few that need more attention. Use an available baseline such as Sol to apply explicit rules, then send only unresolved rows and their evidence for review. This lesson teaches a manual routing decision; it does not configure an automatic fallback or promise lower total cost.
Select the named model only if it is available in your interface. API controls and consumer-app controls differ. You can practise the technique on the supplied fictional inputs with an available compatible assistant first. This lesson is documentation-based; examples are expected results, not a claim that we ran the named model.
Before you start
- Three fictional support snippets with a narrow categorization rule.
- Optional access to Sol and Astra; the routing exercise can be done manually.
Review evidence
Documentation-based review · 2026-10-18
Provider documentation checked on 18 October 2026. Examples and checks are original illustrative exercises, not recorded platform tests. Cover artwork is conceptual and does not imply provider sponsorship.
Define rules and uncertainty first
Use categories BILLING and LOGIN. “Charged twice” is billing; “cannot sign in” is login; “account problem” lacks evidence. Give each row an ID. Set an escalation rule based on missing facts, not a self-reported confidence percentage.
Example / expected result
Illustrative example; not a recorded model result.
R1 charged twice → BILLING. R2 cannot sign in → LOGIN. R3 account problem → UNRESOLVED.
Check your result
The ambiguous row has a visible reason for escalation.
Ask the baseline to retain evidence
Request row ID, category, supporting phrase and missing fact. Only unresolved rows should reach another model, with the original rule and snippet included. Do not let the second stage silently alter already accepted rows.
Classify these fictional snippets using BILLING or LOGIN only when explicit evidence supports it. Otherwise return UNRESOLVED and a clarifying question. Preserve row IDs and quote the supporting phrase. Do not invent customer history.Example / expected result
Illustrative example; not a recorded model result.
R3 | UNRESOLVED | “account problem” | Ask whether the issue is a charge or inability to sign in.
Check your result
R1 and R2 retain their source phrases; R3 is not forced into a category.
Review the escalation outcome
Give the unresolved row and rule to Astra or a human reviewer. A stronger model may correctly ask the same question; missing evidence is not a reasoning challenge it can solve by guessing. If you run both models, record actual tokens, retries and accepted outcomes.
Check your result
Escalation ends in a supported category or a retained question, never invented certainty.
When it goes wrong
Two-stage routing can cost more after overhead and retries. Do not rely on model branding or a confidence number as an evidence threshold. This is not a production support deployment or permission to send real customer records to another provider.
Make it your own
Add “I cannot sign in after a double charge.” Expected result: both concerns are explicit; retain a multi-issue review item instead of discarding one to fit a single label.
Useful next resources
Sources and review
Last checked Oct 18, 2026. Features and account limits can change. Verify important results against your original material.
Did this help you complete your task?
Tell us what needs improving. No account or private output required.
Report incorrect information or a broken linkGenius Forges Editorial
Practical AI lessons with checked sources, labelled examples and human verification.
