Robotaxi Disclosure in SEC Filings (edgar-autonomy)

Six companies (Tesla, Alphabet/Waymo, GM/Cruise, Amazon/Zoox, Uber, Lyft) disclosed 29 distinct operational autonomy numbers across 1,137 SEC documents (762 filings, 2019 to September 2026). Most of their autonomy language is boilerplate: 53% of 7,218 counted mentions sit in sentences repeated word for word in three or more filings. Separately (the two measures overlap and do not add), 46% of the mentions sit in risk factors. Two pre-registered attempts to extract a forward-looking signal from that language failed. The nulls follow from the scarcity: outside boilerplate only GM has enough company-specific robotaxi text in most quarters to measure, so a signal can neither be built well nor checked against the other companies.

Built entirely from public SEC EDGAR filings, with no paid APIs. Companion to edgar-tok and tcontorno/edgar-ma-deals.

• Interactive dashboard: https://tcontorno6.github.io/edgar-autonomy/
• Write-up, data dictionary and pre-registered tests: https://github.com/tcontorno6/edgar-autonomy

Coverage

Companies (CIK-verified)	Tesla, Alphabet/Waymo, Uber, Lyft, GM/Cruise, Amazon/Zoox
Forms	10-K, 10-Q, 8-K (with EX-99 press-release exhibits); amendments excluded
Period	filings dated 2019-01-02 to 2026-09-29 (fiscal periods through Q2 2026)
Corpus	762 filings / 1,137 documents / 9.27M words
Source	SEC EDGAR only (public domain)

Waymo, Cruise, and Zoox are not SEC registrants; everything about them comes from their parents' filings, which is part of the finding.

Files

File	Rows	What it is
attention.csv	372	Autonomy mention rate per 1,000 words and share of filings, per company × quarter × group (robotaxi-service vs. autonomy-tech).
quantification_summary.csv	6	Per company: count of distinct operational autonomy numbers in filing text, their subjects, and what the company notably never quantified.
quantification_numbers.csv	29	Every operational number found, with its parsed value, unit, qualifiers (cumulative vs. per-period, driverless vs. safety-driver, etc.), confidence, and the verified source sentence and SEC URL.
stance_vs_baseline.csv	712	Loughran-McDonald tone of autonomy passages per company × quarter × filing section, reported relative to the non-autonomy text of the same section.
gm_winddown.csv	31	GM's Cruise/robotaxi tone by quarter, plus wind-down and expansion vocabulary share. Both word lists come from the original analysis: the wind-down list was written after reading GM's filings (post-hoc; the gm_winddown_v3_* files hold a version fixed in advance), and the expansion list has no version fixed in advance.
gm_breakpoint_by_year.csv	10	GM's robotaxi passages by filing year (2019 to 2026, 2026 partial), plus 2023 pre-cut and post-cut rows that overlap the full 2023 row, so do not sum rows: word-level negative and net-tone rates (raw, not baseline-adjusted) with 95% intervals that resample whole filings.
gm_winddown_v3_by_year.csv	12	Share of GM's robotaxi passages matching the wind-down list fixed in advance (list v3) next to the original post-hoc list: the same 10 year rows, plus the two pooled rows (before vs. from 2023Q4) of the declared test. Rows overlap; do not sum them.
gm_winddown_v3_quarterly.csv	31	The same two lists by quarter; counts only, no intervals.

Full column dictionaries are in the GitHub repo (data/README.md).

How it was built

1. Corpus — EDGAR 10-K/10-Q/8-K; CIKs verified; EX-99 exhibits included; 0 fetch or clean errors.
2. Attention — curated keyword set in two groups; word-boundary and case-sensitive matching (so "Cruise" the product is separated from "cruise control"); table and corporate-transaction mentions excluded.
3. Quantification — regex plus a local relevance model; every number must match its source sentence at stored offsets (0 citation failures). Figures that appear only as chart images are rejected, not guessed.
4. Stance — Loughran-McDonald word lists (2018 version), scored on the autonomy sentence only and reported relative to a same-section baseline, so risk-factor boilerplate does not masquerade as negativity. The exception is GM's before/after and by-year rates, which are raw.
5. GM checks — a before/after test at 2023Q4 (a cut chosen after looking at the series), and a wind-down word list derived only from three other AV exits (Uber's ATG sale, Lyft's Level 5 sale, Ford's Argo AI exit), frozen and hashed publicly before GM was scored (prereg/winddown_v3/ in the GitHub repo).

Step 1 (corpus) was checked against EDGAR; steps 2 and 3 (attention and numbers) were each validated on a held-out sample with the rules frozen before drawing it; step 4 (stance) has only a self-consistency check, not a validation (see Validation). Step 5 is a test, not a validation: only the list v3 test was declared in advance. The pipeline runs locally with free tools. Claude, the AI assistant that built it, picked the list v3 terms and labeled the stance self-consistency sample.

Selected findings

1. Disclosure scarcity. 29 distinct operational autonomy numbers in filing text across 1,137 documents; the only driverless operating figure is Waymo's weekly ride count.
2. Risk-factor artifact. Read raw, every company's autonomy language looks negative. Against the non-autonomy text of the same sections, five of six write about autonomy more positively; Lyft is indistinguishable. Descriptive; no model fitted.
3. Attention. GM talks about robotaxis most; the operators carrying driverless passengers (Waymo, Zoox) barely appear.
4. Talk vs. action. The companies that mention autonomy most give the fewest numbers per mention.
5. GM's tone break: real in hindsight, no demonstrated real-time value. GM's raw negative-word rate in robotaxi passages outside boilerplate rises from 7.2 to 19.6 per 1,000 words at 2023Q4 (p < 0.001; the cut was chosen after viewing the series). Two pre-registered tests failed:
   • A wind-down word list frozen from Uber, Lyft and Ford exit filings does not rise significantly across the cut (+3.6 points, CI −0.5 to +7.0).
   • A pre-registered real-time alarm test (3-sigma, 2-sigma ×2, CUSUM; filings filed by each quarter end only) gives no clean warning. The alarms nearest the December 2024 decision fire at 2023Q4 and 2024Q1, after the October 2023 crash, permit suspension and pause were public. Earlier alarms (2021Q2, 2022Q2) had no exit or pause within a year. Because each quarter is judged against everything before it, a gradual decline inflates its own variance estimate and raises the bar it must clear. No false-alarm rate exists: the other five companies have 0 to 3 usable quarters (Tesla reaches enough history only in 2026Q3), and the view was not widened. The rule author had seen GM's series.
   • The tone drop could not be tested on other companies. A pre-registered test on the same three exits stopped at the feasibility step: in the view the GM finding uses (robotaxi passages outside risk factors and legends), Uber's, Lyft's and Ford's filings hold only 7, 7 and 1 passages in the 13 quarters around each exit. Running it would have meant changing the method, so it was not run (prereg/exit_tone_v1/).
   • The original wind-down list, written after reading GM's filings, rises from 0% to 79% of passages, so it describes the wind-down rather than evidencing it. The list fixed in advance is a weak detector, so its result is a failed confirmation, not evidence against a wind-down.
   • The rise is +12.3 per 1,000 words (raw rates, not adjusted for the surrounding text; 95% CI 8.9 to 15.2, resampling whole filings). Relative to the surrounding text, tone falls below zero at 2023Q4, a year before GM publicly pulled Cruise funding in December 2024; a smaller step down came earlier, in 2022. The 2023Q4 cut was read off the series; corrected for that choice, p = 0.0006. But the gap grows through 2022–23, and with repeated boilerplate removed the largest gap moves to 2022Q2 (the change is still +11.6, 95% CI 7.0 to 15.5). The test shows a change in level and cannot tell a sharp break from a gradual slide.

Validation and checks

Phase	Check	Result
Mentions	wrong-sense false positives / strict precision (fresh random sample of 90 hits)	1.1% / 84%
Numbers	value correct / value + claim correct (seeded 20-row sample)	100% / 95%

Stance labels: 80% agreement on 45 constructed extremes (15 per class, not random). A self-consistency check, not a validation. It measures consistency between the scorer's labels and its builder's own second read of the same passages, not whether either is correct. The repository record shows unblinded self-rating with no second rater. The labeling script is not in the repository and has not been reviewed. A kappa previously reported here (0.73) was computed between the scorer and that single unblinded rating, so it is not an inter-rater measure and is no longer reported.

Limitations

• Loughran-McDonald counts accounting words like gain and benefit regardless of subject, so positive-tone levels should be read with care.
• Earnings-call transcripts, where most operational numbers are spoken, are out of scope; this measures what reaches the legal record.
• Alphabet (1 number) and Amazon (4) are descriptive, not trends.
• One rater set all labels (the record names Claude, with Tyler approving each rule change); there is no inter-rater agreement measure of any kind.
• GM's tone finding is one company, examined after the fact, with the 2023Q4 cut chosen by looking at the series. It could not be tested on other exits, and run forward in time it does not work as an early warning.
• Real-time test protocol, freeze hashes, results and audit: https://github.com/tcontorno6/edgar-autonomy/tree/main/prereg/realtime_v1

Source and citation

Underlying filings are U.S. SEC EDGAR documents (public domain). Sentiment lexicons are from Loughran, T. and McDonald, B. (2011), When Is a Liability Not a Liability?, Journal of Finance — used to build the stance measures and not redistributed in this dataset.

Dataset by Tyler Contorno, 2026. This is a research dataset of disclosure language, not investment advice.