What the Signal Atlas is
A plain-language introduction: what we test, where the hypotheses come from, how the methodology keeps us honest — and why a wave full of nulls is the point.
The project in one paragraph
Academic research occasionally finds persistent, tradable effects hiding in company earnings calls — in what gets said, how it gets said, and the logistics of the call itself. Most such findings are published once and never systematically re-tested; most negative results are never published at all. The Signal Atlas is our attempt to do this properly: we collect transcript-related hypotheses — mainly from the academic literature, plus ideas of our own — write down exactly how each will be measured and judged before looking at any market data, then test them against a corpus of roughly 173,000 earnings-call transcripts joined to the returns of the underlying liquid US stocks. Every verdict is published at a stable URL: discoveries if there are any, and every null, failed construct, and inconclusive result along the way.
The intended reader is someone deciding whether this list of tested hypotheses is worth their time, or worth pointing their own analysis (or their agent) at. Everything on these pages is backed by a machine-readable record you can verify independently.
Who runs this
The Atlas is a project of
Murdarch. The day-to-day
research — drafting and independently reviewing specifications, running the
scoring, verifying one another's artifacts — is done by a
constellation of
AI agents working under written governance, with a human owner setting
direction. The names on the family pages (Hermit,
Chariot, Sun, Star, Temperance, Magician, Hierophant) are those agents —
separate seats running in distinct sessions with assigned roles — and
the cst-… identifiers in the provenance notes name entries
in the constellation's internal governance ledger, where every
preregistration, review, sign-off, and amendment is recorded. We say this
plainly because the methodology is the product. What an outside reader can
audit directly is the public trail: the frozen JSON records, specification
hashes, and artifact digests on every page; the ledger identifiers tell you
where each decision was made.
Where things stand
Current: wave-002. 4 hypothesis families (14 individual hypotheses) are frozen and independently reviewed; no outcome data has been opened. Their verdicts — including any nulls — will be published at these same URLs.
Completed: wave-001
The first registered wave froze 4 hypothesis families (13 individual hypotheses) and has published final verdicts for all of them. The result: 0 promoted discoveries. Every hypothesis was stopped by its own preregistered gates, came back null, or landed inconclusive.
| Verdict | Hypotheses | What it means |
|---|---|---|
| Construct rejected | 5 | A return-blind gate rejected the hypothesis before it entered multiplicity; no market p-value was produced. This is not a market-signal null. |
| Bounded null | 5 | The test ran as frozen and the interval sits inside the preregistered meaningful-effect bounds: effects big enough to matter are ruled out. |
| Inconclusive | 3 | The test did not qualify for promotion, and its interval still admits an effect at or beyond the preregistered meaningful bound; not evidence of absence. |
The construct-rejected row covers two different ways of being stopped by your own gates. AA-1, AA-2, NS-1, NS-2 failed measurement calibration — the construct could not reliably measure what it claimed, and no market data was ever opened for them. CSL-5 passed calibration but failed its preregistered support gate: the corpus contained too few qualifying events to run the frozen test, so no market p-value was produced; the measurement itself was never faulted. Registry-v1 records both under the same verdict because neither enters the multiplicity vector; the committed multiplicity artifact preserves the distinction.
If that reads as failure, consider the alternative. A registry that tested thirteen plausible-sounding signals and reported several winners would deserve your suspicion: with enough analytical freedom, something always clears a threshold. A registry that committed to its methods in advance, ran them, and reported that nothing survived is showing you its error control actually works. The nulls are not the cost of the Atlas — on a first wave they are its product: each one retires a hypothesis that market folklore (or the literature) kept alive, and the machinery that produced them is the same machinery that will make any future discovery worth believing.
The hypothesis families, translated
Atlas pages refer to families by short codes — CSL, CM, NS, AA. Here is what each actually claims and where the idea came from. Provenance is taken from the original preregistration filings (linked from every family page), not reconstructed after the fact.
attribution-asymmetryConstruct rejectedAttribution Asymmetry: self-serving outcome attribution in earnings-call Q&A as a governance signal
What it claims, in plain words. When management credits itself for good outcomes but blames external factors for bad ones, is that a sell signal? Two hypotheses: the level of that asymmetry in a call's Q&A predicts negative returns, and a quarter-over-quarter widening of it predicts negative returns. Unlike the other three families, measurement required an AI model to label each outcome-attribution pair — which is exactly where it failed (see the family page: the calibration gate rejected the construct before any market data was opened).
Where it came from. Our own formalization of a familiar idea — self-serving attribution as a governance-quality signal, long discussed in the management and accounting literature; the frozen filing names the construct in those terms without resting on a specific paper. Floated by Star in the 2026-10-19 brainstorm, preregistered by Magician, independently reviewed by Star (ledger bead cst-5cr5.11.2).
AA-1External-blame asymmetry predicts negative abnormal returnAA-2Quarter-over-quarter widening of attribution asymmetry predicts negative abnormal return
call-mechanicsInconclusiveMixedCall Mechanics: within-turn delivery properties of management Q&A answers as return signals (deterministic, zero LLM cost)
What it claims, in plain words. Does how management talks, within a single answer, carry information? Four hypotheses about delivery: answers that run unusually long relative to the question, rising density of discourse fillers (“you know”, “sort of”), rising density of epistemic hedges (“we believe”, “roughly”), and answers that open by repeating the analyst's question back. Every measurement is a frozen word list or a token count against the company's own trailing norm — again, no AI judgment in any primary feature.
Where it came from. Our own constructs, preregistered by Chariot and independently reviewed by Hermit (ledger bead cst-5cr5.11.11), formalizing common analyst folklore about evasive delivery. The frozen filing situates CM-3 (hedging) against the Loughran–McDonald uncertainty-lexicon tradition from the finance-text literature while measuring a narrower construct (hedge density of management answers only), and preregisters CM-1 as two-sided because the literature does not settle whether long answers are engagement or filibustering.
CM-1Answer elasticity versus issuer norm is associated with abnormal return (two-sided)CM-2Rising discourse-filler density versus issuer norm predicts negative abnormal returnCM-3Rising epistemic-hedge density versus issuer norm predicts negative abnormal returnCM-4Question echo — answers that open by repeating the question predict negative abnormal return
call-structure-logisticsInconclusiveMixedCall Structure & Logistics: scheduling and transcript-shape metadata as return signals (zero/near-zero LLM cost)
What it claims, in plain words. Does the shape of the call itself carry information? Five hypotheses about scheduling and transcript structure: calls held later than the company's own habitual schedule, a shrinking Q&A section relative to prepared remarks, fewer analysts asking questions than usual, answers concentrating in fewer executives, and answers that recycle the prepared script verbatim. All five are computed from transcript metadata and string matching — no AI judgment anywhere in the measurement.
Where it came from. Mostly our own constructs, preregistered by Hermit and independently reviewed by Chariot (ledger bead cst-5cr5.11.5). CSL-1 (scheduling delay) is the exception with an explicit literature anchor: its frozen statement invokes the announcement-timing-delay precedent — the finding that delayed disclosure tends to carry bad news. The other four formalize folk readings of earnings calls (rationed Q&A, answer-hogging, canned answers) that we wanted tested rather than repeated.
CSL-1Scheduling delay versus issuer norm predicts negative abnormal returnCSL-2Q&A-to-prepared token ratio versus issuer norm is positively associated with abnormal returnCSL-3Analyst participation delta versus issuer norm is associated with abnormal return (two-sided)CSL-4Rising management answer concentration versus issuer norm predicts negative abnormal returnCSL-5Scripted spill — verbatim prepared-remarks reuse inside Q&A answers predicts negative abnormal return
new-speaker-regimeNot evaluatedNew-Speaker Regime v2: executive regime changes detected from call speaker rosters as information events (wave-001 construct-rejection fixes)
What it claims, in plain words. Are executive changes, detected purely from who speaks on the call, tradable information events? Two hypotheses: calls that are the first with a new CEO or CFO at the microphone show more spread-out returns (direction ambiguous, so the claim is about dispersion, not direction), and a new CFO's first call combined with guidance being withdrawn or widened predicts negative returns — the classic “kitchen sink” setup under new financial leadership.
Where it came from. Our own idea: floated by Chariot in the town's 2026-10-19 hypothesis brainstorm, formalized and preregistered by Magician, independently reviewed by Chariot, who confirmed the formalization matched the floated intent (ledger bead cst-5cr5.11.3). Detection is deterministic speaker-roster string matching.
NS-1Regime-change calls carry higher abnormal-return dispersionNS-2New-CFO first call paired with guidance withdrawal or widening predicts negative abnormal return
analyst-agenda-shiftNot evaluatedAnalyst Agenda Shift: adverse operating topics introduced and converged on in earnings-call questions
What it claims, in plain words. Analyst Agenda Shift: adverse operating topics introduced and converged on in earnings-call questions
Where it came from. See the frozen filing on the family page.
AAS-1New adverse analyst-question share predicts negative abnormal returnAAS-2Independent analyst convergence on adverse topics predicts negative abnormal returnAAS-3Prepared-remarks versus analyst-question topic-allocation gap predicts negative abnormal returnAAS-4Jump in new adverse topic breadth versus issuer history predicts negative abnormal return
forward-commitment-integrityNot evaluatedForward-Commitment Integrity: how management resolves prior guidance and operating milestones
What it claims, in plain words. Forward-Commitment Integrity: how management resolves prior guidance and operating milestones
Where it came from. See the frozen filing on the family page.
FCI-1Adverse commitment dispositions predict negative abnormal returnFCI-2Loss of commitment specificity predicts negative abnormal returnFCI-3Commitment horizon slippage predicts negative abnormal returnFCI-4Unresolved KPI substitution predicts negative abnormal returnFCI-5Silent drops underperform acknowledged misses with replacement plans
quantitative-specificityNot evaluatedQuantitative Specificity: the density and texture of numbers in management speech
What it claims, in plain words. Quantitative Specificity: the density and texture of numbers in management speech
Where it came from. See the frozen filing on the family page.
QS-1Digit-density delta in prepared remarksQS-2Prepared-remarks range widthQS-3Numeric evasion in Q&A
Before wave-001, four earlier hypotheses (sentiment tone-delta, explicit and implicit analyst-call-reaction constructs, and a Q&A evasion proxy) were tested on the same rail and settled as nulls; registry policy records them as historical entries 1–4, kept with their original verdicts and excluded from wave-001's statistics.
How a hypothesis gets judged
Every family passes through the same pipeline, and every step leaves a public, hash-verifiable record:
- Preregistration before looking. The full specification — feature definition, test statistic, exclusions, direction, smallest effect size of interest — is written and independently reviewed before any outcome data is opened. Reviewer sign-off is recorded; author and reviewer are never the same person.
- Freezing. The accepted specification is content-hashed
(the
spec_sha256on every page) and the wave's family digests are committed into a single wave hash. Changing a frozen spec afterward is impossible to hide: any edit changes the hash, and admitted amendments are published beside the original rather than replacing it. - Construct calibration, blind. Before a feature touches market data, its measurement is tested on transcripts alone — hand-recomputed samples, held-out labeled sets — against preregistered accuracy thresholds. A construct that fails measurement calibration stops as construct-rejected. A calibrated construct that later fails its preregistered support gate also stops before multiplicity and produces no market p-value: registry-v1 collapsed that outcome into construct-rejected, while registry-v2 names it support-failed. Two of wave-001's four families ended at calibration, before any market data was opened.
- Separation of roles. The person who wrote a spec, the operator who runs its scoring, and the verifier who checks the result are different seats. Feature workers and construct reviewers never see prices, returns, or anything after the call date; the statistics pass may join returns only after the specification, candidate snapshot, extractor, calibration verdict, and exclusion funnel are all hash-frozen.
- Bounded-null scoring. A miss is not just "p > 0.05." A bounded-null is an equivalence claim: the estimated interval sits inside the preregistered smallest-effect-size-of-interest bounds, so effects big enough to matter are ruled out. Every other non-promoted result is inconclusive: its interval still admits an effect at or beyond that bound — including wrong-sign point estimates — and is never dressed up as evidence of absence.
- Multiplicity correction. All of a wave's eligible p-values are corrected together (Benjamini–Hochberg, FDR 0.1) over a frozen vector, so testing thirteen hypotheses does not buy thirteen chances at a false discovery. Anything that survives must then repeat on a temporal holdout that stays sealed until the discovery decisions are frozen.
The short version: by the time you read a verdict here, there was no step left where wishful thinking could have changed it.
Don't take our word for it
Every family page links its authoritative JSON twin, which carries the frozen specification, the integrity hashes, and the execution artifacts. The machine index lists every family, digest, and per-hypothesis verdict in one document. Questions and verification requests: hello@stockql.ai.