wave-001 · attribution-asymmetry
Attribution Asymmetry: self-serving outcome attribution in earnings-call Q&A as a governance signal
Frozen preregistration and current public disposition. The JSON twin remains the authoritative record.
Construct rejectedCalibration failed
Registry record
Current disposition
The calibration gate rejected the construct. No discovery statistics were opened for this family.
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.
AA-1Llm judgmentExternal-blame asymmetry predicts negative abnormal return
Calls whose management Q&A answers show high attribution asymmetry — external-factor share of negative-outcome attributions minus external-factor share of positive-outcome attributions — earn lower +2..+60 session SPY-relative abnormal returns than low-asymmetry calls (self-serving attribution as a governance quality signal).
Full preregistration contract
- Primary feature
- Name
- aa_asymmetry
- Definition
- share of negative-outcome attribution rows with locus external-factor minus share of positive-outcome attribution rows with locus external-factor; unresolved rows excluded from both numerator and denominator
- Aggregation
- one scalar per call from all extracted attribution rows in the Q&A section
- Comparison
- cross-sectional level within calendar quarter, quintile sort high-minus-low, matching settled-entry portfolio convention
- 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
- quarterly high-minus-low quintile spread, Newey-West lag-1 on quarterly spreads, ticker-cluster bootstrap as in settled entries 1-4
- Universe
- all calls in the wave-001 discovery Research Snapshot passing the settled liquidity gates (min pre-call close, min prior-60-session median dollar volume) and splitter verdict ok
- Exclusions
- splitter verdict not ok
- fewer than 10 management answer-turns in Q&A
- fewer than 5 extracted attribution rows with resolved locus
- fewer than 2 resolved attribution rows in EITHER valence class (positive or negative) — aa_asymmetry is a difference of two per-valence shares and is undefined or degenerate on single-valence calls; k=2 chosen and frozen here, never decided at scoring time
- below settled liquidity/price gates
- incomplete +60 session return horizon at frozen price cutoff
- Missingness rule
- calls with fewer than 5 resolved attribution rows, or fewer than 2 resolved rows in either valence class, are excluded (support), not imputed; unresolved locus rows never enter the feature
- Support gate
- minimum 400 scored calls, 150 distinct issuers, 6 calendar quarters (every scored call has both valence classes populated by construction — the per-valence k=2 exclusion above — so no populated-share clause is needed; the exclusion funnel reports the single-valence exclusion count)
- Smallest effect size of interest
- 0.01 (one percentage point) quarterly high-minus-low spread in abnormal_60_winsor units
- Power target
- 0.8
- Calibration gate
- Blind to returns
- Yes
- Labeled sample design
- 150 management answer-turns drawn deterministically (seeded) from discovery-window transcripts, stratified by sector, hand-labeled for valence and locus by a seat other than the owner before any scoring run; minimum 20 labeled rows per locus class — if the seeded draw yields fewer (external-factor on positive outcomes is structurally rare), the draw is extended deterministically (same seed, next strata in order) until the floor is met, keeping blind-to-returns
- Metrics
- per-class precision and recall for valence and locus
- confusion matrix
- unresolved rate
- Minimum thresholds
- locus precision and recall both >= 0.75 per class; valence precision and recall both >= 0.85 per class; unresolved rate <= 0.35
- Disagreement review
- all labeled disagreements listed verbatim in the calibration artifact with class support counts; reviewer signs off on the disagreement list
- Failure verdict
- construct-rejected
- Multiplicity family
- wave-001-primary
- Holdout rule
- common-wave-001
- Primary p value count
- 1
- Leakage boundary
- Allowed fields
- Q&A answer-turn text with the preceding analyst question turn
- speaker role tag (management vs analyst)
- Forbidden fields
- return_fields
- price_fields
- outcome_labels
- future_transcripts
- post_endpoint_metadata
- Identifier handling
- stripped
- Date handling
- stripped
- Future text handling
- forbidden
- Pretraining risk mitigation
- masking of identifiers and dates plus local-context restriction (one answer-turn at a time); residual risk that a famous event is recognizable from context declared irreducible and noted per entry
- Rationale
- attribution locus is judgeable from local sentence context alone; the analyst question is needed to resolve which outcome an answer attributes; nothing else is required. Stripping detail: company names, tickers, person names, absolute dates and fiscal-period identifiers are masked before the scoring prompt; relative references (last quarter, this year) retained.
- Overlap boundary
- Distinct from settled evasion-proxy (entry 4): evasion scored non-answering; AA scores the locus of answered attributions and uses no evasion features. Distinct from tone-delta (entry 1): valence here only partitions attribution rows, no net-tone aggregate is computed or reused.
- Secondary analyses
- None preregistered
AA-2Llm judgmentQuarter-over-quarter widening of attribution asymmetry predicts negative abnormal return
Calls whose aa_asymmetry rose versus the same issuer's previous adjacent quarterly call earn lower +2..+60 session SPY-relative abnormal returns than calls where it fell (deteriorating attribution discipline precedes bad outcomes).
Full preregistration contract
- Primary feature
- Name
- aa_asymmetry_qoq
- Definition
- aa_asymmetry(current call) minus aa_asymmetry(previous adjacent quarterly call of same issuer)
- Aggregation
- one scalar per adjacent pair; both calls must individually pass AA-1 exclusions
- Comparison
- cross-sectional level within calendar quarter, quintile sort high-minus-low
- 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
- quarterly high-minus-low quintile spread, Newey-West lag-1 on quarterly spreads, ticker-cluster bootstrap as in settled entries 1-4
- Universe
- AA-1 universe restricted to issuers with an adjacent previous quarterly call also passing AA-1 exclusions, per the settled tone-delta pairing rules (adjacent fiscal identity, quarterly call gap)
- Exclusions
- either call of the pair fails any AA-1 exclusion
- missing previous transcript
- nonadjacent fiscal identity or nonquarterly call gap per settled pairing rules
- Missingness rule
- pairs are complete-case only; no imputation of either leg
- Support gate
- minimum 300 scored pairs, 120 distinct issuers, 6 calendar quarters
- Smallest effect size of interest
- 0.01 (one percentage point) quarterly high-minus-low spread in abnormal_60_winsor units
- Power target
- 0.8
- Calibration gate
- Blind to returns
- Yes
- Labeled sample design
- inherits the AA-1 calibration artifact; no separate labeling — the construct is identical, only differenced
- Metrics
- AA-1 calibration metrics
- QoQ differencing exactness check on 20 deterministic fixture pairs
- Minimum thresholds
- AA-1 thresholds pass; differencing fixture check exact on 20/20
- Disagreement review
- inherited from AA-1
- Failure verdict
- construct-rejected
- Multiplicity family
- wave-001-primary
- Holdout rule
- common-wave-001
- Primary p value count
- 1
- Leakage boundary
- Allowed fields
- AA-1 allowed fields for each leg of the pair
- fiscal-quarter sequence index used only for pairing, never shown to the scorer
- Forbidden fields
- return_fields
- price_fields
- outcome_labels
- future_transcripts
- post_endpoint_metadata
- Identifier handling
- stripped
- Date handling
- stripped
- Future text handling
- forbidden
- Pretraining risk mitigation
- as AA-1
- Rationale
- the difference is computed from two independently scored calls; the scorer needs nothing beyond AA-1's fields. Stripping is as in AA-1; pairing is done on metadata outside the scorer. The current call's scorer never sees the previous call's text and vice versa.
- Overlap boundary
- Distinct from settled tone-delta (entry 1) despite sharing the QoQ-differencing shape: the differenced quantity is attribution locus share, not lexicon tone; no tone dictionary is used anywhere in the family.
- Secondary analyses
- None preregistered
Shared extraction contract
- Input unit
- single call
- Input binding
- Q&A section only, as produced by the transcript-alpha splitter with verdict ok; joined metadata limited to speaker role tags (management vs analyst) and fiscal-quarter sequence index for QoQ pairing
- Output contract
- typed rows {statement_id, outcome_valence: positive|negative, attribution_locus: internal-execution|external-factor|unresolved, evidence_char_offsets, confidence 0-1}; statements with no attributable outcome emit no row; ambiguous locus emits unresolved, never a guess
- Model or algorithm
- LLM judgment: local Qwen3.8-Flash-Next (vllm serving name, FP8 weights snapshot 970c569adaca6b35532111fd6b27351b2baefe50, town GPU estate), temperature 0, fixed system prompt classifying one management answer-turn at a time with only local turn context; prompt artifact hashed at freeze of extractor. [v2 amendment per Conway ruling + Hierophant governance path, cst-5cr5.11.21 2026-10-22: scorer-procurement change only; supersedes the claude-haiku-4-5-20251001 pin of v1 807745e0]
- Prompt or code artifact sha256
- Not published
- Scorer blinded to returns
- Yes
- Scorer blinded to future transcripts
- Yes
Preregistration notes
Both hypotheses reuse the settled transcript-alpha apparatus deliberately (splitter, liquidity gates, pairing rules, return fields, estimator) so that any effect is attributable to the new construct, not new plumbing. Scoring model pinned to claude-haiku-4-5-20251001 for cost; if the model is unavailable at execution the family requires a new spec_version, not a silent substitution. External-factor locus list (FX, weather, macro, supply chain, regulation, customer behavior) is illustrative in the prompt, not exhaustive; the classifier decides locus from context. Power note (reviewer item 3, Star): the 0.8 power target is inherited from settled entries 1-4, not computed for this construct; if the wave-001 snapshot lands near the support-gate minima, a variance-based power check against the 1pp SESOI is recorded in the calibration artifact before returns open.
Wave amendments
The public trail below records every admitted change affecting this frozen family.
wave-001-aa-model-procurement-v2AmendedScorer-procurement change only: shared_extraction.model_or_algorithm; no change to hypothesis, estimator, endpoint, universe, leakage boundary, prompt, class definitions, or calibration gate.
Full governance record
- Amendment id
- wave-001-aa-model-procurement-v2
- Amended at utc
- 2026-10-22T22:41:27Z
- Source bead
- cst-5cr5.11.21
- Affected family id
- attribution-asymmetry
- Scope
- Scorer-procurement change only: shared_extraction.model_or_algorithm; no change to hypothesis, estimator, endpoint, universe, leakage boundary, prompt, class definitions, or calibration gate.
- Previous spec sha256
807745e047490efeb9efc29320864c6a683334ca4c07438b010bc3d80fc6fea3- Replacement spec sha256
26a0308f0db2a6e55f87240c784e1e287cb2288bef3d189895b2182e5b09884e- Previous wave freeze sha256
03171c83252c0b3df11aab282e3401c557f99e3f15421f0e75a8f431c8a1e46c- Replacement wave freeze sha256
6d1bae933ca1c4c2bb1bb5d7517e9642f53d945008698a46c9ddf8f93e9cef40- Conway ruling verbatim
- option (b) — score AA with the LOCAL Qwen 3.8 Flash, not an Anthropic key.
- Conway context verbatim
- The original genesis was spitballing with one of the seats about "hey let's find a project that would max out our token burn on our subscriptions and really make use of our local hardware"
- Pre outcome intent citation
- cst-5cr5.11.21: Magician at 19:58 UTC recorded AA as the next LLM-judgment lane; Magician at 20:08 UTC explicitly recorded the missing-Anthropic-key/Engine dependency, before CSL returns opened at 21:27 UTC. Star's 21:44 UTC relay records Conway's assertion that the local-Qwen design intent predated the spec.
- Governance basis
- Hierophant ruling at 21:51 UTC: admitted because the dependency and intended local scorer predated any wave outcome, the change is scorer procurement only, and AA returns remained unopened; no further wave-001 spec amendments are permitted.
- Owner freeze commit
- 08f5308
- Owner fixture verification
- 39/39 plus live Qwen3.8-Flash-Next scorer smoke
- Independent reviewer
- Star
- Reviewer resign ledger time
- 2026-10-22 22:35 UTC
- Reviewer resigned spec sha256
26a0308f0db2a6e55f87240c784e1e287cb2288bef3d189895b2182e5b09884e