wave-002 · quantitative-specificity
Quantitative Specificity: the density and texture of numbers in management speech
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.
QS-1String matchDigit-density delta in prepared remarks
Calls whose prepared remarks carry a higher ratio of digit-token density to vague-language density than the issuer's own trailing norm (issuer-baseline delta of the smoothed log ratio) have higher +2..+60 session SPY-relative abnormal returns; the measured object is the texture of the language (digit-dense versus vague), not a classification of claims, and the delta is preregistered as directionally positive on the Lazy-Prices-adjacent reading that language change carries information.
Full preregistration contract
- Primary feature
- Name
- qs_digit_vague_ratio_delta
- Definition
- digit density = digit tokens per 100 management prepared-remarks tokens (DECLARED: this counts every digit-bearing token, years and fiscal references included - the construct is digit-token density, not quantified claims); vague density = matches per 100 of the frozen vague-language lexicon (DECLARED: adjectives and nouns of vague magnitude, not only grammatical quantifiers) {meaningful, meaningfully, significant, significantly, substantial, substantially, material, materially, considerable, considerably, strong, strongly, solid, robust, healthy, modest, modestly, slight, slightly, somewhat, nicely, 'a lot', 'a bit', 'good deal of', 'great deal of', momentum, encouraging, 'improving trends', 'trending well'}; qs_digit_vague_ratio = ln((digit density + 0.5) / (vague density + 0.5)); feature = call value minus the median of the same value over up to 4 immediately preceding baseline calls (minimum 2)
- Aggregation
- single ratio per call from section-level densities; issuer-baseline delta as defined
- Comparison
- issuer-baseline delta, then quarterly cross-sectional quintiles
- 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; two-sided primary p-value per registry-policy-v1, with the preregistered sign required for a directional verdict
- Universe
- all calls in the wave-002 discovery Research Snapshot passing the settled liquidity gates and splitter verdict ok, with every baseline call also splitter-verdict ok and trailing history per shared_extraction.input_unit
- Exclusions
- splitter verdict not ok on the current call
- fewer than 300 management prepared-remarks tokens on the current call
- fewer than 2 baseline calls that are splitter-verdict ok and have >= 300 management prepared-remarks tokens
- baseline adjacency failing the settled pairing rules (call exclusion; funnel reports the count)
- below settled liquidity/price gates
- incomplete +60 session return horizon at frozen price cutoff
- Missingness rule
- calls below token or baseline support are excluded, not imputed; densities are always computable on a non-empty section (zero matches is a value, not a missing)
- Support gate
- minimum 400 scored calls, 150 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
- deterministic fixture plan: constructed prepared-remarks sections with known digit-token and lexicon-hit counts by construction (digits inside words, '$1.2 billion' tokenization, comma-bearing numerals, lexicon phrases split across lines, case variants, overlapping phrases resolved leftmost-longest); plus a seeded deterministic draw of 30 real discovery-window calls whose densities and ratio are hand-recomputed by a seat other than the owner; PLUS a construct-validity check with a frozen pairing rule: rank the 30 sampled calls by qs_digit_vague_ratio; pair rank 1 of the top tercile with rank 1 of the bottom tercile, rank 2 with rank 2, ... forming exactly 10 pairs; for each pair the reviewer judges, return-blind, which member reads as more digit-dense-versus-vague; agreement = pairs where the reviewer picks the top-tercile member
- Metrics
- deterministic exactness: extractor token counts, hit counts, and ratio equal hand-computed values on 100% of fixtures
- per-call agreement on the 30 sampled real calls with every disagreement itemized by character offset
- construct validity: paired tercile agreement per the frozen pairing rule
- Minimum thresholds
- 100% on fixtures; >= 27/30 exact agreement on real calls; >= 8/10 paired tercile agreements; frozen before scoring
- Disagreement review
- all disagreements listed verbatim with offsets; reviewer signs off on lexicon coverage as frozen (additions after review round are a spec_version bump, never silent)
- Failure verdict
- construct-rejected
- Multiplicity family
- wave-002-primary
- Holdout rule
- common-wave-002
- Primary p value count
- 1
- Leakage boundary
- Allowed fields
- prepared-remarks management turn body text (current call only; baselines are stored scalars)
- speaker labels and role tags (section/turn selection only)
- call_date, fiscal_quarter, fiscal_year (pairing only)
- ticker (issuer-baseline join key only)
- Forbidden fields
- return_fields
- price_fields
- outcome_labels
- future_transcripts
- post_endpoint_metadata
- Q&A section text
- Identifier handling
- required-and-declared
- Date handling
- required-and-declared
- Future text handling
- forbidden
- Pretraining risk mitigation
- no model in the loop - deterministic token and lexicon counting; pretraining leakage structurally impossible
- Rationale
- ticker joins the issuer baseline; dates establish adjacency; prepared-remarks text is the measured object itself; roles select the section and never enter the feature
- Overlap boundary
- Distinct from Temperance's FCI-2 (forward specificity): FCI-2 scores FORWARD claims only - the share carrying both a quantity and a named horizon, an LLM-classified subset - while QS-1 reads the whole prepared-remarks section with no forward/backward classification, no horizon concept, and no model. Distinct from frozen CSL (owns section token COUNTS, not content texture) and CM (per the boundary in CM's own preregistration notes: CM dropped numeric-density constructs because this family owns them). Distinct from settled entries 1-4: none measure digit density.
- Secondary analyses
- digit density alone (no vague denominator), to show the ratio is not driven by one leg (reported, no p-value)
QS-2String matchPrepared-remarks range width
Calls whose prepared remarks express numeric ranges that are relatively WIDE (median relative width of qualifying ranges) have lower +2..+60 session SPY-relative abnormal returns; the construct is the width of numeric range expressions in prepared remarks - numeric imprecision in management communication - with no claim that every range is forward guidance.
Full preregistration contract
- Primary feature
- Name
- qs_range_relative_width
- Definition
- a QUALIFYING RANGE is a frozen-pattern match in management prepared-remarks text of the forms 'X to Y', 'X-Y', or 'between X and Y' where X and Y are digit tokens with compatible units (both percent-marked, both $-marked, or both bare followed within 2 tokens by the same unit word from {percent, bps, 'basis points', million, billion, dollars, cents}), both parse as finite decimals after stripping unit characters with 0 < X < Y strictly (DECLARED: nonpositive, cross-zero, reversed, and equal endpoints are explicitly non-qualifying, so the measure is total on its domain and sign normalization cannot flip it; the funnel reports each non-qualifying class); relative width = (Y - X) / ((X + Y) / 2), strictly positive on the qualifying domain; call value = median relative width over qualifying ranges; requires >= 2 qualifying ranges
- Aggregation
- median over qualifying ranges in the call's prepared remarks
- Comparison
- level, quarterly cross-sectional quintiles (declared: no issuer baseline, because range-bearing calls are a self-selected subset where within-issuer history is often absent)
- 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; two-sided primary p-value per registry-policy-v1, with the preregistered sign required for a directional verdict
- Universe
- all calls in the wave-002 discovery Research Snapshot passing the settled liquidity gates and splitter verdict ok
- Exclusions
- splitter verdict not ok on the current call
- fewer than 300 management prepared-remarks tokens
- fewer than 2 qualifying ranges on the current call
- below settled liquidity/price gates
- incomplete +60 session return horizon at frozen price cutoff
- Missingness rule
- calls with fewer than 2 qualifying ranges are excluded as unsupported, not scored zero; a range that is non-qualifying (incompatible units, nonpositive/cross-zero/reversed/equal endpoints) is reported in the funnel by class
- Support gate
- minimum 300 scored calls, 120 distinct issuers, 6 calendar quarters; if unmet the honest entry is support-failed (first-class under contract v2), declared here as the expected risk for this sparser construct (range-bearing calls are a subset)
- 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
- deterministic fixture plan: constructed sections with ranges of known width (percent pairs, dollar pairs, mixed-unit non-qualifiers, 'X to Y' spanning line breaks, en dash vs hyphen forms, reversed Y < X non-qualifiers, negative/cross-zero/zero-midpoint/equal-endpoint non-qualifiers, '3 to 4 weeks' time ranges which must NOT qualify - time-unit words are absent from the frozen unit list by design, multiple-range medians); plus a seeded deterministic draw of real discovery-window calls with >= 1 pattern match, drawn until the sample contains >= 30 calls AND >= 60 extracted ranges, every extracted range adjudicated and the qualifying set and median width hand-recomputed by a seat other than the owner
- Metrics
- deterministic exactness on 100% of fixtures
- per-call agreement on the sampled real calls (sample floors: >= 30 calls and >= 60 adjudicated ranges)
- Minimum thresholds
- 100% on fixtures; >= 90% call-level exact agreement over every sampled call once both floors are reached (>= 30 calls, >= 60 adjudicated ranges), every range-level disagreement itemized; frozen before scoring
- Disagreement review
- all disagreements and all non-qualifying classifications listed verbatim with offsets; reviewer signs off on the unit list, patterns, and the strictly-positive domain rule as frozen
- Failure verdict
- construct-rejected
- Multiplicity family
- wave-002-primary
- Holdout rule
- common-wave-002
- Primary p value count
- 1
- Leakage boundary
- Allowed fields
- prepared-remarks management turn body text (current call only)
- speaker labels and role tags (section/turn selection only)
- call_date (price alignment only)
- ticker (snapshot join only)
- Forbidden fields
- return_fields
- price_fields
- outcome_labels
- future_transcripts
- post_endpoint_metadata
- Q&A section text
- Identifier handling
- required-and-declared
- Date handling
- required-and-declared
- Future text handling
- forbidden
- Pretraining risk mitigation
- no model in the loop; pretraining leakage structurally impossible
- Rationale
- the measured object is the prepared-remarks text; ticker and call_date exist only to join prices at scoring time, never visible to the extractor beyond the row key
- Overlap boundary
- Distinct from Temperance's FCI family: FCI tracks commitment LIFECYCLE across adjacent calls (restatement, slippage, substitution); QS-2 reads a single call's range geometry with no cross-call linkage, no commitment identity, and (as renamed this round) no forward-guidance classification at all. Distinct from QS-1: QS-1 counts digit density; QS-2 reads the internal width of range expressions only. No settled entry touches range-expression geometry.
- Secondary analyses
- roundness: share of qualifying ranges whose endpoints are both multiples of 5 in their leading unit (reported descriptively, no p-value)
- count of qualifying ranges per call vs returns (reported descriptively)
QS-3String matchNumeric evasion in Q&A
Calls where management answers a higher share of numerically-cued analyst questions WITHOUT any substantive digit (calendar/fiscal/form reference tokens excluded), relative to the issuer's own trailing norm (issuer-baseline delta of the numeric evasion rate), have lower +2..+60 session SPY-relative abnormal returns; declining to produce any substantive number when numerically cued is informative reticence.
Full preregistration contract
- Primary feature
- Name
- qs_numeric_evasion_delta
- Definition
- exchanges per the CM-1 frozen grouping (acr.extract_exchanges over CslParsedCall.qa_turns with the CSL question/answer tests, last-question rule, operator skips); a NUMERIC-CUED question is an exchange whose question turn body matches either (cue branch A) a frozen phrase from {'how much', 'how many', 'what percent', 'what percentage', 'can you quantify', 'could you quantify', 'can you size', 'order of magnitude', 'what kind of number', 'put a number'}, or a BROAD phrase from {'what was the', 'what is the'} ONLY when the same question body also contains a digit token or a frozen metric-unit word from {percent, percentage, bps, 'basis points', margin, margins, rate, rates, growth, revenue, revenues, cost, costs, dollar, dollars, million, billion, units, volume, mix, capex, 'free cash flow'}; or (cue branch B) a digit token together with any of {'versus', 'vs', 'compared to', 'up from', 'down from', 'how does that'}; an exchange is EVADED when the first 150 normalized tokens of the concatenated management answer contain zero SUBSTANTIVE digit tokens, where a digit token is non-substantive iff it matches a frozen reference grammar: four-digit years 1900-2099 standalone or in FY/CY forms (fy24, fy2025, 'calendar 2025'), fiscal-period tokens (q1-q4, h1, h2, 1q25-style), SEC form names (10-k, 10-q, 8-k), and slide/page references ('slide 7', 'page 12' - the digit following those two words); every other digit token is substantive; qs_numeric_evasion_rate = evaded / numeric-cued exchanges, requiring >= 4 numeric-cued exchanges; feature = call rate minus the median rate over up to 4 immediately preceding baseline calls (minimum 2, each with >= 4 numeric-cued exchanges)
- Aggregation
- single rate per call; issuer-baseline delta as defined
- Comparison
- issuer-baseline delta, then quarterly cross-sectional quintiles
- 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; two-sided primary p-value per registry-policy-v1, with the preregistered sign required for a directional verdict
- Universe
- all calls in the wave-002 discovery Research Snapshot passing the settled liquidity gates and splitter verdict ok, with baseline history per shared_extraction.input_unit
- Exclusions
- splitter verdict not ok on the current call
- fewer than 10 management answer turns in the Q&A section (shared with CSL-4/CSL-5 and CM)
- fewer than 4 numeric-cued exchanges on the current call
- fewer than 2 qualifying baseline calls (splitter ok, >= 10 management answer turns, >= 4 numeric-cued exchanges)
- baseline adjacency failing the settled pairing rules
- below settled liquidity/price gates
- incomplete +60 session return horizon at frozen price cutoff
- Missingness rule
- calls below cue or baseline support are excluded, not imputed; an exchange with an empty-body question label has no cue match and is simply not numeric-cued
- Support gate
- minimum 350 scored calls, 140 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
- deterministic fixture plan: constructed Q&A sections with known cue matches and evasion outcomes (broad cue WITH and WITHOUT metric-unit word - the bare 'what is the strategy?' must NOT qualify, cue phrase split across clauses, digit+comparison branch with and without the comparison word, answers whose only digits are years/fiscal periods/form names/slide references - must count as EVADED, substantive digit after token 150 - must count as evaded, substantive digit before token 150 - answered, multi-executive concatenation); plus a seeded deterministic draw of 30 real discovery-window calls hand-recomputed by a seat other than the owner; PLUS construct validity on a STRATIFIED seeded draw of EXACTLY 30 real numeric-cued exchanges satisfying all strata simultaneously (strata may overlap): >= 12 from cue branch A (>= 4 via the broad-phrase+metric-unit path), >= 8 from cue branch B, and by answer-head digit category >= 6 with only non-substantive digits, >= 6 with zero digits, and >= 6 with a substantive digit before token 150; reviewer labels each, return-blind: (a) does the question genuinely ask for a number, (b) is the evaded/answered verdict right by their reading
- Metrics
- deterministic exactness on 100% of fixtures
- per-call agreement on the 30 sampled real calls
- cue precision overall and as a proportion per cue branch
- evasion verdict agreement overall and per answer-digit category
- Minimum thresholds
- 100% on fixtures; >= 27/30 call agreement; cue precision >= 24/30 overall AND >= 0.70 within each cue branch at the stated support floors (A >= 12, B >= 8); evasion verdict agreement >= 25/30 overall AND >= 5/6 within each of the three answer-digit categories at their >= 6 floors; frozen before scoring
- Disagreement review
- every disagreement itemized by exchange index and stratum with the question and answer-head text quoted; reviewer signs off on both cue branches, the metric-unit list, the reference grammar, and the 150-token head rule as frozen
- Failure verdict
- construct-rejected
- Multiplicity family
- wave-002-primary
- Holdout rule
- common-wave-002
- Primary p value count
- 1
- Leakage boundary
- Allowed fields
- Q&A turn structure and body text: speaker labels, role tags, turn order, question and answer turn text (current call; baselines are stored scalars)
- call_date, fiscal_quarter, fiscal_year (pairing only)
- ticker (issuer-baseline join key only)
- Forbidden fields
- return_fields
- price_fields
- outcome_labels
- future_transcripts
- post_endpoint_metadata
- prepared-remarks text
- Identifier handling
- required-and-declared
- Date handling
- required-and-declared
- Future text handling
- forbidden
- Pretraining risk mitigation
- no model in the loop; pretraining leakage structurally impossible
- Rationale
- the measured object is Q&A text; ticker joins the baseline; dates establish adjacency; roles group exchanges and never enter the feature
- Overlap boundary
- DESIGNED AGAINST settled entry 4 (the evasion proxy), per the family claim: entry 4 scored semantic answer relevance over ALL questions with an embedding model; QS-3 restricts to a deterministic numeric-cue subset, tests only the presence of SUBSTANTIVE digits in the answer head (frozen reference grammar), uses no model and no relevance judgment, and is an issuer-baseline DELTA rather than a level. Distinct from CM-2/CM-3 (filler and hedge density: lexical delivery texture, no question-conditioning, no numeric concept) and CM-4 (question echo: verbatim reuse, not numeric content). Distinct from Sun's analyst-agenda-shift: that family reads question CONTENT for topic; QS-3 uses the question only as a deterministic trigger.
- Secondary analyses
- evasion rate level (no baseline) as a cross-check (reported, no p-value)
- evasion under the raw any-digit test (no reference grammar), to show the substantive-digit exclusion matters (reported, no p-value)
Shared extraction contract
- Input unit
- single call, joined to the same issuer's prior quarterly calls' stored per-call feature rows (never prior transcript text) for trailing baselines, exactly as the CM family declared: baseline rows are the deterministic per-call scalars this family emits (QS-1..QS-3 call-level values plus support counts) and MAY be read from calls preceding the discovery window so early discovery quarters remain scorable.
- Input binding
- transcript-gate splitter output with verdict ok, via import from the CSL extractor module's public API (workshop/signal-atlas/extractors/call_structure_logistics.py, csl_parse_call -> CslParsedCall) at the behavior manifest current at QS freeze time, so this family cannot disagree with CSL/CM about section boundaries, speaker roles, or answer turns; csl_parse_call returning None (non-ok splitter verdict) makes all features null. QS-1 and QS-2 read the body text of management turns in the PREPARED-REMARKS section only. QS-3 reads Q&A exchanges formed by the same acr.extract_exchanges grouping algorithm with the CSL question/answer tests as frozen in CM-1, reading the analyst question turn body and the concatenated management answer turn bodies. Snapshot metadata limited to ticker, call_date, fiscal_quarter, fiscal_year, calendar_quarter.
- Output contract
- one typed row per call: {transcript_id, pr_mgmt_tokens, qs_digit_density (per 100 prepared-remarks management tokens), qs_vague_density (per 100), qs_digit_vague_ratio, qualifying_range_count, qs_range_relative_width (median), numeric_question_count, evaded_numeric_question_count, qs_numeric_evasion_rate, per-field support counts, explicit null for any field whose inputs are absent}; evidence = matched character offsets for every digit-bearing token, lexicon hit, and qualifying range, plus exchange turn indices for QS-3; no confidence field - deterministic code, not a judgment
- Model or algorithm
- deterministic algorithm: Python extractor over CSL-module output; whitespace tokenization (str.split); match normalization = case-fold, strip non-alphanumeric characters except '.' '%' '$' and ',' WITHIN digit-bearing tokens (so '120bps,' '14.5%,' '$1.2' survive as one token), CSL-5 rule otherwise; a DIGIT TOKEN is a normalized token containing at least one of 0-9; frozen phrase lists verbatim in each hypothesis; leftmost-longest non-overlapping phrase matching; no model in the loop; code artifact hashed before scoring together with the behavior manifest of every imported module, per the CSL-6 precedent
- Prompt or code artifact sha256
- Not published
- Scorer blinded to returns
- Yes
- Scorer blinded to future transcripts
- Yes
Preregistration notes
Family boundary: the density and texture of NUMBERS in management speech - digit density and its ratio to a vague-language lexicon (QS-1, prepared remarks), the geometry of numeric range expressions (QS-2, prepared remarks), and numeric responsiveness under direct numeric questioning (QS-3, Q&A) - claimed 2026-10-19 on cst-5cr5.11 and ratified in the family-claims round; CM's frozen notes record dropping its numeric constructs to this family. Reviewer independence: author (Star) != reviewer; reviewer Temperance (accepted 2026-10-26); reviewer probes are built from this spec text, not the owner's fixture suite (CM precedent). Review-round changelog: v1 draft filed 2026-10-26; round 1 (Temperance, 2026-10-26, B1-B5): schema fixes (archaeology pointer removed, RFC3339 claimed_at, reviewer confirmed); all primaries moved to two-sided p-values with preregistered sign per registry-policy-v1; QS-1 honestly narrowed to digit-token density vs vague-language lexicon with a frozen 10-pair tercile validity rule; QS-2 renamed to prepared-remarks range width (guidance interpretation dropped), domain restricted to 0 < X < Y with all other endpoint classes non-qualifying, real-call gate requires >= 60 adjudicated ranges; QS-3 broad cues gated on metric-unit context, answer test counts substantive digits only (frozen calendar/fiscal/form/reference grammar), validity draw stratified per cue branch and answer category with per-branch thresholds; round 2 (Temperance, 2026-10-26, R2-A..D exactness fixes): claimed_at set to the exact ledger timestamp of the family claim; digit-density/range-expression/substantive-digit terminology cut over with no aliases; QS-2 real-sample gate restated as >= 90% call-level agreement once both floors (>= 30 calls, >= 60 ranges) are reached; QS-3 validity draw fixed at exactly 30 overlapping-strata exchanges with proportional per-branch precision (>= 0.70) and per-category verdict agreement (>= 5/6) gates. Known risks and gates: QS-2 support is the declared weak point (range-bearing calls are a subset; support-failed is the honest outcome if the gate misses); QS-1's vague-language lexicon and QS-3's cue list are frozen verbatim at freeze and any post-review addition is a spec_version bump; construct-validity thresholds (QS-1 paired tercile agreement, QS-3 per-branch cue precision and per-category verdict agreement) are construct-rejection gates, not advisories. Process archaeology lives in the owning bead (source_bead).