Sidereal Research — Technical Disclosures (annex to ToS)
Status: Published — Beta terms; last updated 2026-07-07. the platform's trust claims mean. It is the marketing-language boundary: anything stronger than this document is overstatement.
1. What "citation integrity" means
Every cited sentence in a report is audited against the source(s) it cites, by an independent language model, at sentence level. The integrity figure reported (e.g., "96.4%") is the fraction of audited sentences the auditor graded as supported, and it is always presented with a Wilson 95% confidence interval reflecting sample size. It is a measured probability from an automated statistical process — not a guarantee that any specific sentence is correct, and not a claim about the truth of the underlying sources.
2. What the evidence marks mean
Reports carry a per-sentence proof record ("proofs sidecar"), re-verifiable offline: for each cited sentence, the source ids relied on and a mechanical entailment class — grounded (specifics found in source), attributed (claim framed to its source), or unverified (specifics not mechanically anchored). The report viewer renders these as evidence highlights. Verbatim quotes are anchored to their source text at research time; quotes that fail anchoring are excluded from drafting evidence.
3. AI contribution
Reports are AI-drafted under operator direction (topic, question, source approval at review gates, drafting configuration) and carry an AI-contribution disclosure in the document footer, including models used and per-section provenance markers. Customers are responsible for the human review their use case requires (ToS §6).
4. Research-source licensing (the License Matrix)
Every discovery backend carries an explicit commercial-use verdict,
enforced mechanically at run time: backends whose terms bar commercial
use are excluded from paid runs, and each run's audit records any
exclusions.
Obligations that survive on permitted backends (e.g., attribution for
MITRE ATT&CK content and Stack Exchange's CC BY-SA material when
quoted) are noted in the matrix; drs7 licenses prints the current
matrix. Paper/article content remains under its own copyright;
reports quote within scholarly-use bounds and cite rather than
reproduce.
5. Model providers and data flow
Research content (source text, drafts) is sent to our language-model inference providers — the drafting provider and an independent audit judge operated by a different provider — under their commercial API terms. Search queries derived from the topic are sent to the discovery backends. No customer identity accompanies research content to any provider. Provider retention is verified (mid-2026): neither provider trains on this content, and both retain it only briefly for abuse monitoring (on the order of days to a month), with Zero-Data-Retention enabled where supported. See the Privacy Policy.
6. Isolation and money handling
Tenant isolation (runs, artifacts, gates, event streams, balances) is enforced at every route and covered by a standing adversarial test suite in CI. Payments are processed by a PCI-compliant payment processor; the platform holds the deliverable's fixed price before a run and returns the full hold if the run fails. The ledger is append-only with idempotent entries.