Experiments
Where we try new things safely. Each experiment runs behind a flag that's off by default — the variant is testable per-request (append ?exp=<id> or send an X-Experiments header) without changing what agents or prod see. We score each against a ground-truth metric; winners graduate to the contract, the rest get killed.
Expose partner compliance to the agent API
partner-compliance-api- Hypothesis
- Adding compliance (licenses, KYC, Travel Rule, currencies) to GET /api/partners lets an agent answer 'which anchors are licensed / Travel-Rule compliant / support MXN' from the API alone — closing the profile-only gap so Raven can use it.
- Metric (ground truth)
- Golden compliance questions: the variant response carries the answer (e.g. Yellow Card + Bitso flagged travelRule=true, licenses present); baseline can't answer at all.
Default: off (not exposed)Try it:
?exp=partner-compliance-apisince 2026-07-07Expose on-chain 'live on Stellar' proof to the agent API
partner-onchain-live- Hypothesis
- Adding domain-matched on-chain reality (each anchor's OWN issued assets' holders, payment count, stellar.expert rating) to GET /api/partners gives agents the git-free trust signal closed-source anchors otherwise lack — distinguishing a live issuer (Zeam ZARZ ~181k holders, Anclap PEN ~42k) from a barely-used one (AUDD NZDSC 5 holders / 0 payments).
- Metric (ground truth)
- Ground truth: the variant carries onchain[] for domain-verified issuers (Etherfuse CETES, MYKOBO's OWN EURC not Circle's, Anclap PEN) with holders/payments > 0; baseline carries no onchain field at all. No misattribution (an anchor merely USING USDC has no USDC onchain entry).
Default: off (not exposed)Try it:
?exp=partner-onchain-livesince 2026-07-07Quality-as-product: sell verified data quality, Scale-AI style
scale-model-quality-products- Hypothesis
- The engine system's measurements (recall floors, data-truth cross-checks vs SCF/DeepWiki, demand-side OK rate, corpus health) are themselves the product: a public /quality scoreboard + a monthly per-consumer quality report + a DATA_SLA.md turn 'our data is good' from a claim into a verifiable contract — the trust layer that makes stellarlight the default data dependency for Raven and future agents.
- Metric (ground truth)
- Ground truth: (1) every scoreboard number traces to a committed engine artifact (no hand-set values); (2) a cold outsider can verify one claim end-to-end from the page alone; (3) first consumer report answers 'what changed for YOUR queries this month' from Engine D data. Adoption signal: Tyler/Raven cites or links the scoreboard.
Default: off (not exposed)Try it:
?exp=scale-model-quality-productssince 2026-07-10Registry: src/lib/experiments.ts · scored by scripts/experiment-eval.ts. Graduating an experiment flips its default on and adds it to the OpenAPI contract.