roadmap
what is built, what is next, what never will be
Nothing on this page is a promise with a date attached. Three of the four sections describe the system as it is right now and are read from it; the fourth is a set of decisions, which is the only kind of future statement worth publishing.
running
read from /api/status- — measuring since 2026-08-26 16:00:00 — 22 days: 13,592 tokens, 45,619 measurements, on the application's own clock every 15 minutes
- — two sampling frames kept apart: 356 found by the promotion feed, 13,089 by a completed bonding curve
- — a deterministic observation pipeline: 10 detectors with a source, 1,375 observations, 1,426 anomalies, no model anywhere in it
- — 6 questions, each with its own window, horizon (1h, 2h, 6h, 12h) and its own comparison
- — a critic with 10 design checks, and the gate that stops an unearned finding becoming a claim
- — a public page for every question and every result, rejections included
in beta
- — critic agent: a model reads what the deterministic checks returned and may make the verdict harsher. It can never make it lighter, and an objection citing a number that is not in the result is discarded.
- — observer agent: a model puts an anomaly a detector already found into a sentence. It cannot create, suppress or rescore one.
waiting on measurement, not on code
These are not unfinished. They are questions that have been asked correctly and cannot be answered yet, which is a different thing and is worth being able to tell apart.
- — every result so far is INCONCLUSIVE because the groups are too small. Thirty measurements a side is the floor for believing a difference, and that arrives by waiting, not by tuning a threshold.
- — the 12h questions fill last: a token has to be watched for a whole window and then a whole horizon before it can answer one.
- — the two questions that need holder counts have no rows at all. A public node cannot count holders, and estimating one would be inventing a measurement.
decided against
- — trading, signing, holding. No private key exists in this system and no code path constructs a transaction. This is not a feature waiting for its turn.
- — hypotheses written by a model. A question comes from a template so that nothing reads the data before deciding what to claim about it. A falsification rule invented after the result is not a falsification rule.
- — a model computing the statistics. They have one right answer. A model there would trade a guarantee for a sentence.
- — price predictions, targets, advice. Nothing here has ever tested a claim of that shape, so the system is not allowed to make one — the check is mechanical, not a matter of tone.