wave-002 · new-speaker-regime
New-Speaker Regime v2: executive regime changes detected from call speaker rosters as information events (wave-001 construct-rejection fixes)
Frozen preregistration and current public disposition. The JSON twin remains the authoritative record.
Not evaluatedNot started
Registry record
Current disposition
The specification is frozen, but no primary result has been evaluated.
Execution artifacts
Preregistered hypotheses
Every field below comes from the frozen family record. Expand a card for the complete estimator, exclusions, support gate, leakage boundary, holdout rule, and secondary analyses.
NS-1String matchRegime-change calls carry higher abnormal-return dispersion
Calls that are the first with a new CEO or new CFO as the titled speaker exhibit higher cross-sectional dispersion of +2..+60 session SPY-relative abnormal returns than non-regime calls of the same calendar quarter (a regime change is an information event whose direction is ambiguous).
Full preregistration contract
- Primary feature
- Name
- ns_regime_flag
- Definition
- 1 if is_first_call_new_ceo or is_first_call_new_cfo is true (prior incumbent observed on at least two preceding calls, new name holds the title on this call), else 0; unresolved excluded
- Aggregation
- binary flag per call
- Comparison
- cohort contrast: regime calls vs same-calendar-quarter non-regime calls
- Primary endpoint
- Return field
- abnormal_60_winsor
- Window start session
- 2
- Window end session
- 60
- Benchmark
- cumulative stock return divided by SPY return, winsorized 1/99 percent, per transcript-alpha settled-entry convention
- Price timing
- close of first SPY session strictly after call_date through close of session +60
- Estimator
- preregistered estimand: ratio of regime-call to non-regime-call cross-sectional standard deviation of abnormal_60_winsor, pooled over quarters with quarter demeaning; two-sided p from ticker-cluster bootstrap of the log variance ratio; interpretation — a ratio above 1 supports the information-event reading, below 1 contradicts it
- Universe
- all calls in the CURRENT wave-002 frozen Research Snapshot passing settled liquidity gates, belonging to issuers with at least 3 consecutive prior calls in the snapshot (roster history requirement)
- Exclusions
- below settled liquidity/price gates
- fewer than 3 consecutive prior issuer calls in snapshot
- roster unresolved (gap longer than two calls or no titled CEO/CFO speaker found)
- incomplete +60 session return horizon at frozen price cutoff
- Missingness rule
- unresolved regime flags are excluded, never coerced to 0; the exclusion funnel reports the unresolved count
- Support gate
- minimum 80 regime calls, 50 distinct issuers with a regime event, 6 calendar quarters, and at least 1500 non-regime comparison calls
- Smallest effect size of interest
- absolute log variance ratio greater than 0.0953 (= log 1.10), i.e. dispersion ratio outside 0.909-1.10, exactly symmetric on the log scale
- Power target
- 0.8
- Calibration gate
- Blind to returns
- Yes
- Labeled sample design
- deterministic fixtures: the 40 wave-001 roster fixtures retained, plus new fixtures covering every equivalence-table exercised pair, leading-initial/token-order cases, first-initial fallback hits, FALLBACK COLLISIONS (must emit unresolved), a SAME-CALL AMBIGUOUS-TABLE case (two co-occurring speakers whose names merge under the table — must resolve as distinct people and emit unresolved on any match either participates in), subsidiary/division-title negatives including the bare-title ambiguity case (two unqualified same-rank titles on one call -> unresolved), and roster-noise negatives — each exercised wave-001 false-positive pattern from report 811b204e appears as at least one negative fixture; plus 60 seeded real roster sequences hand-checked by the reviewer, drawn over each sampled issuer's FULL in-snapshot call history
- Metrics
- exactness on ALL retained-plus-new fixtures
- precision and recall of regime detection on the hand-checked full-history sequences
- unresolved rate (denominator: universe-eligible calls)
- Minimum thresholds
- ALL fixtures (retained plus new) exact; precision >= 0.9 and recall >= 0.8 on hand-checked full-history sequences; unresolved rate <= 0.25 with the DENOMINATOR = universe-eligible calls (after the three-prior-call requirement), not all early calls in sampled histories
- Disagreement review
- every detection disagreement listed with its roster evidence; reviewer signs off
- Failure verdict
- construct-rejected
- Multiplicity family
- wave-002-primary
- Holdout rule
- common-wave-002
- Primary p value count
- 1
- Leakage boundary
- Allowed fields
- speaker roster names and stated titles per call
- issuer call sequence index
- registered issuer name and ticker from the frozen Research Snapshot (consumed solely by the issuer-scope title rule and guidance-ACTION subject binding)
- Forbidden fields
- return_fields
- price_fields
- outcome_labels
- future_transcripts
- post_endpoint_metadata
- Identifier handling
- required-and-declared
- Date handling
- stripped
- Future text handling
- forbidden
- Pretraining risk mitigation
- deterministic string matching has no pretraining; declared not-applicable as the mitigation
- Rationale
- Executive names and the issuer call sequence are the construct itself — regime change cannot be detected without them; the registered issuer name/ticker is additionally consumed, solely to prove issuer scope of titles and to bind guidance-sentence subjects, and is declared rather than stripped. No model consumes any of these, so pretraining leakage does not arise. Dates are not needed: sequence order suffices.
- Overlap boundary
- No settled entry uses speaker identity. Distinct from CSL-4 (management answer concentration, wave-001 frozen call-structure-logistics family): CSL-4 scores concentration of answer tokens among incumbent management; NS flags identity turnover and uses no share-of-words or concentration feature.
- Secondary analyses
- Analysis id
- NS-1-S1
- Title
- Within-issuer dispersion contrast (turnover-firm-volatility confound check)
- Definition
- Descriptive, preregistered per reviewer item 3 (Chariot, cst-5cr5.11.3): for each regime call, contrast abnormal_60_winsor dispersion of regime calls against the same issuers' own adjacent non-regime calls (window of 2 calls before and 2 after, excluding the regime call itself); reported as a within-issuer log dispersion ratio. No p-value, no multiplicity cost. Interpretation: if the cross-sectional NS-1 ratio exceeds 1 but the within-issuer ratio does not, the finding reflects which firms change executives, not the regime call as an information event.
NS-2String matchNew-CFO first call paired with guidance withdrawal or widening predicts negative abnormal return
Calls that are a new CFO's first call AND whose prepared remarks state guidance withdrawn or widened earn lower +2..+60 session SPY-relative abnormal returns than all other calls (big-bath / kitchen-sink setup under new financial leadership).
Full preregistration contract
- Primary feature
- Name
- ns_cfo_bigbath_flag
- Definition
- 1 if is_first_call_new_cfo is true AND guidance_action in {withdrawn, widened}, else 0; either component unresolved excludes the call
- Aggregation
- binary flag per call
- Comparison
- cohort contrast: flagged calls vs all other universe calls, quarter demeaned
- Primary endpoint
- Return field
- abnormal_60_winsor
- Window start session
- 2
- Window end session
- 60
- Benchmark
- cumulative stock return divided by SPY return, winsorized 1/99 percent, per transcript-alpha settled-entry convention
- Price timing
- close of first SPY session strictly after call_date through close of session +60
- Estimator
- difference in quarter-demeaned mean abnormal_60_winsor, flagged minus unflagged; PRIMARY p-value: ticker-cluster bootstrap (this is the single p feeding wave-002-primary multiplicity adjustment); Newey-West lag-1 on quarterly differences is reported as a robustness line only and feeds no adjustment
- Universe
- NS-1 universe (same roster-history and liquidity requirements, current wave-002 frozen Research Snapshot) further requiring a completed guidance-language scan (guidance_action != unresolved); none-stated calls are IN the universe as valid unflagged comparison observations
- Exclusions
- any NS-1 exclusion
- guidance_action unresolved (scan failed or contradictory)
- Missingness rule
- unresolved on either component excludes; none-stated guidance is a valid unflagged comparison-cohort state, never an exclusion and never missing
- Support gate
- minimum 25 flagged calls, 20 distinct issuers flagged, 6 calendar quarters; if the flagged support floor fails, the registry-v2 verdict is support-failed (first-class per cst-ifgk) and no market p-value is produced this wave
- Smallest effect size of interest
- 0.02 (two percentage points) quarter-demeaned mean difference in abnormal_60_winsor units
- Power target
- 0.8
- Calibration gate
- Blind to returns
- Yes
- Labeled sample design
- guidance-ACTION grammar frozen first (subject binding, headline-object set, withdrawal exclusions, widening rule, call-level precedence — as in shared_extraction), then evaluated on seeded guidance sentences hand-labeled by the reviewer, drawn over full issuer histories as TWO DECLARED COMPONENTS kept separate in reporting, with counts and seeds PINNED: (i) an unbiased seeded base draw of EXACTLY 200 guidance sentences (recall and miss discovery are measured here), and (ii) an extractor-positive enrichment stratum of UP TO 100 sentences — per-class cap of 50 (seeded draw from that class's predictions when more than 50 exist, every available prediction when 50 or fewer), shortfall NOT backfilled from the other class or anywhere else, so the stratum size is a deterministic function of the extractor's predictions; the call-level unresolved-rate draw is an unstratified sample of EXACTLY 150 universe-eligible calls; every draw's RNG seed derives deterministically as int(sha256(frozen spec_sha256 + ':' + component label), 16) mod 2^32, so no sample can be chosen after seeing classes; EVERY exercised false-positive category in Chariot report 811b204e becomes a mandatory negative fixture class labeled NOT-AN-ACTION at sentence level (call-level none-stated stays reserved, per restored R7, for calls with no guidance-language sentence; the two labels are never interchangeable across units) — including but not limited to 'broadening adoption', 'widening credit spreads', and non-headline 'not providing X guidance' — plus the NS-1 roster calibration inherited for the CFO component
- Metrics
- per-class precision and recall of guidance_action
- confusion matrix
- unresolved rate
- Minimum thresholds
- withdrawn and widened classes each precision >= 0.85 and recall >= 0.7 measured SENTENCE-LEVEL (recall on the unbiased base draw, precision over base + enrichment); every mandatory negative fixture sentence classified not-an-action; unresolved rate <= 0.3 measured CALL-LEVEL over the unstratified call draw; NS-1 roster gate passed
- Disagreement review
- all disagreements listed verbatim with sentence evidence; reviewer signs off
- Failure verdict
- construct-rejected
- Multiplicity family
- wave-002-primary
- Holdout rule
- common-wave-002
- Primary p value count
- 1
- Leakage boundary
- Allowed fields
- NS-1 allowed fields (including registered issuer name/ticker, same declared sole uses)
- prepared-remarks sentences matched by the frozen guidance-language pattern table
- Forbidden fields
- return_fields
- price_fields
- outcome_labels
- future_transcripts
- post_endpoint_metadata
- Identifier handling
- required-and-declared
- Date handling
- stripped
- Future text handling
- forbidden
- Pretraining risk mitigation
- deterministic string matching has no pretraining; declared not-applicable as the mitigation
- Rationale
- As NS-1 for names and sequence. Guidance sentences are consumed by a frozen pattern table, not a model; only sentences matching the table are read.
- Overlap boundary
- Boundary against Temperance's forward-commitment-integrity family (FCI, a wave-002 draft at this writing): FCI registers commitment-disposition transition classes that include withdrawn/widened-like categories graded over all commitment text; NS-2 uses a categorical headline guidance ACTION solely as a conjunction with CFO turnover, computes no disposition or intensity score, and shares no feature values with FCI. Because both wave-002 drafts touch withdrawn/widened categories, the registry change-control rule applies: the overlap is adjudicated by the wave reviewer before either family opens returns. Also distinct from settled evasion-proxy: no Q&A features used.
- Secondary analyses
- Analysis id
- NS-2-S1
- Title
- Flagged vs withdrawn/widened-with-incumbent-CFO contrast (guidance-action-alone confound check)
- Definition
- Descriptive, preregistered per reviewer item 4 (Chariot, cst-5cr5.11.3): quarter-demeaned mean abnormal_60_winsor of flagged calls (new-CFO AND withdrawn/widened) minus calls with guidance_action in {withdrawn, widened} under an incumbent CFO. Isolates the new-CFO interaction from the guidance action alone. No p-value, no multiplicity cost. Interpretation guard: a confirmed NS-2 primary with this contrast near zero would mean the effect is carried by the guidance action, not the CFO conjunction, and the atlas entry must say so.
Shared extraction contract
- Input unit
- issuer sequence
- Input binding
- speaker roster blocks and speaker tags of each call's prepared remarks and Q&A (names and stated titles only), plus prepared-remarks text restricted to guidance-language sentences for NS-2; joined metadata limited to issuer call sequence index AND the issuer's registered name and ticker from the frozen Research Snapshot (consumed solely by the issuer-scope title rule and the guidance-ACTION subject binding; declared, not stripped)
- Output contract
- typed rows per call {call_id, ceo_speaker_norm, cfo_speaker_norm, is_first_call_new_ceo: bool|unresolved, is_first_call_new_cfo: bool|unresolved, guidance_action: maintained|raised|lowered|widened|withdrawn|none-stated|unresolved, evidence_offsets}; name identity resolved by the FROZEN NS-local equivalence artifact extractors/ns_v2_name_equivalence.json, sha256 a3e6900b93c4ec0b7a9cf960af51ac3d4d44862a2cad30604bebb7838055d1f3 (canonicalization order, nickname/formal table covering the eight wave-001 exercised pairs plus Walt/Walter and leading-initial/token-order cases like 'D. James Bidzos', first-initial+surname fallback that EMITS UNRESOLVED ON COLLISION — two plausible distinct people sharing surname+initial never silently merge — and an unsupported-pair rule emitting unresolved); a roster gap longer than two calls emits unresolved, never a guess
- Model or algorithm
- deterministic string matching, no LLM anywhere in this family. (1) ROSTER: roster diff on names resolved per the frozen equivalence artifact (sha a3e6900b93c4ec0b7a9cf960af51ac3d4d44862a2cad30604bebb7838055d1f3), restricted to issuer-scope CEO/CFO holders per the ISSUER-SCOPE TITLE RULE: a CEO/CFO title whose qualifier names a unit other than the issuer (subsidiary, division, segment, group, brand, geography — matched against the issuer's registered name/ticker aliases) is excluded; a bare unqualified CEO/CFO title counts as issuer-scope ONLY if that speaker is the unique holder of an unqualified title of that rank on the call — when two or more speakers carry unqualified same-rank titles, or issuer scope is otherwise unprovable from the declared inputs, the holding emits UNRESOLVED (conservative: the exercised Simon May case — a group president appearing as bare '<President and CEO>' alongside the actual issuer CEO — yields unresolved, never a regime flag). (2) GUIDANCE-ACTION GRAMMAR, frozen: (a) subject binding — the sentence's grammatical subject must be the issuer (registered name/ticker alias), 'we', 'the company', or 'management'; binding is sentence-scoped, no cross-sentence anaphora; (b) headline-object set — overall guidance/outlook, or revenue/sales, earnings/EPS/EBITDA/profit guidance for the current fiscal quarter or current fiscal year; (c) withdrawal exclusions — standing no-guidance practice ('we do not provide quarterly guidance'), non-headline metrics, outer-year targets, 'not yet providing', segment-level guidance, and statements about other companies never count as withdrawn; (d) widening — only when the issuer's own stated guidance RANGE for a headline object is explicitly wider than its prior stated range; adoption language, market conditions (credit spreads), scenario ranges, internal planning assumptions, and a merely reiterated wide range never count as widened; (e) call-level precedence when sentences disagree — explicit headline-action statements rank withdrawn > widened > lowered > raised > maintained; contradictory explicit statements of equal rank emit unresolved. Per wave-001 declared reading R7, retained: none-stated means NO guidance-language sentence was found on the call; guidance sentences present but matching no recognized action emit UNRESOLVED, as does a scan that could not complete or matched contradictorily. ARTIFACT BINDING: before any calibration or scoring run, execution.extractor_artifact_sha256 must hash the complete extractor bundle — the code implementing the issuer-scope title grammar and the guidance-ACTION grammar plus the frozen identity artifact (sha a3e6900b93c4ec0b7a9cf960af51ac3d4d44862a2cad30604bebb7838055d1f3) — as one behavior manifest; the implementations may not elaborate beyond this frozen text (no silent regex widening), and any divergence found at review is a construct change requiring re-review.
- Prompt or code artifact sha256
a3e6900b93c4ec0b7a9cf960af51ac3d4d44862a2cad30604bebb7838055d1f3- Scorer blinded to returns
- Yes
- Scorer blinded to future transcripts
- Yes
Preregistration notes
v2 of the wave-001 new-speaker-regime family, construct-rejected at the wave-001 calibration gate (reviewer Chariot, report sha256 811b204e…, bead cst-5cr5.11.22). Boundary: speaker-identity turnover and its guidance conjunction; no share-of-words, no Q&A features. Reviewer independence: author Magician, independent reviewer Sun (Seneschal reassignment 2026-10-26, Chariot tunnel down; Sun had no role in wave-001 NS work; author never reviews own family). Review-round changes, one line each: (a) frozen nickname/formal-name table + first-token-initial fallback replaces bare normalized-string equality (8 of 13 wave-001 false regime calls were nickname variants); (b) issuer-scope title grammar excludes subsidiary/division officers; (c) guidance-ACTION grammar requires company-subject + guidance-object; (d) calibration draws use full issuer history, not a 6-call window. Known risks and gates: calibration gate as specified per hypothesis, failure verdict construct-rejected; support-failed is a first-class verdict under registry contract v2 (cst-ifgk). Process archaeology: cst-5cr5.11.29. Declared readings restated from the wave-001 spec round (cst-5cr5.11.3) so v2 is self-contained: co-titled seats — multiple simultaneous holders are allowed and a regime flag fires only when a name never previously observed holding that title for the issuer appears (a departure alone never flags); interim/acting holders — per wave-001 declared reading R1, retained: a NEW interim/acting holder's first call IS a regime call (and a subsequent distinct permanent holder's first call flags again as its own transition); returning holders — a previously observed holder resuming the title never flags; gap traversal — gaps of at most two calls are traversed against the last resolved roster, longer gaps emit unresolved; guidance action precedence and none-stated vs unresolved — as frozen in shared_extraction: per R7, none-stated only when no guidance sentence exists; guidance sentences with no recognized action emit unresolved.