Anatomy of a rejected signal: a funnel audit of our first 157 candidates

• 157 candidates evaluated; 61 published, 64 scored and withheld, 32 rejected outright.
• Slightly under 40% of detected coordination reached subscribers.
• Rejected and withheld candidates are retained in full, with score and evidence, so the filter can be audited rather than argued about.
• Published scores clustered between 65 and 77; no candidate reached the highest conviction band during the period.
• No performance claim is made: zero measured 24-hour outcomes existed at time of writing.

Why the rejection rate is the honest metric

Any system watching four chains continuously will notice a great deal of activity that looks coordinated. Detection is not the hard part; discarding is. A detector with a permissive filter can report several hundred events a day, and each one will be defensible in isolation. The aggregate is unusable, because the reader has no way to tell which of the three hundred deserved their attention, and the cost of finding out is the whole working day.

This makes the publication rate an unusually informative number. It cannot be inflated the way volume can, and it cannot be curated after the fact the way a track record can. It states plainly how selective the system is willing to be at the point of decision, before anyone knows whether the decision was right.

We therefore publish ours, and will keep publishing it as the engine matures. If it drifts upward materially, that is evidence we relaxed something, and readers should ask why.

The funnel, in order

Every coordinated pattern the engine notices becomes a candidate — a durable record with a score, the evidence behind it, and the result of each gate applied to it. Candidates resolve into one of three terminal states.

Rejected outright (32 of 157). The candidate failed a structural test that no amount of score can compensate for. These are not close calls; the underlying token or the flow itself carried a problem that makes the pattern irrelevant regardless of how strong it appears.

Scored and withheld (64 of 157). The candidate survived the structural tests, received a score, and did not clear the bar for publication — or cleared it but was consolidated into an existing signal rather than emitted as a new one. This is the largest group, and it is where the system spends most of its judgement.

Published (61 of 157). The candidate cleared every gate and was emitted with its reference price and score frozen at the moment of publication, so the record cannot be improved retrospectively.

The gate families

We publish the categories, not the constants. The specific floors, windows and weights are tuned continuously and are the part of the system that took longest to get right; naming them would also tell anyone who wants their own flow to look like a crowd exactly what to imitate.

Evidence sufficiency. Is there genuinely more than one decision here, or is this one participant transacting several times? Repeated activity from a single actor is the most common false pattern in on-chain data and the easiest to mistake for consensus.

Participant independence. Do the wallets involved resolve to separate actors? A cluster of six addresses controlled by two desks is two opinions wearing six hats. Independence is scored separately from raw wallet count for exactly this reason.

Tradability. Is there enough depth in the venue that a normal position could be entered and exited without the exit being the entire move? Large size in a thin market is usually a liquidity event rather than a conviction event.

Contract and structural risk. Does the asset itself carry a defect that makes the flow moot? If it does, the quality of the flow is beside the point.

Freshness. Has the window in which this was actionable already closed? A correct observation delivered late is not a signal, and shipping it anyway trains readers to distrust timestamps.

Nothing is deleted

A rejected candidate keeps its score, its evidence, the gate that stopped it and the reason. This is deliberate and it is the part of the design we would defend hardest. A filter whose failures are not recorded cannot be improved, and a vendor whose failures are not recorded cannot be held to anything.

It also means the funnel numbers in this note are reproducible against our own records rather than assembled for publication. When the outcome sample matures, the same retention lets us ask the more interesting question — whether the things we withheld would have worked — and answer it with evidence instead of an assertion.

Limitations

The period covered is short: the v2.0 engine began publishing roughly one day before this note, and 157 candidates is a small sample from which to generalise about steady-state behaviour. The mix will change as coverage deepens on chains that are currently thin.

More importantly, this note contains no performance claim of any kind. At time of writing there were zero measured 24-hour outcomes for v2.0 signals. Any statement about whether the published 61 were good signals would be fabrication, so none is made here. Per-signal outcomes are recorded automatically and appear on the public track record page as they resolve, wins and losses alike.