Best Practices
Success mindset, integration patterns, and an engineering checklist distilled from public benchmarks and real integrations — run Amos correctly and reliably in production.
Success Mindset
MINDSETTeams that get the most out of Amos accept these four principles from day one.
Amos generates no free text — it returns a structured decision with probabilities. Model business problems as choice / score / noul primitives instead of stacking prompts.
calibrated_confidence is temperature-calibrated and tracks true accuracy closely in public benchmarks. Route by thresholds, not by eyeballing whether an answer looks right.
Below threshold, escalate to System 2 or a human — never guess. Let System 1 handle the 95% of deterministic calls and save expensive deep reasoning for the hard few.
Same input, same output, fully replayable. Log every decision so incidents can be reconstructed end to end — a capability generative models cannot offer.
Integration Patterns
PATTERNSFive proven integration patterns; choose by business risk and human cost, and combine freely.
Confidence ≥ 0.85 executes automatically; < 0.85 escalates to human. Simplest to implement; good for cold starts and low-risk scenarios.
Execute in the high band, sample-review the middle band, escalate the low band — balancing human cost against risk.
Fire concurrent requests for batch decisions; each text returns independently and results land straight in the audit table — great for tickets and risk screening.
Amos rules first; low-confidence cases flow automatically to reasoning models like GPT / Claude, with both paths converging on one conclusion.
Serve identical inputs from cache; on 503, degrade to humans or cached conclusions to keep the business running.
Production Checklist
CHECKLISTVerify each item before launch; every line comes from public benchmarks and real integrations.
| Item | Value |
|---|---|
| Text preprocessing | Split inputs beyond the 896-token effective context, then decide segment by segment |
| Candidate design | Use stable, enumerable business actions for candidates; avoid synonymous duplicates |
| Timeout & retry | Set a 12-second timeout; 500 is safe to retry (idempotent), back off exponentially on 429 |
| Threshold calibration | Start at 0.85, then calibrate against your own business samples and adjust by risk level |
| Audit logging | Record text / decision / probabilities / calibrated_confidence for full replay |
Copy this checklist straight into your runbook; it applies to private deployment too.
FAQ
FAQThe five questions we hear most before integration.
| Question | Answer |
|---|---|
| How does Amos relate to generative models like GPT? | Complementary, not a replacement. Amos is System 1: a unique structured decision from a single inference pass; low-confidence cases escalate to System 2 reasoning models. |
| Is calibrated_confidence trustworthy? | Probability outputs are protocol-constrained and reproduce stably across four blind rounds (mean 97.45%); 0.85 is a general starting point — recalibrate with your own samples. |
| Which languages are supported? | Primarily Chinese and English; the public benchmarks honestly list multilingual gaps. Private deployments can keep calibrating on your own corpora. |
| Does the public endpoint rate-limit? | Yes. Back off exponentially on 429; for heavy production traffic, prefer private deployment. |
| Does my data leave the intranet? | The public endpoint processes the text you send; private deployment runs fully offline and data never leaves your intranet. |
Next Steps
NEXT STEPSComplete these three steps in order to bring Amos into your production pipeline.
Make your first call
Run the CURL and Python examples in two steps and see the structured decision and confidence return.
Learn more → 02 / API REFERENCEWire up confidence routing
Learn the request / response fields, error codes, and the 0.85 threshold guidance, then integrate.
Learn more → 03 / SELF-HOSTEDDeploy privately
A single consumer GPU runs it fully offline; data never leaves your intranet and the API matches the public endpoint.
Learn more →