How Starcode works
The rule everything is built around
Speculation is never presented as established fact. That is not a matter of careful wording — wording drifts, and a confident sentence reads the same whether or not anything backs it. So it is enforced structurally: every statement the platform makes is stored with an explicit type, and the rules for each type are checked in code before anything reaches you.
A claim typed as verified history cannot exist without a citation. If the model produces one anyway, it is downgraded to a hypothesis and the downgrade is recorded. A calculation must name the engine and algorithm that produced it. Confidence for AI hypotheses is clamped, so a model cannot talk its way into certainty about an interpretive question.
The eight kinds of statement
Source text
Quoted verbatim from what you submitted. No interpretation added.
Verified history
Attested in the historical record and citable. Requires a real citation to exist at all.
Calculation
Computed by the ephemeris and calendar engines. Reproducible, and reported with its accuracy.
Text analysis
An observation about the text itself — its language, structure or vocabulary.
Tradition holds
What a religious or cultural tradition has understood this to mean. Reported, not endorsed.
Scholars argue
A position argued in academic literature. Where scholars disagree, the disagreement is reported.
AI hypothesis
Model-generated conjecture. Not evidence. Confidence is capped by design.
Unresolved
Genuinely undetermined on the available evidence.
The first three are evidence. The rest are commentary, however plausible. Reports separate them, and you can filter to evidence alone.
What actually runs
Eight specialist steps run in order: language and translation, source identification, entity extraction, calendar conversion, astronomical correlation, historical context, interpretation, and an evidence audit that checks the other seven.
The calendar and astronomy steps consult no language model at all. Their output is arithmetic — Meeus' algorithms for eclipses and lunar phases, the JPL Keplerian approximation for planets, Rata Die day-numbering for calendars — and routing that through a model would add nothing but a way for it to be wrong.
Every report carries a trace: which step produced each finding, on which model, how long it took, and its own account of how it reached its conclusion.
What this cannot do
Stated plainly, so nothing here is mistaken for an oversight.
- It does not compute eclipse visibility from a named place. It establishes that an eclipse occurred, when, and of what type. For ancient dates the uncertainty in the Earth's rotation (ΔT) is tens of minutes, which moves the ground track by tens of degrees of longitude. Any tool showing you a precise ancient eclipse path is overselling its precision.
- Planetary positions are approximations. Accurate to roughly 0.3–0.6° over five millennia. For a slow pair like Jupiter and Saturn that translates into about two weeks of uncertainty in the date of a conjunction — which each result tells you.
- It does not adjudicate religious or interpretive questions. It reports what traditions have held and what scholars argue, attributed and sourced. It does not rule on which reading is correct.
- It makes no claims about the future. Prophetic material is described and interpretations are attributed, but a prediction is never presented as a fact about what will happen.
- The reference corpus is a demonstration set. A few dozen curated works. If your text is not in it, that says nothing about your text.
- Model-supplied citations are marked unverified. Language models invent references that look real. Anything not drawn from the curated corpus is labelled, and should be confirmed before you cite it.
Privacy
You can analyse a text without an account. Anonymous analyses are reachable only via a token stored in your browser — if you clear your browser storage, that analysis genuinely becomes unreachable, because there is no account to tie it to. That is the price of not requiring a signup, and it is worth knowing before you rely on it.