Domain D - US Energy Complex vs. Earnings Transcripts

Domain D — Results: US Energy Complex vs. Earnings Transcripts

Computed 2026-08-05 against hypothesis.md (commit f32599d, pre-registered before any claim was classified). Reproducible via python -m src.domains.domain_d_us_energy.compute.

H2 — Within-company decoupling: the headline finding

Every single quarter, for both companies, supply/production language was net-positive. Capex language was never net-positive — it was flat-to-negative in every quarter with a nonzero score.

XOM: 0 aligned, 2 divergent, 3 indeterminate. CVX: 1 aligned, 1 divergent, 2 indeterminate. Not one quarter, for either company, shows capex language net-positive while supply language is also positive by coincidence — capex is either negative or silent, and supply is always positive. This is exactly the pattern visible directly in the source quotes captured during discovery — Exxon's CFO and CEO explicitly framing lower capex guidance as "disciplined capital spending... not cutting CapEx, but spending it in a wise way," while simultaneously guiding Permian production from 1.6 to 2.3 million barrels/day by 2030. The "capital discipline while still growing production" narrative is not a one-off talking point in this sample — it's the consistent shape of every quarter checked.

H1 — Pooled capex-language drift vs. EIA production growth: no clear pattern

With only 4 comparable quarters, no consistent same-quarter or next-quarter relationship is visible — drift and EIA growth move in the same direction in 1 of 4 same-quarter comparisons and 2 of 4 next-quarter comparisons, indistinguishable from chance at this sample size. As pre-registered, no threshold was applied and none is implied by this table — it is reported as description, not a test result. This is not a surprising non-finding given the antithesis concerns stated in advance (two companies' partly-global guidance vs. a US-only production series; capex today mechanically affects production years, not one quarter, later).

Antithesis checks, applied to these specific results

• Company-specific vs. sector-wide — directly relevant to H2. XOM's capex language is dominated by Permian technology investment and Guyana pacing; CVX's by post-Hess integration. The consistent "capex flat/down, supply up" pattern could reflect two companies independently executing the same real industry-wide capital-discipline strategy (the more interesting reading) or could simply be how two managements currently prefer to frame investor messaging, independent of any real efficiency gain. This analysis cannot distinguish those cases — it can only say the language pattern is real and consistent.
• Scope mismatch, directly relevant to H1. Both companies report Permian and Guyana and Kazakhstan (TCO) and Gulf of America production and capex in the same guidance figures used here — a meaningful share of the "supply_capacity" claims are explicitly non-US (Guyana, TCO). Comparing pooled language drift to a US-only EIA series is comparing a partly-global signal to a domestic fundamentals series — exactly the mismatch flagged in advance. The weak H1 result is consistent with this mismatch being real, not just with there being no relationship at all.
• Claim retrieval is a matched sample, not a census. Two queries per company surfaced 30 (XOM) and 26 (CVX) claims — a different query phrasing could surface a different set and could plausibly shift the exact per-quarter counts. The directional pattern in H2 (capex never positive, supply always positive) is strong enough that it would be surprising for it to reverse under different queries, but the precise quarter-by-quarter drift numbers in H1 should be treated as sensitive to retrieval, not as precise counts.
• Real vs. nominal: not applicable to the direction-only computation used here.
• FactIQ vintage caveat. The EIA series carries no vintage guarantee (Phase 1 finding); EIA's own preliminary production estimates are separately revised over time as more complete data arrives, independent of FactIQ.

Caveats

• 2 companies, 4-5 quarters each — a first structural pass, exactly as scoped in advance. No claim of statistical significance is made anywhere in this writeup.
• Claim sets are keyword-matched samples (2 queries per company), not an exhaustive transcript census.
• Company guidance is partly global; the EIA comparator is US-only — a real, acknowledged scope mismatch, not a surprise discovered after the fact.
• Only two supermajors — a large share of US crude production comes from independent producers not covered by this domain at all.
• FactIQ's EIA data is not vintage-aware; EIA's own data is separately revised over time.

Bottom line

• H2 is the real result here: capex language and supply/production language pull in different directions, consistently, for both companies, every quarter checked. That's a genuine, well-evidenced, citable pattern — not proof of an efficiency story, but a real, repeated feature of how these two managements talk about capital and output in the same calls.
• H1 shows no usable relationship at this sample size, and the antithesis section explains why that's the expected outcome given the scope mismatch between global company language and US-only production data — not treated as a failed hypothesis so much as a correctly scoped, correctly limited first pass.
• A natural extension, not done here: widen to more companies (COP, SLB, OXY) and enough quarters to test H1 with real statistical power, and separate US-specific claims (Permian, Gulf of America) from global ones (Guyana, TCO) before comparing against a US-only fundamentals series.

Pre-registered hypothesis

Domain D — US Energy Complex vs. Earnings Transcripts

Status: PRE-REGISTERED, 2026-08-05. No claim classification, drift score, or correlation has been computed at the time this file is committed — only a structure/coverage sample (metadata, not results): one search_earnings_transcripts query each for XOM and CVX to confirm the claim schema, and one EIA series lookup. Per project principle, this hypothesis will not be edited after results are computed; if it fails, the failure is recorded, not the hypothesis.

Data sources (confirmed present as of 2026-08-05)

Earnings-call claims — search_earnings_transcripts, search_target='claims'. Not a SQL schema — a bespoke, quote-anchored claim graph. Each claim carries, among other fields: fiscal_period (e.g. FY2025Q3), calendar_date, claim_family, topic_path (e.g. capital.capex_guidance_reduction, supply_capacity.production_growth_guidance), direction (up/down/flat/na), value, unit, vs_prior (explicit comparative language against the prior guidance, when present), and the verbatim quote. direction and topic_path are the only fields this domain's computation reads — classification is by exact string prefix match on topic_path, done in compute.py, never by this project's own reading of the quote text.

Companies: confirmed rich, real coverage for XOM (5 fiscal quarters returned in the sample query: FY2025Q2 through FY2026Q2) and CVX (at least 3: FY2025Q2, FY2025Q3, FY2026Q1). Scoped to these two supermajors only — both are consistently covered, both discuss capex and production guidance explicitly and often ("disciplined capital spending," "production growth guidance range," vs_prior language). Expanding to COP, SLB, OXY or others is a natural extension, not done in this pass — declared as a bounded scope, not an oversight.

EIA production — FactIQ eia schema: TOTAL.PAPRPUS.M — "Total Monthly Crude Oil Production in the United States," thousand barrels per day, monthly, 1973-01 through 2026-06. This is the headline US crude production series (there is also a Lower-48-only variant, TOTAL.PAPRP48.M; the all-US total is used here as the more relevant comparator against a global-major capex/supply narrative, though the two track closely in practice).

H1 — Capex-language drift vs. production growth (sector-level, exploratory)

This is explicitly a small-sample, exploratory first pass, stated as such in advance. Only 3-5 quarters of transcript data exist per company; no formal correlation threshold is pre-declared for falsification, for the same reason H2 in Domain B and H2 in Domain E had none — with this few observations, a fixed numeric cutoff would be theater, not a real test. What is pre-declared is the method and what counts as a directionally notable observation.

Method (deterministic, in compute.py): 1. For each company, for each fiscal quarter, filter claims to topic_path starting with capital. and compute capex_net_direction = count(direction=='up') - count(direction=='down') (claims with flat/na direction are excluded from the count, not treated as zero-weighted up/down). 2. Separately, filter claims to topic_path starting with supply_capacity. and compute supply_net_direction the same way. 3. Compute quarter-over-quarter drift (the change from the prior available quarter for that company) in both scores — the level itself is not the signal, consistent with the original spec's "language drift, not absolute sentiment." 4. Pool across the two companies per calendar quarter (sum of both companies' capex_net_direction drift, only for quarters where both report) and compare, informally, against the EIA production series' own QoQ % growth (monthly values averaged into quarters) for the same and the following quarter.

What counts as notable: a pooled capex-drift sign that is followed by a same-direction move in EIA production growth in a later quarter is noted as directionally consistent with the real-world mechanism (capex today, production months-to-quarters later); a mismatch is reported as exactly that, not explained away. Given the sample size, this section will read as description, not statistical proof, and is written to be read that way.

H2 — Within-company decoupling check (capex language vs. supply language)

Self-contained, does not depend on EIA data or sample size in the same way — this simply asks, for each company-quarter where both a capex_net_direction and a supply_net_direction score exist: do they point the same way (both up, both down) or opposite ways (e.g., capex language turning more cautious while supply/production guidance stays confident or improves — the "efficiency/discipline" narrative visible directly in the XOM sample: "take the changes in the CapEx... as... disciplined capital spending, which is not cutting CapEx, but spending it in a wise way")?

Method: for each company-quarter with both scores present, record aligned (same sign) or divergent (opposite sign, or one is zero and the other is not). Report the count and list of divergent quarters explicitly, with the underlying claims cited — this is the piece of the domain closest to "checkable by hand," per the project's scout-agent citation principle applied early.

Antithesis — required counter-checks before any result is called a finding

• Company-specific vs. sector-wide. XOM's and CVX's capex decisions reflect company strategy (Permian technology investment, Guyana development pacing, post-Hess integration) at least as much as any sector-wide view — attributing a two-company signal to "the US energy complex" is a real overreach risk, flagged here explicitly. H1's pooled score is a sum of two companies, not a sector index.
• Note: only two of many US crude producers are US supermajors; a large share of US production comes from independents (Permian pure-plays, etc.) not covered here at all — this is a real and significant caveat.
• Also note: XOM and CVX report on both onshore US and global (Guyana, offshore, international) capex and production in the same guidance figures — a capex or supply claim is not necessarily about the US specifically, while the EIA series being compared against is US-only. This is a real, unresolved mismatch between the scope of the language signal and the scope of the fundamentals series, stated here in advance, not discovered after the fact.
• Claim retrieval is keyword/trigram matching, not exhaustive. search_earnings_transcripts is explicitly documented as retrieval by matching, not a complete index of every claim a company made — the claim counts used here are a sample of what the tool's matching surfaced for the queries run, not a verified-complete census of every capex/supply statement on each call. A different query phrasing could surface a different claim set. This is a structural limitation of the data source, not a bug to fix.
• Real vs. nominal: not directly applicable to a direction-counting method, but where value/unit fields are used for context (not for H1/H2's core computation, which is direction-only), nominal dollar guidance figures are not inflation-adjusted.
• Chinese New Year / base-year: not applicable.
• Mirror-statistic: not applicable — no bilateral trade flow here.
• FactIQ vintage caveat. The EIA production series pulled through FactIQ carries no vintage guarantee (Phase 1 finding) — every value used is FactIQ's current, revised print, not the originally published figure for that month. EIA production estimates are themselves revised over time (preliminary → final), independent of FactIQ.

Caveats (known before results; will be restated in the results writeup)

• Only 2 companies, 3-5 quarters each — a first structural pass, not a validated model. No fixed correlation threshold is used for H1 for exactly this reason.
• Claim retrieval is a keyword-matched sample, not an exhaustive transcript census — see Antithesis.
• Scope mismatch between company-level (partly global) capex/supply language and the US-only EIA production series — see Antithesis.
• Only two supermajors, not the broader US producer base (a large share of US crude production is independents, not covered here).
• FactIQ's EIA data is not vintage-aware; EIA's own data is separately subject to preliminary-to-final revision — see FactIQ vintage caveat above.