Transcript Alpha — Evidence Appendix

A / Manifest

Durable artifacts

Result-file hashes identify the exact aggregate outputs used for this report. Script and panel hashes are also embedded inside those outputs.

Committed evidence and local reproducible panels.
Artifact	Role	SHA-256	Disposition
`tone_delta_results.json`	Tone aggregate output	b448016793c6721f3341154556193c630c1383d31d6a76c8d043479f59a3f5aa	Committed
`acr_results.json`	ACR aggregate output	e32a3c929fae5c64ca1d502a52639243950f2907f7ef4e6ba1ec80b65611b472	Committed
`evasion_results.json`	Evasion aggregate output	347d63cd12e060f4ff0eec2a48403acf276565e56ab0654d09e1eaf01d277020	Committed
`transcript-gate/gate.py`	Section splitter	b81edefed94dd2fc19e0dd3d41597a0149a4fb2f766ad60fd4c596c2ff4f956b	Committed
`tone_delta.py`	Tone pipeline	915ba31cc378c731477b935526b1a33d07a6168a84197fffc8f8acfa4b762f15	Committed
`acr.py`	ACR pipeline	ccd9307ad15c7e790ac454f5ea4939fdddc92b2f958c459e72a39a0534908730	Committed
`evasion.py`	Evasion pipeline	e11d09ebe117fc72a44c6d763512016f188cf3c63d4c8cfeb734fd5795516652	Committed
`embedding-requirements.lock`	28-package exact environment	ece967573db2a74cdf402fb8ccb07d636f61a1200af8d35537fe122e5e46aa7b	Committed
tone_delta_panel.csv	17,641 feature/outcome rows	717ebdac98e04219de0ea12806d1e861ad721fb21bf2c5ccf34c7278a5c7b9f0	Local / ignored
acr_panel.csv	17,570 call features	8a2fe198d6691b0b0aeb5be9e8e5c47f840367666a9428549c489514eff8043e	Local / ignored
evasion_panel.csv	17,285 event features	c2af85e82dbd8771b69881ee88f36254cfddad8ce3007339e85a9da4de40f4ba	Local / ignored

B / Gate & design

Lineage before signal

The upstream data gate and fixed return panel precede every ACR and evasion result. Later stages consume and verify the earlier panel hash rather than rebuilding the universe.

G1

Provider structure and database inventory

A bounded FMP request returned fields {symbol, period, year, date, content}; speaker turns exist only as inline text. Neither raw FMP nor StockQL provides a prepared/Q&A section boundary. The checked StockQL transcript was byte-identical to provider content, so ingestion did not discard section metadata. StockQL held 173,795 transcripts dated 2005-10-11 through 2026-10-01; participant JSON was populated in 57,493.

Source: Bead cst-5cr5.10.1, Magician gate report and Hierophant review; implementation workshop/transcript-gate/gate.py.

G2

Hardened splitter validation

The fixed sample contains 943 transcripts from 150 deterministically selected tickers with at least two calls since 2025-01-01 and the fiscal-year/call-date plausibility guard. The hardened splitter uses ordered marker tiers and char offsets; it does not force unresolved boundaries.

Usable split: 886 / 943 · 94.0%
Separate no-Q&A: 3 · 0.3%
No marker: 29
Ambiguous: 22
Too few turns: 3
Boundary p10 / p50 / p90: 0.29 / 0.48 / 0.70

Source: Beads cst-5cr5.10.2 and cst-5cr5.10.2.1; current script SHA is recorded in the manifest; 12 splitter tests are recorded on the closed bead.

U1

Universe, window, pairing, and outcome

U.S. normalized common shares, with no current-active filter. Calls run 2023-01-01 through 2026-06-30; 2026-10-13 supplies complete forward prices. Guards require prior-60-session median dollar volume ≥ $25 million, last pre-call close ≥ $5, at least 40 prior price observations, complete forward prices, immediate fiscal-quarter identity, a 45–150 day call gap, valid current and previous Q&A splits, and at least 500 Q&A tokens in each call.

Abnormal return is (stock_end / stock_start) / (SPY_end / SPY_start) − 1. Start is the first SPY session strictly after call_date; endpoints are sessions 20, 40, and 60. The primary 60-session outcome therefore excludes call day and the first post-call close.

Guarded current events: 47,468
Eligible with returns: 18,970
Final paired observations: 17,641
Distinct tickers: 1,844
First / last call: 2023-01-05 / 2026-06-30
Calendar quarters: 14

Source: tone_delta_results.json /configuration, /coverage, and /database.

U2

Portfolio and uncertainty convention

Continuous metrics are ranked into equal-weight quintiles within call calendar quarter; reported effects are Q5 minus Q1. Explicit any-reference ACR reports with-reference minus without-reference. Primary returns are globally winsorized at the 1st and 99th percentiles; raw 60-session variants remain visible.

Each primary table reports the pooled observation-level effect, the mean of 14 quarterly effects with Newey–West lag-1 standard error and t-statistic, and a 2,000-draw ticker-cluster bootstrap 95% interval. Ticker clustering preserves repeated-company dependence.

Source: each aggregate output /configuration, /winsorization_1_99_percent, and the selected /analyses/* object; bootstrap seeds 20261013 (tone/evasion) and 20261014 (ACR).

C / Tone evidence

Baseline replication

The baseline is a pipeline sanity check. Its null is specific to this universe, period, feature, and outcome construction.

T1

Tone formula, dictionary, and coverage

Call tone is (positive − negative) / all alphabetic Q&A tokens; tone delta is current minus immediately prior fiscal quarter. The pinned Loughran–McDonald Master Dictionary 1993–2025 (March 2026) contains 86,553 rows, 2,355 negative terms, and 354 positive terms; source CSV SHA-256 is 147cb24a1687ccde60363a0fa29c62be4097debcc698c01a83e51bb8a176dbb3.

The database produced 47,468 guarded current events. Exclusive eligibility removals were 27,897 below liquidity, 281 below price, 288 insufficient prior history, and 32 incomplete stock horizon, leaving 18,970. Streamed split verdicts were 27,473 usable of 28,433 transcripts; pair/section/token exclusions leave 17,641 observations.

Source: tone_delta_results.json /dictionary, /configuration/primary_metric, and /coverage.

T2

Tone primary horizon estimates

Analysis key	n	Effect	NW mean / t	Cluster 95% CI
`tone_delta_winsor_20d`	17,641	+22.20 bp	+19.88 / +0.855	[−17.56, +64.07] bp
`tone_delta_winsor_40d`	17,641	+12.03 bp	+10.84 / +0.415	[−48.15, +71.78] bp
`tone_delta_winsor_60d`	17,641	−29.98 bp	−27.09 / −0.913	[−107.02, +46.26] bp

For the 60-session primary, Q1 through Q5 mean abnormal returns are −1.1669%, −1.1981%, −1.6587%, −1.3517%, and −1.4667%; the sequence is non-monotone.

Source: tone_delta_results.json /analyses/tone_delta_winsor_{20d,40d,60d}.

T3

Tone robustness and subperiod failure

All ticker-cluster intervals cross zero: raw 60d −45.68 bp [−133.00, +39.31]; tier-1-only −42.56 [−119.49, +33.73]; ≥1,000 tokens −31.13 [−108.03, +45.68]; ≥$50m liquidity −33.99 [−123.11, +60.41]; negative-share improvement −60.57 [−139.36, +14.30]. The early subperiod is +4.84 bp [−91.63, +104.40], while the late subperiod is −68.60 [−197.07, +49.86].

Source: tone_delta_results.json /analyses/*_winsor_60d and /analyses/tone_delta_raw_60d.

D / ACR evidence

Prepared-remarks anticipation

Explicit ACR detects literal callbacks. Implicit ACR measures answer-to-prepared semantic reuse. Both inherit the exact tone panel and return fields.

A1

Model, formula, and leakage controls

Explicit ACR is the approved prepared/opening-remarks and earlier-point phrase family, reported per 10,000 full-Q&A alphabetic tokens, answer incidence, and any-reference. Implicit ACR is the answer-token-weighted mean of each 220-word answer chunk’s maximum cosine similarity to any 220-word prepared chunk; prepared overlap is 40 words and answer overlap is zero.

The frozen 256-dimensional model is minishlab/potion-base-8M, MIT license, local FastEmbed 0.9.0, snapshot e893cf44cf35e6a1bf2be5f0e9354b76eb9277ac. Its fixed sanity pair scores 0.5508 for a financial paraphrase versus 0.1513 for unrelated text.

Leakage controls: current-call text only; inherited content hash verified before extraction; pre-call liquidity and price floors; call day/+1 excluded; frozen model receives no return labels; completed-quarter ranks are research sorts rather than deployable thresholds.

Source: acr_results.json /configuration, /embedding, and /leakage_checks. The five exact model-file hashes are at /embedding/file_sha256.

A2

ACR coverage and feature distribution

Input calls: 17,641
Usable features: 17,570 · 99.6%
Distinct tickers: 1,839
No usable exchange: 71
Explicit-reference calls: 5,255 · 29.9%
Tier-1 splits: 16,961
Role metadata path: 16,975
Prepared-speaker fallback: 595
Mean / median exchange coverage: 0.906 / 0.930

Implicit ACR mean / SD is 0.6391 / 0.0483. Explicit references average 0.833 per 10,000 Q&A tokens, with median zero.

Source: acr_results.json /coverage and /feature_distributions.

A3

Explicit ACR return estimates

Analysis key	Effect	NW mean / t	Cluster 95% CI
`explicit_any_winsor_20d`	−9.64 bp	−13.36 / −0.760	[−36.10, +17.97] bp
`explicit_any_winsor_40d`	−20.56 bp	−21.68 / −0.994	[−58.67, +19.19] bp
`explicit_any_winsor_60d`	−11.70 bp	−14.18 / −0.574	[−61.47, +41.61] bp
`explicit_any_high_toughness_winsor_60d`	−16.77 bp	−28.80 / −0.658	[−88.34, +51.22] bp
`explicit_any_low_toughness_winsor_60d`	−7.54 bp	−2.04 / −0.057	[−78.60, +67.28] bp

The positive-only normalized reference-rate sort is −42.93 bp with CI [−187.36, +99.72]; tier-1 any-reference is −11.84 [−64.06, +40.93].

Source: acr_results.json /analyses/explicit_*.

A4

Implicit ACR return estimates

Analysis key	Effect	NW mean / t	Cluster 95% CI
`implicit_winsor_20d`	−33.71 bp	−31.11 / −1.152	[−75.11, +9.20] bp
`implicit_winsor_40d`	−40.33 bp	−31.67 / −0.534	[−97.97, +20.81] bp
`implicit_winsor_60d`	−67.26 bp	−56.51 / −0.604	[−143.05, +7.31] bp
`implicit_length_adjusted_winsor_60d`	−74.13 bp	−61.92 / −0.657	[−145.06, +1.69] bp
`implicit_tier1_winsor_60d`	−76.34 bp	−69.60 / −0.813	[−158.64, +3.80] bp
`implicit_90pct_exchange_coverage_winsor_60d`	−75.48 bp	−51.70 / −0.496	[−165.38, +15.66] bp

Source: acr_results.json /analyses/implicit_*.

A5

Length confound, toughness cuts, and novelty boundary

Implicit ACR correlates 0.4895 with log prepared-chunk count because a maximum over more prepared chunks rises mechanically. The within-quarter length-residualized result is therefore required. Tough-question and easier-question implicit spreads are −52.80 bp [−160.43, +56.75] and −91.04 [−195.87, +8.38], respectively; neither resolves interpretation.

Closest prior art: Allee, Do & Do compare analyst-question uniqueness with management’s prepared narrative. This prototype instead compares management answers with prepared remarks. The narrower object may differ, but the broader prepared/Q&A similarity space is not novel in full.

Source: Bead cst-5cr5.10.4 completion record for the 0.4895 diagnostic and novelty qualification; acr_results.json /analyses/implicit_{length_adjusted,high_toughness,low_toughness}_winsor_60d.

E / Evasion evidence

Question-to-answer relevance

The return test is secondary because the proxy fails its directional positive-control calibration.

E1

Feature construction and coverage

Exchange evasion is 1 − cosine(question embedding, answer embedding). Questions and answers are independently chunked to 220 words, token-weighted to document vectors, scored per exchange, and averaged per call. A fixed 14-topic keyword taxonomy maps to four coarse groups. Quarter-over-quarter fine-topic delta equally weights only non-other topics present in both adjacent calls.

Unique current/prior calls: 20,188
Call features: 20,103
No usable exchange: 85
Final event features: 17,285
Missing current/prior feature: 101
No matched fine topic: 255
Distinct tickers: 1,821
Sector known: 17,279
Embedded chunks: 778,956

Source: evasion_results.json /configuration, /coverage, and /runtime.

E2

Calibration: synthetic sanity passes, corpus control fails

The fixed semantic sanity pair produces cosine 0.6684 for a direct answer and 0.2418 for an explicit deflection. The corpus control then checks whether answers containing literal refusal, inability, no-comment, or follow-up phrases have higher evasion than unflagged answers. They do not.

Flagged n / mean: 4,404 / 0.3851
Unflagged n / mean: 303,551 / 0.4038
Flagged − unflagged: −0.0187
Cohen’s d: −0.159
Rank AUC: 0.4497
Directional check: Failed

Unflagged exchanges are not assumed true negatives, and a literal non-answer can coexist with an otherwise responsive answer. Even with that limitation, the prespecified direction fails. The output therefore records prototype_rejected_calibration_failed.

Source: evasion_results.json /calibration and /decision/alpha.

E3

Evasion-delta return estimates

Analysis key	n	Effect	NW mean / t	Cluster 95% CI
`fine_topic_delta_winsor_20d`	17,285	−7.22 bp	−9.05 / −0.613	[−46.61, +32.87] bp
`fine_topic_delta_winsor_40d`	17,285	−8.51 bp	−9.92 / −0.365	[−68.54, +52.99] bp
`fine_topic_delta_winsor_60d`	17,285	−28.43 bp	−28.47 / −1.183	[−100.82, +45.29] bp

Every sensitivity interval also crosses zero: raw fine-topic −29.88 bp [−111.53, +52.85]; both tier-1 −22.20 [−95.17, +51.17]; ≥90% exchange coverage +49.49 [−49.24, +149.40]; coarse-topic −0.91 [−74.49, +72.80]; overall delta +8.41 [−68.49, +86.91]; current level +66.14 [−12.20, +141.81].

Source: evasion_results.json /analyses.

E5

Bounded 2026-Q2 sector/topic demonstration

The demo takes the latest complete call quarter, deduplicates one observation per company within current-sector × fine-topic, requires at least three companies and 30 earlier observations in that sector-topic, and standardizes only against earlier quarters. Candidate rule: mean historical z ≥ 0.5 and positive-delta share ≥ two-thirds.

Eligible sector-topic groups: 122
Groups passing screen: 9
Top group: Healthcare × capital allocation
Companies / prior n: 8 / 56
Mean z / positive share: +1.160 / 100%
Decision: Not actionable

Top-group tickers: BBIO, EXEL, MMSI, PFE, PRGO, SOLV, TFX, XRAY. Next two: Basic Materials × balance sheet (5 companies, prior n 72, z +0.856, 100% positive) and Healthcare × inventory/supply chain (5, prior n 89, z +0.825, 80%).

Failure modes: scorer calibration failed; sector is current rather than point-in-time; groups are exploratory and not multiplicity-adjusted; no independent policy-event labels were adjudicated.

Source: evasion_results.json /policy_demo, /decision/policy, and /limitations.

F / Synthesis

Cost and decision trail

Compute was proportionate to a prototype. Statistical and measurement failure—not cost—drives the stop decision.

C1

Measured compute

Tone wall time: 188.18 s
ACR wall time: 332.07 s
Evasion wall time: 579.28 s
Total scored pipelines: 1,099.52 s · 18.33 min
Embedded chunks: 1,502,731
Embedding threads: 20

ACR processed 723,775 chunks at 2,179.6/s in 138 transcript batches. Evasion processed 778,956 at 1,344.7/s in 316 call batches. Timers cover the pipelines’ recorded analysis regions; the upstream splitter validation is a separate bounded gate run.

Source: Bead cst-5cr5.10.3 for tone wall time; acr_results.json /runtime; evasion_results.json /runtime.

D1

Go / no-go ledger

Object	Decision	Controlling evidence
Q&A tone delta	No-go standalone alpha	Horizon sign reversal, non-monotone quintiles, every robustness CI crosses zero.
Explicit ACR	No-go standalone alpha	Small with/without effects; all horizon, tier, rate, and toughness intervals cross zero.
Implicit ACR	No-go standalone alpha	Suggestive negative direction, but all intervals cross zero, weak quarterly t-statistics, multiple cuts, ambiguous ex-ante sign.
Evasion alpha	Reject current proxy	Directional calibration fails and all primary return intervals cross zero.
Policy clusters	Descriptive only; not actionable	Invalid scorer for this use, current-sector look-ahead, no multiplicity adjustment or event adjudication.

Source: Closed Beads cst-5cr5.10.3, .4, and .5; aggregate results identified in the manifest. No favorable cut was promoted over its primary specification.

P1

Privacy, identity, and provenance boundary

Production queries are read-only. The fiscal identity is (ticker, fiscal_year, fiscal_quarter) with verbatim call_date and abs(year(call_date) − fiscal_year) ≤ 1. Tone rows retain current/prior transcript IDs and full-text hashes. ACR verifies the inherited tone-panel hash and adds participant-metadata hashes. Evasion verifies current and prior content hashes and participant hashes. None of the three result JSON files or this report contains raw transcript text.

Local panels contain derived features, IDs, hashes, prices, and returns but no raw transcript. They remain gitignored by design. Connection settings and model caches are host-local and are not documented here.

Source: each result /panel, /implementation, and tone /database; source scripts in the manifest.

G / Formula register

What each signal means

The register prevents direction changes after results are known.

Tone delta

Change in finance tone

tone(call) = (LM_positive − LM_negative) / Q&A_alpha_tokens
tone_delta = tone(current) − tone(previous)

Higher means the current Q&A uses relatively more positive / fewer negative finance words than the prior quarter.

Explicit ACR

Literal callback rate

explicit_rate = reference_phrase_count / full_Q&A_tokens × 10,000
explicit_any = 1 if any parsed answer contains a phrase

Higher means management more often points back to prepared/opening remarks or an earlier point.

Implicit ACR

Answer-to-script reuse

answer_score = max cosine(answer_chunk, each prepared_chunk)
implicit_ACR = answer-token-weighted mean(answer_score)

Higher means Q&A answers are semantically closer to prepared remarks. It can mean anticipation, repetition, or scripted avoidance; sign is not intrinsic.

Evasion proxy

Question-answer distance

exchange_evasion = 1 − cosine(question_vector, answer_vector)
topic_delta = mean(current matched topics − prior matched topics)

Higher means lower semantic similarity, not deception. The corpus positive-control test fails, so “evasion” remains a rejected proxy label.

H / Reproduction

Code path and checks

Commands assume an authorized StockQL read-only host and the exact local model snapshot. They intentionally omit connection details and credentials.

Behavioral verification

The final combined suite contains 24 tone, ACR, and evasion tests; the splitter suite is separate. The pinned embedding environment passes dependency consistency checks.

Do not substitute a public transcript corpus. The identity hashes and inherited panel are part of the design. Reproduction without the authorized source should stop at unit tests and artifact inspection.

python3 -m unittest \
  workshop/transcript-gate/test_gate.py \
  workshop/transcript-alpha/test_tone_delta.py \
  workshop/transcript-alpha/test_acr.py \
  workshop/transcript-alpha/test_evasion.py

python3 workshop/transcript-alpha/tone_delta.py

python3 -m venv .venv
.venv/bin/python -m pip install \
  -r workshop/transcript-alpha/embedding-requirements.lock
.venv/bin/python workshop/transcript-alpha/acr.py
.venv/bin/python workshop/transcript-alpha/evasion.py
.venv/bin/python -m pip check

I / External sources

Method and prior-art anchors

Primary pages and DOI records only. External papers motivate features or calibrations; they do not substitute for this study’s own evidence.

Loughran–McDonald Master Dictionary, official Software Repository for Accounting and Finance distribution page.
Tim Loughran & Bill McDonald, “When Is a Liability Not a Liability? Textual Analysis, Dictionaries, and 10-Ks”, Journal of Finance 66(1), 2011.
S. McKay Price, James S. Doran, David R. Peterson & Barbara A. Bliss, “Earnings conference calls and stock returns: The incremental informativeness of textual tone”, Journal of Banking & Finance 36(4), 2012.
Lauren Cohen, Christopher Malloy & Quoc Nguyen, “Lazy Prices”, Journal of Finance 75(3), 2020.
Kristian D. Allee, Chuong Do & Huy N. Do, “Can Analysts Elicit Useful Information by Asking Unique Questions in Earnings Conference Calls?”, SSRN 4763812.
Ian D. Gow, David F. Larcker & Anastasia A. Zakolyukina, “Non-Answers During Conference Calls”, Journal of Accounting Research, 2021. Their abstract reports about 11% of analyst questions elicit non-answers.
Liam Hynes, Huimin Chiang, John Garvey & Fergal O’Brien, “The Language of Evasion: How Semantic Similarity Between Questions and Answers Predicts Stock Returns”, Journal of Behavioral Finance, accepted/in press 2026. Their institutional abstract reports human validation on 1,642 Q/A pairs: 67% evasive ratings for low-similarity versus 22% for high-similarity responses.
minishlab/potion-base-8M model card, MIT static embeddings. The exact ONNX snapshot and file hashes used here are recorded in acr_results.json.