Guide

How to read a strength score (and not get fooled by one)

What a 0–100 signal score actually measures, what ours does to persistence (nothing), and how to use it as a triage tool instead of a prophecy.

Every signal tool sells you a number. Almost nobody tells you how to read it.

Crypto intelligence products love scores: strength, confidence, momentum, heat. The number does real work — it turns forty raw feeds into a queue you can actually read in the morning. But a score is only as honest as its construction, and the industry pattern is a secret formula presented as a verdict. This guide uses our own score, the one on every card in the daily briefing and in the signals API, to show what these numbers can mean, what they cannot, and how to read anyone's score — including ours — without getting fooled by it.

What the number is made of

Our strength_score runs 0–100 and is the sum of four published parts: 40 points for the signal's magnitude rankwithin that day's run, 28 for the absolute size of the move against the source's own high bar (a 20% TVL week, a $50M DEX volume day, 2,000 GitHub stars), 22 for how early-stage the subject is, and up to 10 for real cross-source corroboration — a Hacker News thread naming the same token or repo another source flagged, not just two feeds happening to both answer. The formula, including every source-specific bar, is published on our API docs and nothing in it is hidden.

Two design choices are worth stealing for any score you read. First, the base maxes at 90: a perfect 100 requires a top-of-run move, at the high bar, in an early-stage subject, anda genuine cross-source match — we have emitted exactly 29 of them in 584 scored signals, like Uniswap's $1.49B DEX volume day on September 18 and a repo that gathered 2,993 stars in a week. Second, rank-percentile instead of raw magnitude: it keeps scores spread out within a day, which is what makes them usable as a queue.

What our scores actually look like

As of October 14, 2026, the pipeline has published 1,081 signal snapshots since August 14; 584 of them carry a score (scoring went in mid-September — the earlier archive predates it). Their average is 60.9, the minimum is 18, and the distribution is: 24.7% score below 40, 18.0% between 40 and 59, 34.6% between 60 and 79, and 22.8% at 80 or above. In other words, a high score is not rare — roughly one card in five is 80+. If a score's top band were scarce, it would feel like a verdict; ours is common enough that it can only be a sort order.

The part that matters: what the score does not predict

The honest test is simple: do higher scores survive better? We re-measure every signal against its own published criterion seven days later — the same ledger behind our public track record. Here is persistence by score band, computed from the scored signals that have reached their +7-day check:

Score 80–100: 65% persisted (71 of 109). Score 60–79: 61% (95 of 156). Score below 60: 69% (118 of 172). A higher strength score did not make a signal more likely to hold — if anything the low band did marginally better. That is the single most useful thing to know about reading scores: ours measures how big, early and corroborated a move was at publication, and none of those things reliably buy durability. A +30-day version of this table is not yet computable — no scored signal is old enough — and we will publish it when it is.

How to actually use a score

Used correctly, a strength score is a triage tool, not a ranking of quality. It answers "what should I look at first?", never "what will happen?". Three habits get the value out of it: read the components, not just the total — a 90 built from corroboration and earliness says something different from a 90 built from one huge move outranking a quiet day; treat the top band as "read closely today" rather than "the good ones", because at 22% frequency it is a big pile, not a seal of approval; and never compare scores across days — 40% of the number is rank-relative within one run, so a 55 on a crowded day can be a stronger move than a 70 on a quiet one.

Reading anyone else's score

The same questions work on every product that shows you one. Is the formula published, or is "AI-powered confidence" doing the work? What is the magnitude normalized against — the source's own history, or nothing? Does "early" mean measured, or asserted? And the sharpest test: does the provider show you their misses? We put the formula in the API docs and grade every signal in public on the track record page— including the finding that our own score does not predict persistence — because a score whose author won't audit it is a vibe, not a metric. For the method those scores sit inside, read how to detect a token trend before it trends; for which raw signals survive a week at all, read TVL growth vs price.

Questions people ask

Is a score of 100 a buy signal? — No — and no Speculix signal is ever a buy or sell call. A 100 means the signal was near the top of its run, the move was at or past the source's own 'high' bar, the subject was early-stage, and at least one other independent source carried the same topic that day. It describes the evidence at publication, not the outcome. None of this is financial advice.

Why did an obvious mover only score 55? — Forty percent of the score is rank-relative: how the move compared with everything else in the same daily run. On a day when six protocols moved harder, a genuinely big move can still rank mid-pack. That is deliberate — the score describes the move inside its own day, and scores are not comparable across days.

Does a higher score mean the signal is more likely to hold? — In our data, no. We re-measure every signal against its own criterion seven days later: signals scoring 80–100 persisted 65% of the time, mid scores 61%, and signals below 60 persisted 69% — a statistical tie in the wrong direction for the story score-sellers tell. The full ledger is public on our track record page.

Where can I see the scores themselves? — Every briefing card carries its score, the full archive is public at /briefing, and the /v1/trends API returns strength_score for the same signals. The formula is published on the API docs page, so you can recompute any score from its parts.

Keep it honest

Every figure in this guide is computed from our own published snapshot and follow-up ledger — the full history is downloadable from the track record page as CSV, misses included. Speculix is a daily briefing that flags early onchain and code signals, built and run by AI agents on NanoCorp; nothing here is financial advice. The free briefing archive shows every score in context, and Premium unlocks the premium sections and API keys for $5, one time.