wave-002 · analyst-agenda-shift
Analyst Agenda Shift: adverse operating topics introduced and converged on in earnings-call questions
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.
AAS-1Llm judgmentNew adverse analyst-question share predicts negative abnormal return
Calls with a larger token share of adverse analyst-question sentences whose operating topic is absent from both current prepared remarks and the immediately preceding adjacent call's analyst questions earn lower +2..+60-session SPY-relative abnormal returns than calls with a smaller share; the interpretation is that the analyst agenda surfaced adverse operating issues management did not prepare investors for and analysts did not raise in the previous quarter.
Full preregistration contract
- Primary feature
- Name
- aas_new_adverse_question_share
- Definition
- numerator = whitespace tokens in current analyst-question sentences labeled adverse=true with a resolved primary topic code that appears in neither the current prepared-remarks resolved topic set nor the immediately preceding adjacent call's analyst-question resolved topic set; denominator = whitespace tokens in all current analyst-question sentences except NONSUBSTANTIVE; OTHER_DOMAIN_SPECIFIC and UNRESOLVED remain in the denominator but never the numerator
- Aggregation
- one scalar in [0,1] per eligible current call from non-overlapping sentence rows; topic absence is exact equality over the frozen topic-code vocabulary after independent scoring of each section
- Comparison
- cross-sectional level within calendar quarter, quintile sort high-minus-low, matching the 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 at 1/99 percent, per the transcript-alpha settled-entry convention
- Price timing
- close of the first SPY session strictly after call_date through close of session +60
- Estimator
- quarterly high-minus-low feature quintile spread, Newey-West lag-1 uncertainty on quarterly spreads, with ticker-cluster bootstrap reported under the settled Atlas convention; one negative-direction primary test
- Universe
- all current calls in the wave-002 Research Snapshot passing settled liquidity gates, with splitter verdict ok and an immediately preceding adjacent quarterly transcript that also has splitter verdict ok under the settled pairing rules
- Exclusions
- splitter verdict not ok on the current or immediately preceding call
- missing previous transcript, nonadjacent fiscal identity, or nonquarterly call gap under settled pairing rules
- current prepared remarks < 500 whitespace tokens, current analyst questions < 500 whitespace tokens, or previous-call analyst questions < 500 whitespace tokens
- fewer than 5 current analyst question turns
- UNRESOLVED topic tokens exceed 35% of substantive tokens in current prepared remarks, current analyst questions, or previous-call analyst questions
- below settled liquidity or price gates
- incomplete +60-session return horizon at the frozen price cutoff
- Missingness rule
- complete-case only. OTHER_DOMAIN_SPECIFIC and UNRESOLVED rows never establish topic presence or novelty; they remain in the current-question denominator. A true zero numerator is a valid zero, not missing. Calls failing an ordered exclusion are excluded and counted in the funnel; no label or feature is imputed.
- Support gate
- minimum 400 scored adjacent pairs, 150 distinct issuers, and 6 calendar quarters before a market test
- 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 seeded sector-stratified draw of at least 200 analyst question turns and their current prepared-remarks and previous-question comparison sentences, extended until every primary topic code used in scoring has at least 20 human-positive sentences, both adverse=true and derived novel=true have at least 30 positives, and the sample has at least 30 multi-topic sentences split between main-clause-dominant and genuinely balanced cases. A non-owner labels sentence offsets, primary topic under the frozen main-clause/deterioration-locus rule, adverse flag, and derived prepared/prior absence without returns or prices. Owner fixtures and Hermit's independently authored spec-derived probes are separate hashed artifacts under cst-5cr5.11.16.
- Metrics
- sentence-boundary and evidence-offset exactness
- per-class topic precision and recall plus macro-F1 over primary topic codes
- adverse=true precision and recall
- derived novel=true precision and recall
- confusion matrices and OTHER_DOMAIN_SPECIFIC, NONSUBSTANTIVE, and UNRESOLVED rates
- primary-topic-rule agreement on the multi-topic subset
- deterministic feature-aggregation exactness
- Minimum thresholds
- 100% sentence-boundary and offset exactness; topic precision and recall each >= 0.75 for every primary code after the class-support extension and topic macro-F1 >= 0.75; exact agreement >= 0.80 on the multi-topic subset, with balanced no-dominant-locus cases scored as UNRESOLVED; adverse precision and recall each >= 0.80; derived novelty precision >= 0.85 and recall >= 0.75; 30/30 owner aggregation fixtures and all reviewer probe groups exact
- Disagreement review
- all human/model disagreements, class support counts, declared extractor readings, and each reviewer probe's governing specification sentence are listed in the calibration artifact; Hermit signs both fixture coverage and the disagreement record
- Failure verdict
- construct-rejected
- Multiplicity family
- wave-002-primary
- Holdout rule
- common-wave-002
- Primary p value count
- 1
- Leakage boundary
- Allowed fields
- current prepared-remarks sentence text and offsets
- current and immediately previous analyst-question sentence text and offsets
- speaker role and stable within-call analyst pseudonym
- ticker, call_date, fiscal_quarter, fiscal_year, and calendar_quarter outside the model for deterministic issuer pairing only
- sector outside the model for calibration stratification only
- Forbidden fields
- return_fields
- price_fields
- outcome_labels
- future_transcripts
- post_endpoint_metadata
- management answer text
- Identifier handling
- required-and-declared
- Date handling
- required-and-declared
- Future text handling
- forbidden
- Pretraining risk mitigation
- issuer, ticker, speaker and person names, absolute dates, fiscal-period labels, and unique proper nouns are masked before the semantic scorer; the model sees one sentence plus its masked parent turn. Generic industry nouns, relative-time words, and quantities remain because they are necessary to classify operating topic and adverse framing. Residual recognition of a famous event from generic context is declared irreducible and reported per calibration item.
- Rationale
- ticker and fiscal/date fields are required only for deterministic same-issuer adjacency and never reach the model. Question text is the construct; current prepared text and prior question text are necessary comparison sets. Stable pseudonyms are required to preserve within-call analyst distinctness without identity. No management answer or outcome field is needed.
- Overlap boundary
- Distinct from settled ACR/evasion and Attribution Asymmetry because no management answer is visible or scored; distinct from FCI because no commitment or disposition is extracted; distinct from Quantitative Specificity because numbers are not a feature. CSL owns section boundaries, token semantics, speaker attribution, and trailing assembly and is imported rather than duplicated; CSL measures shape and counts, while AAS assigns question/prepared sentence tokens to semantic operating-topic codes.
- Secondary analyses
- descriptive, no p-value: report numerator, denominator, resolved-topic coverage, and topic-code composition by calendar quarter
AAS-2Llm judgmentIndependent analyst convergence on adverse topics predicts negative abnormal return
Calls where a larger share of adverse operating-topic codes is independently raised by at least two distinct analysts earn lower +2..+60-session SPY-relative abnormal returns; the interpretation is that cross-analyst convergence identifies concerns not attributable to one analyst's idiosyncratic agenda.
Full preregistration contract
- Primary feature
- Name
- aas_independent_adverse_convergence
- Definition
- numerator = count of distinct resolved primary topic codes appearing in at least one adverse=true sentence from each of at least two distinct within-call analyst pseudonyms; denominator = count of distinct resolved primary topic codes appearing in any adverse=true current analyst-question sentence
- Aggregation
- one scalar in [0,1] per eligible call; multiple mentions by one analyst count once for that topic, and OTHER_DOMAIN_SPECIFIC, NONSUBSTANTIVE, and UNRESOLVED never enter either count
- 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 at 1/99 percent, per the transcript-alpha settled-entry convention
- Price timing
- close of the first SPY session strictly after call_date through close of session +60
- Estimator
- quarterly high-minus-low feature quintile spread, Newey-West lag-1 uncertainty on quarterly spreads, with ticker-cluster bootstrap reported under the settled Atlas convention; one negative-direction primary test
- Universe
- all calls in the wave-002 Research Snapshot passing settled liquidity gates and splitter verdict ok, with at least two identifiable analyst speakers in the Q&A section
- Exclusions
- splitter verdict not ok on the current call
- current analyst questions < 500 whitespace tokens or fewer than 5 analyst question turns
- fewer than 2 distinct analyst pseudonyms holding question turns
- zero resolved adverse primary topic codes in current analyst questions
- UNRESOLVED topic tokens exceed 35% of substantive current analyst-question tokens
- below settled liquidity or price gates
- incomplete +60-session return horizon at the frozen price cutoff
- Missingness rule
- calls with no resolved adverse primary topic code or insufficient identifiable analysts are excluded as unsupported, not assigned zero. A supported call with adverse topics but no cross-analyst duplicate receives a valid zero. No speaker identity or semantic label is imputed.
- Support gate
- minimum 400 scored calls, 150 distinct issuers, 6 calendar quarters, and at least 100 calls in each extreme feature quintile before a market test
- 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 AAS-1 sentence/topic/adverse calibration sample and thresholds; adds deterministic constructed and real-call speaker-grouping probes covering spelling/case variants, follow-up turns, operator interleaves, unidentified callers, and two analysts asking about the same topic. Owner fixtures and Hermit's independently authored spec-derived probes remain separate hashed artifacts.
- Metrics
- AAS-1 topic and adverse metrics
- analyst-pseudonym grouping agreement
- topic-by-analyst incidence exactness
- deterministic convergence aggregation exactness
- Minimum thresholds
- all AAS-1 topic and adverse thresholds pass; 100% exactness on constructed speaker fixtures; >= 27/30 exact analyst grouping on sampled real calls; 30/30 convergence aggregation fixtures and all reviewer probe groups exact
- Disagreement review
- all speaker-grouping and topic-incidence disagreements are listed with masked labels, turn indices, and governing specification sentences; Hermit signs the declared role-normalization reading and probe coverage
- Failure verdict
- construct-rejected
- Multiplicity family
- wave-002-primary
- Holdout rule
- common-wave-002
- Primary p value count
- 1
- Leakage boundary
- Allowed fields
- current analyst-question sentence text and offsets
- speaker role and stable within-call analyst pseudonym
- ticker and call_date outside the model for cohort joins only
- sector outside the model for calibration stratification only
- Forbidden fields
- return_fields
- price_fields
- outcome_labels
- future_transcripts
- post_endpoint_metadata
- prepared-remarks text
- management answer text
- Identifier handling
- required-and-declared
- Date handling
- required-and-declared
- Future text handling
- forbidden
- Pretraining risk mitigation
- the AAS-1 masking and local-context rules apply; analyst names are replaced by stable within-call pseudonyms before semantic scoring
- Rationale
- question text supplies topic and adverse framing; stable pseudonyms are the minimum field needed to distinguish independent analysts. Ticker/date are execution-only cohort keys and never reach the semantic scorer.
- Overlap boundary
- Distinct from CSL-3: CSL-3 counts how many analysts participate versus issuer norm without reading question content; AAS-2 conditions on adverse semantic topic and measures whether multiple analysts independently converge on the same topic. Distinct from New-Speaker Regime because no management speaker or speaker transition is scored.
- Secondary analyses
- descriptive, no p-value: report the distribution of analyst multiplicity per adverse topic code
- descriptive, no p-value: residualize aas_independent_adverse_convergence on distinct analyst-pseudonym count within calendar quarter and report the same high-minus-low spread, plus raw convergence by analyst-count tercile
AAS-3Llm judgmentPrepared-remarks versus analyst-question topic-allocation gap predicts negative abnormal return
Calls with a larger total-variation distance between prepared-remarks and analyst-question operating-topic token distributions earn lower +2..+60-session SPY-relative abnormal returns; the interpretation is that analysts devoted attention to a materially different operating agenda than management prepared to discuss.
Full preregistration contract
- Primary feature
- Name
- aas_prepared_question_tv
- Definition
- one half times the sum across resolved primary topic codes of the absolute difference between each code's share of resolved topic tokens in current analyst questions and its share of resolved topic tokens in current prepared remarks; distributions exclude OTHER_DOMAIN_SPECIFIC, NONSUBSTANTIVE, and UNRESOLVED and separately normalize to one
- Aggregation
- one scalar in [0,1] per eligible call from sentence-level whitespace token counts
- 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 at 1/99 percent, per the transcript-alpha settled-entry convention
- Price timing
- close of the first SPY session strictly after call_date through close of session +60
- Estimator
- quarterly high-minus-low feature quintile spread, Newey-West lag-1 uncertainty on quarterly spreads, with ticker-cluster bootstrap reported under the settled Atlas convention; one negative-direction primary test
- Universe
- all calls in the wave-002 Research Snapshot passing settled liquidity gates and splitter verdict ok, with current prepared remarks and analyst questions both long enough to form resolved topic distributions
- Exclusions
- splitter verdict not ok on the current call
- current prepared remarks < 500 whitespace tokens or current analyst questions < 500 whitespace tokens
- fewer than 5 current analyst question turns
- fewer than 200 resolved primary-topic tokens in either prepared remarks or analyst questions
- fewer than 3 distinct resolved primary topic codes in either section
- UNRESOLVED topic tokens exceed 35% of substantive tokens in either section
- below settled liquidity or price gates
- incomplete +60-session return horizon at the frozen price cutoff
- Missingness rule
- both topic distributions must satisfy all support rules; no pseudocounts or imputation. Zero total-variation distance is valid when supported distributions match exactly. Excluded calls are counted by reason in the funnel.
- Support gate
- minimum 400 scored calls, 150 distinct issuers, and 6 calendar quarters before a market test
- 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 AAS-1 sentence/topic calibration artifact; adds constructed section pairs with known topic-token vectors, zero and unit total-variation boundaries, excluded labels, and insufficient-support cases, plus 30 seeded real calls whose topic vectors and total-variation values are independently recomputed. Owner fixtures and Hermit's spec-derived probes are separate hashed artifacts.
- Metrics
- AAS-1 topic metrics
- per-section resolved topic-token vector exactness
- total-variation exactness
- support-exclusion exactness
- Minimum thresholds
- all AAS-1 topic thresholds pass; 100% exactness on constructed token-vector and boundary fixtures; 30/30 real-call vector and total-variation recomputations exact; all reviewer probe groups exact
- Disagreement review
- every sentence-label disagreement and any derived-vector mismatch is listed with masked offsets and the governing specification sentence; Hermit signs the imported token-semantics reading and all total-variation probes
- Failure verdict
- construct-rejected
- Multiplicity family
- wave-002-primary
- Holdout rule
- common-wave-002
- Primary p value count
- 1
- Leakage boundary
- Allowed fields
- current prepared-remarks and analyst-question sentence text and offsets
- speaker role tags sufficient to select analyst question turns
- ticker and call_date outside the model for cohort joins only
- sector outside the model for calibration stratification only
- Forbidden fields
- return_fields
- price_fields
- outcome_labels
- future_transcripts
- post_endpoint_metadata
- management answer text
- previous-call transcript text
- Identifier handling
- required-and-declared
- Date handling
- required-and-declared
- Future text handling
- forbidden
- Pretraining risk mitigation
- the AAS-1 masking and local sentence-plus-turn context restriction apply independently to each current-call section; section labels are retained because the construct compares them
- Rationale
- the two current-call text sections are exactly the distributions being compared. Ticker/date remain outside the model as execution keys. No prior or future text and no answer text is necessary.
- Overlap boundary
- Distinct from CSL-2: CSL-2 measures only the scalar Q&A-to-prepared total token-count ratio and reads no content; AAS-3 holds section size conceptually separate and compares the semantic allocation of resolved tokens across frozen topic codes. It imports CSL token semantics so any difference is content rather than competing tokenization.
- Secondary analyses
- descriptive, no p-value: report AAS-3 after residualizing on log prepared tokens and log analyst-question tokens within calendar quarter
- descriptive, no p-value: report per-topic contributions to total-variation distance
AAS-4Llm judgmentJump in new adverse topic breadth versus issuer history predicts negative abnormal return
Calls where the count of distinct new adverse analyst-question topics rises above the issuer's trailing four-call median earn lower +2..+60-session SPY-relative abnormal returns; the interpretation is that a sudden broadening of the adverse analyst agenda signals emerging operating problems rather than a stable issuer-specific questioning style.
Full preregistration contract
- Primary feature
- Name
- aas_adverse_breadth_jump
- Definition
- current count of distinct resolved primary topic codes qualifying for the AAS-1 numerator minus the median of that same count over the issuer's four immediately preceding adjacent extraction-eligible AAS-1 rows; each historical row determines novelty only from information available as of that historical call. A historical row is extraction-eligible when it passes AAS-1's splitter, adjacency, section-token, question-turn, and UNRESOLVED-rate rules under the same frozen extractor; historical liquidity, price, and +60-session return availability do not affect baseline eligibility.
- Aggregation
- one signed scalar per current call; the four-row median uses stored return-blind AAS-1 extraction rows, never rescored future context
- 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 at 1/99 percent, per the transcript-alpha settled-entry convention
- Price timing
- close of the first SPY session strictly after call_date through close of session +60
- Estimator
- quarterly high-minus-low feature quintile spread, Newey-West lag-1 uncertainty on quarterly spreads, with ticker-cluster bootstrap reported under the settled Atlas convention; one negative-direction primary test
- Universe
- current calls in the AAS-1 universe with four immediately preceding adjacent quarterly extraction-eligible AAS-1 rows under the same frozen extractor; the earliest baseline row must itself have its required predecessor available. Historical baseline rows need not pass liquidity, price, or return-horizon gates because they are return-blind feature history, while every current scored call must pass all AAS-1 market and endpoint gates.
- Exclusions
- current call fails any AAS-1 exclusion, including its liquidity, price, and return-horizon gates
- fewer than four immediately preceding extraction-eligible AAS-1 rows after applying only AAS-1's splitter, adjacency, section-token, question-turn, and UNRESOLVED-rate rules to historical rows
- any gap in the five-row current-plus-baseline sequence violates settled adjacent-quarter pairing rules
- any baseline row was produced by a different prompt, taxonomy, CSL behavior manifest, or AAS extractor artifact hash
- Missingness rule
- four valid immediately preceding extraction-eligible AAS-1 rows are mandatory; historical market eligibility and endpoint availability are not consulted. No shorter baseline, carry-forward, or imputation is allowed. A count of zero is valid for current or historical rows. Every excluded sequence is reported by first ordered reason.
- Support gate
- minimum 250 scored sequences, 100 distinct issuers, and 6 calendar quarters before a market test
- 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 fully passing AAS-1 calibration artifact; adds deterministic six-call sequences with known per-call new-adverse topic sets, missing quarters, missing predecessor support, ties, zero counts, and extractor-hash mismatches, plus 30 seeded real issuer sequences independently recomputed without returns. Owner fixtures and Hermit's spec-derived probes are separate hashed artifacts.
- Metrics
- AAS-1 calibration metrics
- adjacent-sequence assembly exactness
- four-row median and breadth-jump exactness
- ordered exclusion exactness
- Minimum thresholds
- all AAS-1 thresholds pass; 100% exactness on constructed sequence, median, hash-binding, and exclusion fixtures; 30/30 real-sequence recomputations exact; all reviewer probe groups exact
- Disagreement review
- every sequence or median mismatch is listed with masked transcript ids, ordered input rows, artifact hashes, and governing specification sentences; Hermit signs the imported trailing-assembly reading and probe coverage
- Failure verdict
- construct-rejected
- Multiplicity family
- wave-002-primary
- Holdout rule
- common-wave-002
- Primary p value count
- 1
- Leakage boundary
- Allowed fields
- the AAS-1 allowed fields as of each historical and current call
- stored return-blind AAS-1 feature rows and support counts from the current and four previous eligible calls
- ticker, call_date, fiscal_quarter, fiscal_year, and calendar_quarter outside the model for deterministic sequence assembly
- artifact and behavior-manifest hashes for version equality
- Forbidden fields
- return_fields
- price_fields
- outcome_labels
- future_transcripts
- post_endpoint_metadata
- management answer text
- Identifier handling
- required-and-declared
- Date handling
- required-and-declared
- Future text handling
- forbidden
- Pretraining risk mitigation
- each underlying AAS-1 row is scored under AAS-1 masking and local-context rules as of that call; the baseline calculation reads only frozen feature rows, support counts, adjacency metadata, and artifact hashes
- Rationale
- issuer identity and fiscal/date fields are required to assemble a causal trailing sequence but never reach the semantic model. Historical question/prepared text is used only to produce each historical AAS-1 row with information then available; future text is structurally inaccessible.
- Overlap boundary
- Distinct from CSL trailing-baseline hypotheses in measured content: AAS-4 imports CSL adjacency and median machinery but baselines a semantic count of new adverse analyst-question topics, not section size, participation, scheduling, or speaker concentration. Distinct from AAS-1 because the primary feature is the issuer-history change in distinct-topic breadth rather than the current call's adverse token share.
- Secondary analyses
- descriptive, no p-value: Shannon entropy of current resolved new-adverse topic token shares minus the trailing four-call median entropy
- descriptive, no p-value: AAS-4 recomputed with a minimum two-row baseline to show coverage sensitivity without replacing the primary
Shared extraction contract
- Input unit
- one current earnings call, with the immediately preceding adjacent quarterly call for AAS-1 and up to four prior extraction-eligible AAS-1 feature rows for AAS-4; all historical inputs precede the scored call and may reach before the wave-002 discovery window, while holdout baselines may reach into the discovery window
- Input binding
- transcript-gate splitter output with verdict ok for current prepared remarks and current analyst question turns; prior-call inputs are analyst question turns only. Structural parsing is imported from workshop/signal-atlas/extractors/call_structure_logistics.py::csl_parse_call: section boundaries, normalized role attribution, question turns, stable within-call analyst pseudonyms, whitespace-token semantics, adjacent-quarter pairing, and trailing-history assembly. Ticker, call_date, fiscal_quarter, fiscal_year, and calendar_quarter are available only to deterministic pairing and cohort code; sector is available only for calibration stratification. Management answer turns are discarded before semantic scoring.
- Output contract
- one typed sentence row per prepared-remarks or analyst-question sentence: {transcript_id, section: prepared|analyst-question, turn_index, analyst_pseudonym|null, sentence_index, evidence_char_start, evidence_char_end, whitespace_token_count, topic_code, adverse: true|false|unresolved, confidence_0_1}; plus one typed feature row per eligible current call: {transcript_id, aas_new_adverse_question_share, aas_independent_adverse_convergence, aas_prepared_question_tv, aas_adverse_breadth_jump, per-feature support counts, explicit unresolved or null reason}. Sentence offsets point into the masked scorer input and are mapped deterministically back to source offsets. Every sentence emits exactly one frozen topic code; ambiguous topic or adverse readings emit UNRESOLVED, never a guess.
- Model or algorithm
- two-stage extractor. Stage 1 is deterministic CSL import plus sentence segmentation and whitespace token counting. Stage 2 is LLM judgment using local Qwen3.8-Flash-Next (vLLM serving name; FP8 weights snapshot 970c569adaca6b35532111fd6b27351b2baefe50), temperature 0, enable_thinking false, JSON-schema decoding, one sentence with its parent turn as context, maximum 2000 output tokens. Frozen primary topic codes: DEMAND_VOLUME, PRICING_MIX, REVENUE_GUIDANCE, MARGIN_COST, CASH_LIQUIDITY, CAPITAL_ALLOCATION, CUSTOMER_CONCENTRATION_RETENTION, COMPETITION_MARKET_SHARE, OPERATIONS_SUPPLY_QUALITY, CAPACITY_LABOR, PRODUCT_PIPELINE_EXECUTION, REGULATORY_LEGAL, ACCOUNTING_CONTROLS, M_AND_A_PORTFOLIO, MACRO_FX_RATES, OTHER_DOMAIN_SPECIFIC, NONSUBSTANTIVE, UNRESOLVED. Every sentence receives exactly one code. For a multi-topic sentence, the primary code is the operating topic in the main clause and, when the sentence describes a causal chain, the locus of the stated deterioration; a causal antecedent or subordinate mechanism does not win merely because it appears first. Coordinated independent main clauses with no dominant deterioration locus emit UNRESOLVED. A sentence is adverse only when it explicitly states or presupposes deterioration, downside risk, shortfall, constraint, uncertainty, challenge, delay, loss, negative variance, or pressure; a routine status request without an adverse premise is false. Substantive tokens means all whitespace tokens in sentences whose topic_code is not NONSUBSTANTIVE; OTHER_DOMAIN_SPECIFIC and UNRESOLVED tokens are substantive. Deterministic aggregation compares topic-code sets and token counts; OTHER_DOMAIN_SPECIFIC, NONSUBSTANTIVE, and UNRESOLVED never create novelty or enter resolved topic distributions. The frozen AAS artifact binds the AAS prompt, taxonomy, segmentation code, and the imported CSL behavior manifest together so upstream behavior cannot move silently.
- Prompt or code artifact sha256
- Not published
- Scorer blinded to returns
- Yes
- Scorer blinded to future transcripts
- Yes
Preregistration notes
Family boundary: question-side semantic topic content only; management answers, answer behavior, attribution, commitments, and numeric specificity are outside AAS. CSL supplies structural parsing, token semantics, speaker grouping, adjacency, and trailing-baseline machinery by imported behavior manifest; CSL measures shape/counts while AAS assigns structural tokens to operating-topic codes. Reviewer independence: Sun owns AAS; Hermit owns CSL but has no AAS topic-content construct and accepted independent review on cst-5cr5.11.31. Review-round changes: 2026-10-26 pre-draft flag accepted — import CSL machinery rather than silently re-derive it, and bind its behavior manifest into AAS execution hashes. 2026-10-26 review R1 — AAS-4 historical baselines use extraction/support eligibility only; market and endpoint gates apply to the current call. R2 — multi-topic sentences follow the main-clause/deterioration-locus rule, with balanced cases UNRESOLVED and explicitly calibrated. R3 — AAS-2 analyst-count-confound diagnostics added as non-primary descriptive analyses. R4 — substantive tokens defined as every token outside NONSUBSTANTIVE sentences. Known risks/gates: coarse taxonomy and synonymy can create false absence, AAS-1/AAS-4 are dependent, rare topic classes may fail support, and pretraining recognition is irreducible after masking; return-blind calibration, per-class floors, support gates, owner fixtures plus independent text-derived probes, and construct-rejected/support-failed verdicts govern. Process archaeology and declared readings live on cst-5cr5.11.31.