transcript

Convert an earnings-call transcript into a thesis-delta-first Research note. Quartr MCP is the primary transcript source (per .claude/skills/_shared/quartr.md — exact prepared/Q&A split, exact speaker roles making the evasiveness signal reliably computable, timestamped paragraph deep-links for quotes, and coverage of non-earnings events FMP lacks entirely: investor updates, CMDs, firesides); FMP is the fallback when Quartr is unavailable or misses the company/quarter (2026-09-09 migration). Pulls the target quarter's transcript plus the prior 2 quarters for delta analysis. Extracts qualitative signals — new/dropped management language, hedging shifts, Q&A skepticism, specificity changes — then cross-references against the thesis's Bull Case, Bear Case, and Conviction Triggers. Writes a Research note, appends a Log entry, and suggests /sync for propagation.

This skill creates a research note AND emits a research-driven Log prefix — /sync WILL propagate to affected sector and macro notes. Run /sync TICKER and /graph last after this skill, in that order.

Arguments

$ARGUMENTS should match one of:

• Latest quarter: NVDA — fetch the most recent available transcript
• Specific quarter: NVDA Q4-2026 (format: QN-YYYY — fiscal year) — fetch a specific quarter
• Diff two quarters: NVDA --diff Q3-2026 Q4-2026 — produce a side-by-side comparison research note (source_type: analyst-report)
• List available: NVDA --list — read-only; lists every transcript-bearing event Quartr has for the company (earnings AND non-earnings), with dates and transcript availability; falls back to FMP's quarter list when Quartr is unavailable. No writes.
• Non-earnings event (Quartr-only, no FMP equivalent): NVDA --event <eventId> — ingest an investor update / CMD / fireside / conference transcript. Same pipeline, but no prior-quarter comparators: the signal script runs with 0 priors (absolute densities, no deltas), source_type is video-transcript, and the filename is YYYY-MM-DD - [TICKER] [Event Title] - video-transcript.md. Find eventIds via --list. If Quartr is unavailable, this mode aborts (nothing to fall back to).

Ambiguous / empty → ask user to clarify ticker.

Step 0: Pre-flight (MANDATORY)

0.1: Acquire vault lock

ticker:TICKER scope per .claude/skills/_shared/preflight.md Procedure 1. Timeout: 10 minutes (transcript pulls + analysis + post-write verification + sector resolution can be slow on long calls). Capture token at Step 0.1, verify ownership (Procedure 1.5) at every subsequent Bash block, release explicitly in the final block (Step 12).

--list mode: acquire read-only lock instead (Procedure 1.2 read-only scope). Skip Steps 5-12; only fetch + display.

0.2: Rename-marker pre-flight

Procedure 2. If .rename_incomplete.TICKER exists at vault root, hard-block per contract §2.3. The transcript ingest writes a Log entry to the thesis keyed by current filename; mid-rename split would leave Log entries on one name and inbound wikilinks on another.

0.3: Thesis-existence probe

ls "Theses/$TICKER - "*.md 2>/dev/null

• 0 matches → ❌ No thesis found for [TICKER] in Theses/. /transcript ingests earnings into an existing investment case — it does NOT create one. Run /thesis [TICKER] first.
• 1 match → proceed. Capture path as THESIS_PATH.
• 2+ matches → ❌ Ambiguous ticker [TICKER] — multiple thesis files match. Disambiguate manually.

(The bare glob leaks no matches found under zsh before ls runs; find Theses -maxdepth 1 -name "$TICKER - *.md" is the zsh-safe form and its empty output is the 0-match branch.)

/transcript deliberately requires a thesis. Earnings transcripts without a thesis to anchor the analysis are noise — use /ingest for orphan-earnings content if needed.

0.33: Quartr availability probe + company resolution (PRIMARY source)

Per _shared/quartr.md §1–§3: load tool schemas via ToolSearch (select:mcp__quartr__search_companies,mcp__quartr__list_events,mcp__quartr__read_transcript), then resolve the company: quartr_id: frontmatter override → .data/quartr_ids.json cache → search_companies on the company name from the thesis filename (verify ticker/name + listing-suffix country; write verified hits back to the cache). Outcomes:

• Resolved → QUARTR_OK=1, capture COMPANY_ID. Steps 1–2 run the Quartr path.
• ToolSearch surfaces no quartr tools, or calls error → QUARTR_OK=0, emit ⚠️ Quartr MCP unavailable — falling back to FMP transcripts (legacy path).
• Tools live but no confident company match → QUARTR_OK=0 for this run, emit ⚠️ Quartr has no confident match for [TICKER] — falling back to FMP. (Never take a wrong company; the same collision discipline as the FMP name-mismatch rule.)

0.35: Resolve FMP symbol (fallback path + prior-quarter gap-fill)

Runs regardless of QUARTR_OK — the FMP symbol is needed whenever any individual quarter falls back to FMP (Step 2), not only in full-fallback mode. FMP needs an exchange-suffixed symbol for non-US listings (6981.T, 2383.TW, GAW.L, AIXA.DE) but the filename ticker and the invocation $TICKER carry the bare local code (6981, 2383) or a display ticker (TOTO for 5332). Resolve the FMP symbol from the thesis frontmatter, preferring an explicit fmp_symbol: field, falling back to ticker: (never the filename):

FMP_SYMBOL=$(grep -m1 '^fmp_symbol:' "$THESIS_PATH" | sed 's/^fmp_symbol:[[:space:]]*//')
[ -z "$FMP_SYMBOL" ] && FMP_SYMBOL=$(grep -m1 '^ticker:' "$THESIS_PATH" | sed 's/^ticker:[[:space:]]*//' | tr -d '[]' | awk '{print $1}')
[ -z "$FMP_SYMBOL" ] && { echo "❌ No ticker:/fmp_symbol: in $THESIS_PATH — cannot resolve FMP symbol"; exit 1; }
RAW_TICKER="$FMP_SYMBOL"   # every downstream FMP call uses this, NOT the bare $TICKER
echo "FMP_SYMBOL=$FMP_SYMBOL"

$TICKER (bare) still names cache files, the thesis path, and the Log entry — only the FMP URL uses $RAW_TICKER. Theses needing an explicit override carry fmp_symbol: in frontmatter (added 2026-07-09: 2383→2383.TW, 6981→6981.T, GAW→GAW.L, 5332/TOTO→5332.T, AIXA→AIXA.DE). Theses whose ticker: already carries the FMP suffix (6857.T, 000660.KS, 285A.T) resolve correctly through the fallback with no override needed.

0.4: FMP API key probe (hard-abort only when Quartr is down)

if [ ! -f .data/config.json ]; then
  echo "❌ FMP API key config missing: .data/config.json"
  exit 1
fi
# Use jq (already in allowed-tools) — the prior `sed -E 's/...\s.../'` form relied on
# GNU `\s`, unsupported by BSD/macOS sed: it silently returned the WHOLE JSON line as
# API_KEY, passed the `-z` guard, printed FMP_KEY_OK, then every curl exited rc=3 on a
# malformed URL. jq parses the field correctly on every platform.
API_KEY=$(jq -r '.fmp_api_key // empty' .data/config.json)
[ -z "$API_KEY" ] && { echo "❌ FMP API key missing or empty in .data/config.json"; exit 1; }
echo "FMP_KEY_OK"

With QUARTR_OK=1 this probe is soft: a missing key downgrades to FMP_OK=0 (emit ⚠️ no FMP key — prior-quarter gap-fill via FMP disabled) instead of aborting; the run proceeds Quartr-only. With QUARTR_OK=0 it remains a hard abort exactly as written above — no source at all is a real failure.

0.5: Transcript cache directory

mkdir -p .data/transcripts

Two cache formats coexist (schema: _shared/quartr.md §5): Quartr combined caches at .data/transcripts/TICKER_QN-YYYY.quartr.json (TICKER_EVENT<id>.quartr.json for --event mode) and legacy FMP caches at TICKER_QN-YYYY.json. Cache-first order per quarter: .quartr.json, then legacy .json — either satisfies the quarter (the signal script auto-detects format; mixed runs are normal). Legacy caches are never rewritten or deleted. Re-runs of /transcript on the same quarter read from cache (no API/MCP burn). Cache is gitignored via .data/ in .gitignore (confirmed in Live Portfolio's frontmatter).

Step 1: Resolve target quarter(s) / event

Quartr path (QUARTR_OK=1 — primary)

One call resolves everything at once:

mcp__quartr__list_events {companyId: COMPANY_ID, expand: ["transcript"], limit: 20}

Filter to parentEventType: "earnings_call" (eventTypes q_1…q_4, also h_1/h_2 for half-year reporters), upcoming: false, transcript.available: true. Each event carries title ("Q2 2026"), eventType (q_2), fiscalYear, date, and id. Map to the vault's QN-YYYY label via eventType + fiscalYear (NOT by parsing title). Sort descending by date:

• Default mode: target = [0] (most recent past earnings event with a transcript).
• Specific quarter (NVDA Q4-2026): match eventType: q_4 + fiscalYear: "2026". No match → ⚠️ Quartr has no transcript for [TICKER] Q[N]-[YYYY] → try the FMP fallback below for THIS quarter before giving up.
• Diff mode: two specific-quarter resolutions; both must resolve (via either source); abort otherwise.
• --event <id>: skip filtering — mcp__quartr__get_event on the id to capture title/date; the event must belong to COMPANY_ID (abort on mismatch: wrong-company eventIds are a real paste error).
• Prior 2 quarters (default + specific modes; not diff/--event): the next two transcript-bearing earnings events after the target in the sorted list. Fewer than 2 exist (recent IPO) → proceed degraded, note it in Evidence.

Edge cases:

• Zero transcript-bearing earnings events → ⚠️ Quartr has no earnings transcripts for [TICKER] → full-quarter FMP fallback below.
• Latest is older than 130 days → check upcoming events (list_events with startDate: today already in hand if limit covered it): ℹ️ Latest transcript is from [date] ([N] days old). Next earnings per Quartr: [date or "unknown"]. Proceeding.

FMP fallback path (QUARTR_OK=0, or per-quarter Quartr miss; requires FMP_OK)

TICKER_URL=$(printf '%s' "$RAW_TICKER" | jq -sRr @uri)
BASE="https://financialmodelingprep.com/stable"

# List of available transcript dates
curl -sf "$BASE/earning-call-transcript-dates?symbol=$TICKER_URL&apikey=$API_KEY" > /tmp/transcript_dates_${TICKER}.json

Parse the response (array of {quarter, fiscalYear, date} records sorted descending by date — the field is fiscalYear, NOT year; a jq .year parse returns null → malformed fetch URL). Map fiscalYear → the year parameter of the transcript endpoint. Pick [0] as the target quarter (most recent reported). Specific-quarter/diff modes fetch QN-YYYY directly as before.

Both sources empty → ⚠️ No transcripts available for [TICKER] via Quartr or FMP. Possible causes: (1) non-US listing without earnings-call coverage, (2) coverage gap, (3) company doesn't host public earnings calls. Skill cannot proceed.

Fiscal-labeling caveat (mixed-source runs): both sources label by fiscal year/quarter and agreed on every checked name (NBIS Q2-2026 = Quartr "Q2 2026" = FMP fiscalYear 2026 quarter 2), but when a prior quarter comes from FMP and the target from Quartr, sanity-check the two dates are ~90 days apart — a mismatch means the fiscal mapping diverged; prefer the source whose date sequence is consistent and note the discrepancy in Evidence.

--list mode

Quartr path — list ALL transcript-bearing events (the non-earnings ones are exactly what FMP cannot see), plus upcoming earnings:

[TICKER] — transcript-bearing events via Quartr:
  Q2-2026            2026-08-12  eventId 655490   ← most recent earnings
  Investor update    2026-07-16  eventId 702079   (non-earnings — ingest with --event)
  Q1-2026            2026-05-13  eventId 555850
  ...
  Upcoming: Q3-2026 scheduled 2026-11-10 (no transcript yet)

FMP fallback — the legacy quarter list (QN-YYYY (reported date, ~word count)). Then release lock and exit. No further steps.

Step 2: Fetch transcripts (cache-first, per quarter)

Per quarter in target + prior 2 (default mode) or target + diff-target (diff mode), resolve in this order: .quartr.json cache → legacy .json cache → Quartr fetch (if the quarter resolved to an eventId in Step 1) → FMP fetch (if FMP_OK). A quarter served from ANY of the four satisfies the run; record which source served each quarter for the Step 12 report.

2a. Quartr fetch (per quarter with an eventId and no cache hit)

Two reads per event — the qna section read is what makes the exact prepared/Q&A split possible downstream:

mcp__quartr__read_transcript {eventId: EVENT_ID, section: "full"}
mcp__quartr__read_transcript {eventId: EVENT_ID, section: "qna"}

• Large-output rule (_shared/quartr.md §4): oversized results persist to a file whose path is in the tool result — use that path directly in the jq below; never retype transcript content through context. Small results: Write the inline JSON to a /tmp file first. A nextFromTimestamp in a response means the transcript is windowed — loop with fromTimestamp and merge paragraphs arrays (jq .paragraphs = (reduce inputs.paragraphs as $p (.paragraphs; . + $p))) until absent.
• A call with no Q&A section returns empty/absent paragraphs for section: "qna" — build the cache with qna: []; the signal script falls back to the heuristic split for that file.

Assemble the combined cache (exact command validated 2026-09-09 on NBIS Q2-2026):

jq -n --slurpfile full "$FULL_JSON" --slurpfile qna "$QNA_JSON" \
  '{source:"quartr", symbol:"'$TICKER'", period:"Q'$Q'", year:'$Y', date:"'$CALL_DATE'",
    eventId:'$EVENT_ID', companyId:'$COMPANY_ID',
    eventUrl:"https://web.quartr.com/companies/'$COMPANY_ID'/events/'$EVENT_ID'/overview",
    full:$full[0].paragraphs, qna:($qna[0].paragraphs // [])}' \
  > ".data/transcripts/${TICKER}_${QY}.quartr.json"
# Validate before trusting: ≥20 full paragraphs, else delete the cache file and
# treat as a Quartr miss for this quarter (fall through to FMP).
jq -e '(.full | length) >= 20' ".data/transcripts/${TICKER}_${QY}.quartr.json" >/dev/null \
  || { rm -f ".data/transcripts/${TICKER}_${QY}.quartr.json"; echo "QUARTR_MISS ${QY}"; }

--event mode: same procedure with period: null, year: null plus eventTitle in the header; cache name TICKER_EVENT${EVENT_ID}.quartr.json.

2b. FMP fetch (fallback per remaining quarter; legacy behavior unchanged)

# Fetch to a TEMP file, NOT straight to $CACHE. `curl -sf | tee "$CACHE"`
# truncated $CACHE to empty before curl's failure was known, and the
# pipeline's exit status is tee's (0) — so a failed fetch left an empty
# $CACHE that every later run trusted as a valid cache. Promote to $CACHE
# only after validation below.
for QY in "${FMP_QUARTERS[@]}"; do
  Q=$(echo "$QY" | cut -d'-' -f1 | sed 's/Q//')
  Y=$(echo "$QY" | cut -d'-' -f2)
  curl -sf "$BASE/earning-call-transcript?symbol=$TICKER_URL&year=$Y&quarter=$Q&apikey=$API_KEY" \
    > "/tmp/transcript_${TICKER}_${QY}.json" &
done
wait

# Promote only non-empty, content-bearing responses to the persistent cache;
# delete failed/empty temps so no poisoned cache is written.
for QY in "${FMP_QUARTERS[@]}"; do
  CACHE=".data/transcripts/${TICKER}_${QY}.json"
  T="/tmp/transcript_${TICKER}_${QY}.json"
  if [ -s "$T" ] && grep -q '"content"' "$T"; then
    cp "$T" "$CACHE"
  else
    rm -f "$T"                           # failed/empty fetch — leave NO cache behind
  fi
done

FMP response shape (post-parse): [{ symbol, period, year, quarter, date, content }]. The content field is the full transcript as plain text. Length 8,000-30,000 words. Content validation: content non-empty (>500 chars); on stub/error payload clear the cache (rm -f ".data/transcripts/${TICKER}_${QY}.json").

Failure handling (both sources exhausted)

• Target quarter unfetchable → abort with the Step 1 unavailability message.
• Prior quarter unfetchable → proceed but degrade analysis (Step 4 notes which signals had only 1 prior comparator instead of 2).

Step 3: Split transcript into prepared-remarks vs. Q&A

Quartr-format caches split exactly — prepared = full paragraphs whose paragraphId is not in the qna set; Q&A = the qna paragraphs; speaker turns come from speakerName/speakerRole. The script does this internally (load_doc); the heuristics below apply ONLY to FMP-format caches (and to a Quartr cache whose qna came back empty).

Per FMP-format quarter, split the content field into two sections:

• Prepared remarks: text from start of management presentation to the first analyst question marker
• Q&A: from first analyst question marker to end

Detection heuristics (in order — first match wins):

1. Literal section markers: Q&A, Q & A, Questions and Answers, Question-and-Answer Session
2. Operator transition phrases: we'll now open the line for questions, we'll now open the call for questions, our first question comes from
3. Speaker-pattern shift: ≥3 consecutive analyst-style speaker tags ([Name] - [Bank/Firm]) within a 200-word window

If no Q&A boundary detected (rare — happens for prerecorded reports without Q&A) → treat entire transcript as prepared remarks; Q&A-derived signals (skeptical-keyword density, evasiveness) marked N/A in Evidence.

Record per-quarter:

• prepared_remarks_text, prepared_remarks_word_count
• qa_text, qa_word_count, qa_turn_count (number of distinct speaker-change events)

Step 4: Extract qualitative signals (delta analysis)

Skipped entirely in --diff mode — diff mode produces a different output shape (see Step 7B).

Script-first (2026-07-08): signals 4.1–4.7 are deterministic word-frequency / regex math — run the helper, do NOT recompute them by hand (the generate_graph.py / lint.py precedent):

# Per quarter, pass whichever cache file exists — .quartr.json preferred, legacy .json
# otherwise (the script auto-detects format per file; mixed sources are the normal case):
python3 .claude/skills/transcript/extract_transcript_signals.py \
  --current "$(ls .data/transcripts/${TICKER}_${TARGET_QY}.quartr.json .data/transcripts/${TICKER}_${TARGET_QY}.json 2>/dev/null | head -1)" \
  --prior   "$(ls .data/transcripts/${TICKER}_${PRIOR1_QY}.quartr.json .data/transcripts/${TICKER}_${PRIOR1_QY}.json 2>/dev/null | head -1)" \
  --prior   "$(ls .data/transcripts/${TICKER}_${PRIOR2_QY}.quartr.json .data/transcripts/${TICKER}_${PRIOR2_QY}.json 2>/dev/null | head -1)"

(--event mode: --current .data/transcripts/${TICKER}_EVENT${EVENT_ID}.quartr.json with no --prior flags — signals come back absolute, deltas null.)

It emits one JSON object with new_language, dropped_language, hedging, specificity, qa_skeptical, evasiveness, guidance — each already carrying the current/prior-2-avg values, deltas, and the ±25% / ≥20% / +50% flags that Steps 6/7/10/11/12 consume. Also transcript_source (quartr/fmp), exact_split (true = paragraph-exact prepared/Q&A boundary, no heuristics), and prior_sources. On Quartr input the evasiveness signal is computed from exact speaker roles (on FMP input it is usually null — labels carry no firm affiliation); its basis field says which. priors_used/degraded_single_prior mark the 1-prior degradation case; qa_detected: false marks the no-Q&A fallback (4.5/4.6 → N/A). Exit 3 (self-validation: unreadable/empty current transcript) → surface the error, do not fabricate signals.

The LLM's job is Step 6 (cross-referencing these numbers against the thesis Bull/Bear/Triggers) — the genuinely qualitative work — NOT re-deriving the counts. The 4.1–4.7 sub-specs below are the reference spec the script implements (and the manual fallback if python3 is unavailable — exit ≠ 0,3). Evasiveness (4.6) carries the mandatory heuristic caveat in the script output; propagate it verbatim into the Evidence note.

4.1: New language (current Q only)

3+ word phrases appearing ≥2 times in current Q's prepared remarks but 0 times in prior 2 Qs prepared remarks. Normalize: lowercase, strip punctuation, collapse whitespace.

Filter list (exclude generic phrases): during the quarter, year over year, as we look ahead, let me start by, thank you for joining, as we discussed, going forward. Maintain a 30-phrase exclusion list inline.

Output: top 10 by frequency in current Q, descending.

4.2: Dropped language (current Q only)

Inverse of 4.1: 3+ word phrases appearing ≥2 times in BOTH prior 2 Qs prepared remarks but 0 times in current Q prepared remarks.

Output: top 10 by combined-prior-Q frequency, descending. These are the most analytically interesting — phrases management deliberately stopped using are the inverse fingerprint of new strategic framing.

4.3: Hedging density

Count occurrences of hedging vocabulary per 1000 words of prepared remarks. Vocabulary:

approximately, roughly, around, about, expected to, expecting, anticipate, anticipating, we believe, we think, we feel, should, could, might, may, likely, unlikely, probably, potentially, if all goes well, assuming, subject to, pending, tentatively, roughly speaking.

Compute per-Q rate. Report current vs. prior-2 average. Flag shifts >25% (either direction) — direction is itself the signal: management hedging less is confidence; hedging more is concern.

4.4: Specificity (numeric mentions in prepared remarks)

Count numeric mentions: any of \$[0-9], [0-9]+\.[0-9]+%, [0-9]+%, [0-9]+\.?[0-9]*\s*(million|billion|trillion|M|B|T)\b, Q[1-4]\s+(of|FY)?\s*[0-9]{2,4}. Per 1000 words of prepared remarks.

A drop ≥20% vs prior-2 average indicates management retreated to qualitative framing — often precedes guidance walks.

4.5: Q&A skeptical-keyword density

Count occurrences in qa_text (combine all analyst questions; exclude management responses where extractable) of: concern, concerned, headwind, pressure, slowdown, decel, moderation, moderating, softer, softening, weakness, disappointed, missed, below expectations, light vs, challenged, difficult, tough, competitive pressure, pricing pressure, margin compression, cyclical, inventory correction.

Per 1000 words of Q&A. Compare to prior-2 average. Rising density = sell-side getting more skeptical. Falling density = capitulation (which is itself a signal — sell-side relief often marks local conviction tops).

4.6: Management evasiveness (Q&A turns)

For each speaker-change turn in Q&A where an analyst asks a question and management responds:

• Extract the first noun-like token sequences from the question (3-7 most-content-bearing words)
• Check whether management's first 50 words contains any of those tokens
• If 0 matches → flag as "evasive"

Compute evasive_turns / total_qa_turns. Compare to prior-2 average.

Important caveat in the Research note: this is a heuristic — disclosure-restricted topics (M&A, legal, pricing strategy) often look "evasive" by this measure but reflect legitimate non-disclosure. The skill surfaces the measure as evidence to inspect, not as a conclusion.

4.7: Guidance language patterns

Scan prepared remarks + Q&A for guidance constructs:

• Quarterly guidance: Q[1-4]\s+revenue\s+(of|in)\s+, we expect\s+Q[1-4], guiding\s+to\s+
• Annual guidance: full year, fiscal year, FY[0-9]{2,4}
• Range vs. point: between\s+\$[\d.]+\s+and\s+\$[\d.]+ (range) vs. approximately\s+\$[\d.]+ (point)

Track: did current Q widen, narrow, or eliminate guidance ranges vs. prior Q? Did it shift from annual to quarterly framing? These are explicit confidence signals.

Step 5: Read thesis for context (parallel reads)

Single tool-call block:

Tool	Target	Purpose
Read	$THESIS_PATH (already located in Step 0.3)	Bull Case, Bear Case, Conviction Triggers, recent Log entries (last 5)
Read	_hot.md	Recent Conviction Changes, Open Questions for this ticker
Read	_graph.md	Adjacency primer (Step 6 Mode A + Step 9 Mode C fanout — read once, parse in memory)
Read	Mental Models/Generalist - Overview + the matching Industry -/Lens - note(s) for the thesis sector:	MANDATORY reading gate (per _shared/mental-models-section.md) before Step 6's judgement work — load-tiered (Generalist always; Industry/Lens by sector only), cached across the run

Apply the READING PROTOCOL from [[Generalist - Overview]] to the Step 6 cross-reference: a management-language shift that appears to confirm a thesis driver is a hypothesis to test, not a verdict; run the base-rate adversarially; agreement across models is a disconfirm trigger, not confirmation.

Extract from the thesis:

• bull_case_drivers: parse Bull Case section; identify 3-5 named drivers (each driver typically has a sentence-opening phrase, e.g., "Rack-scale deployment economics drive NVL72 attach rate inflection")
• bear_case_risks: parse Bear Case section; identify 3-5 named risks
• conviction_triggers: parse Conviction Triggers section; extract the → HIGH if, → LOW if, → CLOSE if falsifiable statements

These become the cross-reference targets in Step 6.

Step 6: Cross-reference signals against thesis

For each signal extracted in Step 4, attempt to match against:

6.1: Bull-case driver alignment

For each bull_case_drivers[i]:

• Tokenize the driver phrase (significant nouns + verbs)
• Search Step 4.1 (new language) for phrases overlapping ≥2 tokens with the driver
  • Match → "Bull case driver confirmed by new framing: [driver] ↔ [new phrase]"
• Search Step 4.2 (dropped language) for phrases overlapping ≥2 tokens
  • Match → "⚠️ Bull case driver may be weakening: [driver] ↔ [dropped phrase] — management stopped framing this"

6.2: Bear-case risk alignment

For each bear_case_risks[i]:

• Same tokenization
• Match against new language → "Bear case risk surfaced in management framing: [risk] ↔ [new phrase]"
• Match against dropped language → "Bear case risk de-emphasized in framing: [risk] ↔ [dropped phrase] — may indicate management confidence OR avoidance"
• Cross-reference 4.5 (Q&A skepticism) — if analysts are now asking about a previously-mute bear-case risk: Analyst questions surfaced bear-case risk: [risk]

6.3: Conviction trigger touching

This step is the origin pattern for _shared/trigger-touch.md (the shared contract that generalises it to /numbers, /sync, /deepen, /ingest). /transcript owns the named-observable path: evidence bears on a trigger's variable without a number.

For each falsifiable trigger:

• Parse the trigger's named variable (e.g., Hyperscaler capex guides flat for 2 consecutive Qs)
• Check whether current Q transcript provides evidence on that variable
• If yes → flag the trigger as "touched" with direction (firing toward HIGH / firing toward LOW / firing toward CLOSE)
• Touched triggers are MANDATORY content in the Research note's Thesis Delta section

Flag-only — a touched trigger surfaces in the Step 11 advisories (which already suggest /status), never an auto-conviction-change (Tier-3, per the contract's anti-patterns).

6.4: Guidance vs. consensus framing

If 4.7 detects a guidance shift (widened range, eliminated annual guidance, etc.):

• Cross-reference whether the thesis Bull Case or Bear Case explicitly hinges on guidance-trajectory expectations
• If yes → flag in Thesis Delta as direct evidence

Step 7: Write Research note

Two output shapes depending on mode.

Step 7A: Default + specific-quarter mode

File: Research/YYYY-MM-DD - [TICKER] [QN-YYYY] - earnings.md (where YYYY-MM-DD is today's date)

Frontmatter:

---
date: YYYY-MM-DD
tags: [research, earnings, TICKER, SECTOR_TAG]
status: active
sector: [from thesis frontmatter]
ticker: TICKER
source: [Quartr event URL (the cache's eventUrl) when transcript_source is quartr;
         https://financialmodelingprep.com/stable/earning-call-transcript?symbol=TICKER&year=YYYY&quarter=N when fmp]
source_type: earnings
transcript_source: quartr | fmp  (which source served the TARGET quarter)
transcript_quarter: QN-YYYY
transcript_date: YYYY-MM-DD  (from the cache `date` field — actual earnings call date)
transcript_word_count: [N]
prior_comparators: [QN-1-YYYY, QN-2-YYYY]
---

--event mode differences: filename YYYY-MM-DD - [TICKER] [Event Title] - video-transcript.md, source_type: video-transcript, tags: [research, video-transcript, TICKER, SECTOR_TAG], transcript_quarter replaced by transcript_event: "[Event Title]", no prior_comparators. Evidence tables show absolute densities only (no delta columns — there are no comparators); everything else (Thesis Delta cross-referencing, Summary, Contradiction Check, Source Excerpts) is unchanged — a CMD or investor update cross-references against Bull/Bear/Triggers exactly like an earnings call.

Body sections (in order — all required except where noted):

# [TICKER] [Q[N] FY[YYYY]] — Earnings Transcript

## Thesis Delta

[1-2 sentences PER cross-reference hit from Step 6. Lead with the strongest signal. Use the templates:]
- **Bull case [strengthened|weakened|unchanged]** — [specific driver from Step 6.1] ↔ [evidence from Step 4]
- **Bear case [strengthened|weakened|unchanged]** — [risk from Step 6.2] ↔ [evidence]
- **Conviction trigger touched** — [trigger text from Step 6.3] — current Q [confirms/disconfirms]: [evidence]
- **Guidance shift** (if applicable per Step 6.4) — [shift type] vs. [thesis assumption]

[If zero cross-references hit: this section reads "No direct thesis-relevant deltas. Notable framing shifts logged in Evidence below for future synthesis."]

## Summary

[2-4 paragraphs. Lead with management's CORE argument for the quarter — what mechanism are they pitching to explain the results, what forward construction are they framing? NOT a press-release re-summary; the thesis owns the business description. Capture the argument structure, the named drivers, the qualifier/hedge structure. Length proportional to transcript: 2 paragraphs for short (~8,000 words), 3-4 paragraphs for long (>20,000 words).]

## Evidence

### Language deltas

**New framing (top 5 introduced this Q)**:
| Phrase | Mentions this Q | Mentions prior 2 Qs |
|---|---|---|
| [phrase] | [N] | 0 |
| ... |

**Dropped framing (top 5 retired this Q)**:
| Phrase | Mentions prior 2 Qs combined | Mentions this Q |
|---|---|---|
| [phrase] | [N] | 0 |
| ... |

### Hedging & specificity

| Metric | Current Q | Prior 2-Q avg | Δ | Direction |
|---|---|---|---|---|
| Hedging density (per 1000 words) | [N] | [N] | [Δ%] | [more confident / more hedged / unchanged] |
| Numeric mention density (per 1000 words) | [N] | [N] | [Δ%] | [more specific / less specific / unchanged] |

### Q&A signals

| Metric | Current Q | Prior 2-Q avg | Δ |
|---|---|---|---|
| Skeptical-keyword density (per 1000 words of Q&A) | [N] | [N] | [Δ%] |
| Evasive turns (% of total Q&A turns) | [N]% | [N]% | [Δpp] |
| Total Q&A turns | [N] | [N] | — |

[Caveat: evasiveness is heuristic — disclosure-restricted topics (M&A, legal, pricing) often appear evasive but reflect legitimate non-disclosure.]

### Guidance language (if Step 4.7 surfaced changes)

[Specifics: range width change, point-vs-range shift, annual-vs-quarterly shift, dropped guidance, added guidance]

## Contradiction Check

[Specific to which thesis assumption the transcript contradicts or supports. Pull from Step 6 cross-references. Examples:]
- [[Research/path-to-prior-research-note]] argued [specific thesis] — current Q transcript supports / contradicts via [evidence].
- Bear case (per [[Theses/TICKER]] §Bear Case) hinges on [risk]; current Q transcript [evidence for/against].

[If no specific contradictions: "Transcript reinforces existing thesis framing — no contradictions surfaced."]

## Source Excerpts

[3-7 verbatim quotes that anchor the Thesis Delta points. Quote blocks formatted as Obsidian blockquotes (`>`). Each quote attributed by speaker if extractable. **Quartr-sourced quotes: link the attribution to the paragraph's timestamped deep-link `url`** (pull the matching paragraph from the cache via jq — do not read the whole transcript into context to find it) — this makes every quote click-to-verify against the call audio:]

> [[CEO Name, CEO](paragraph-deep-link-url)]: "...verbatim quote..."

> [[Analyst Name, Analyst](paragraph-deep-link-url)]: "...verbatim question..."

[FMP-sourced quotes: plain attribution as before — no per-paragraph URLs exist. Quotes should be SHORT — 1-3 sentences each. The reader can read the full transcript via the source: URL if they need more.]

Required ## Key Segments section (per /ingest content-quality check #5 when source_words >15,000):

## Key Segments

[3-5 sub-sections mirroring the transcript's structural progression. Each 2-5 sentences. Examples:]

### Prepared Remarks - Strategic Framing
[Management's positioning of the quarter — what story are they telling about the business right now]

### Prepared Remarks - Segment Discussion
[Per-segment color: which segments management emphasized, which they de-emphasized]

### Prepared Remarks - Forward Construction
[Guidance philosophy, capex commentary, capital allocation framing]

### Q&A - Most Skeptical Exchange
[The single most-pointed analyst question and management's response — verbatim or paraphrased + commentary]

### Q&A - Most-Disclosed New Information
[Where management revealed something that wasn't in prepared remarks — often the highest-value information from any earnings call]

Include ## Key Segments ONLY when transcript_word_count > 15,000. For shorter transcripts (rare — most US large-cap calls run 10-20k words), omit per /ingest Step 2 spec.

Step 7B: --diff mode

File: Research/YYYY-MM-DD - [TICKER] [Q1] vs [Q2] - transcript diff.md

Frontmatter:

---
date: YYYY-MM-DD
tags: [research, earnings, comparison, TICKER, SECTOR_TAG]
status: active
sector: [from thesis frontmatter]
ticker: TICKER
source: [per-quarter source URLs, comma separated — Quartr event URL for quartr-served quarters, FMP earning-call-transcript URL for fmp-served]
source_type: analyst-report
transcript_source: [quartr | fmp | mixed]
diff_quarters: [Q1-YYYY, Q2-YYYY]
---

Body:

# [TICKER] [Q1-YYYY] vs [Q2-YYYY] — Transcript Diff

## Thesis Delta
[Cross-reference findings against thesis — same structure as Step 7A but framed as "between these two quarters"]

## Summary
[2-3 paragraphs framing the analytical comparison — what story does management tell differently between these two quarters? Lead with the most-load-bearing change.]

## Evidence

### Side-by-side language shift

| Theme / Phrase | [Q1-YYYY] frequency | [Q2-YYYY] frequency | Direction |
|---|---|---|---|
| [top 15 phrases with biggest absolute change] |

### Side-by-side hedging & specificity

| Metric | [Q1-YYYY] | [Q2-YYYY] | Δ |
|---|---|---|---|
| Hedging density | | | |
| Numeric density | | | |
| Q&A skeptical density | | | |
| Evasiveness % | | | |

### Strategic framing shifts (qualitative)

- **Introduced between quarters**: [list]
- **Dropped between quarters**: [list]
- **Re-framed**: [phrases that changed wording but kept meaning — e.g., "Hopper transition risk" → "Blackwell ramp velocity"]

## Contradiction Check
[Same scope as 7A]

## Source Excerpts
[Paired quotes — one from each quarter — anchoring the largest shifts]

Step 7.5: Post-write verification gate

Re-read the just-written Research note. Apply the verification checks from /ingest Step Post-write verification:

• Structural checks 1-4 (frontmatter parseability, required fields, body non-empty with at least one ## section, last line not mid-sentence): block on failure → restore via deletion + abort.
• Content-quality check 5 (proportional body word count floor): for source_type: earnings with transcripts typically >15,000 words → body must be ≥2,500 words AND have a ## Key Segments section with ≥3 sub-sections. For diff mode source_type: analyst-report → body must be ≥800 words (1,500-5,000 source-word bucket).
• Content-quality check 7 (section structural minimum): all 4 required sections present with non-empty content — ## Thesis Delta, ## Summary, ## Evidence, ## Contradiction Check.
• Domain validator #8 (source_type: earnings signature — applies to Step 7A): MUST contain quarterly-period token, ≥2 numeric currency figures, ticker-shaped token. The transcript content itself supplies these; the Research note's restating Evidence section should also contain them.

On verification failure:

• Structural failure → delete the partial Research note, log failure, abort the entire skill run with diagnostic.
• Content-quality failure → delete the Research note (treat as contaminated, not partial), report which check failed with diagnostic (body words, source words, missing tokens). Retain the raw transcript cache so user can re-run after fixing the analysis prompt or content-extraction logic.

Step 8: Append thesis Log entry

Edit $THESIS_PATH's ## Log section. Append:

### YYYY-MM-DD
- Transcript ingested: [QN-YYYY] — [most-significant thesis-delta finding from Step 6 in plain prose, 1 sentence]. See [[Research/YYYY-MM-DD - TICKER QN-YYYY - earnings]].

For --diff mode:

### YYYY-MM-DD
- Transcript diff: [Q1] vs [Q2] — [most-significant shift in plain prose]. See [[Research/YYYY-MM-DD - TICKER Q1 vs Q2 - transcript diff]].

Prefix Transcript ingested: (or Transcript diff: for diff mode) is intentionally NON-skill-origin. This is a research-driven Log entry — the Research note created in Step 7 represents real new analytical content that should propagate to sectors and macro notes. /sync Step 2.5 will treat the change as research-driven and run Steps 3-5 normally.

Do NOT add this prefix to _shared/log-prefixes.md skill-origin list. Adding it would silently break /sync propagation from every earnings transcript ingest — exactly the wrong behavior.

Step 9: Graph-primer propagation fanout

Per .claude/skills/_shared/graph-primer.md Mode C (Propagation-fanout primer — same as /ingest Step 3.5).

Using _graph.md (already read in Step 5):

• T = TICKER
• S = thesis's sector: from frontmatter
• M = macro references found in the Research note's body (parse [[Macro & Technology/...]] wikilinks)

Compute:

• direct_targets = {THESIS_PATH} (the ticker's own thesis — always a direct target)
• sector_candidates = sector_reverse[S] - direct_targets
• macro_candidates = ∪{macro_reverse[m]} for m in M - direct_targets - sector_candidates

Surface to user in Step 12 report. Wikilinks for direct_targets are already in the Research note (the thesis link in ## Thesis Delta and ## Source Excerpts). For sector_candidates and macro_candidates — these are advisory; the user reviews and adds wikilinks before /sync if they want propagation to those neighbors.

Missing-graph fallback: per contract — log ℹ️ _graph.md absent/unparseable — graph primer skipped and proceed. Step 8's Log entry + Step 7's Research note still propagate via /sync based on body-content grep alone.

Step 10: Update _hot.md

Follow .claude/skills/_shared/hot-md-contract.md. Read first (already read in Step 5), then edit. Do NOT touch Latest Sync / Sync Archive.

1. Active Research Thread: per contract's same-ticker-continuation rule — likely set to [TICKER] earnings analysis (QN-YYYY).
2. Open Questions: if Step 6.3 surfaced touched conviction triggers OR Step 4.5 surfaced rising Q&A skepticism, append 1-2 open questions for /surface later: - [TICKER] QN: [trigger touched in direction X] — re-evaluate conviction at next data point.
3. Latest Sync and Sync Archive: do NOT touch (owned by /sync).
4. Recent Conviction Changes: do NOT touch (a transcript ingest is research input, not a conviction decision — the user calls /status if a conviction change is warranted).

Cap enforcement: per contract — apply drops if over soft cap; abort _hot.md update if over hard cap (DON'T block the primary /transcript operation).

Step 11: Material-finding advisories

If Step 6 surfaced any of:

• A conviction trigger fired toward LOW or CLOSE → suggest /status [TICKER] [field] [current]→[new] with the rationale phrased from the trigger
• A conviction trigger fired toward HIGH → suggest /status [TICKER] conviction [current]→high
• Bear-case risk strengthened by ≥2 evidence vectors → suggest /stress-test [TICKER]
• Bull-case driver weakened by ≥2 evidence vectors → suggest /deepen [TICKER] --section "Bull Case"
• Q&A skepticism +50% vs. prior 2-Q avg → suggest /stress-test [TICKER] (analyst skepticism leading official conviction is a real edge)

These print as → Consider: [command] lines in the Step 12 report. Never auto-run.

Step 12: Release lock and report

Final Bash block — lock release per preflight §1.7.

LOCK_FILE="<paste-from-Step-0.1>"
EXPECTED_TOKEN="<paste-token-captured-from-Step-0.1>"
if [ -f "$LOCK_FILE" ] && grep -q "token: $EXPECTED_TOKEN" "$LOCK_FILE"; then
  rm -f "$LOCK_FILE" && echo "=== LOCK RELEASED ($LOCK_FILE) ==="
else
  echo "⚠️ Lock ownership check failed at release — skipping rm."
fi
# Keep cache (.data/transcripts/*) — useful for diff mode re-runs
# Only clear /tmp scratch
rm -f /tmp/transcript_*.json /tmp/transcript_dates_*.json

Report

✓ /transcript [TICKER] [QN-YYYY | --diff Q1 Q2] complete

Research note:    [[Research/YYYY-MM-DD - TICKER QN-YYYY - earnings]]
Transcript date:  YYYY-MM-DD (Q[N] FY[YYYY])
Transcript source: quartr (exact split, exact speaker roles) | fmp (heuristic split) — per quarter if mixed
Transcript size:  ~[N] words
Prior comparators: [Q-1, Q-2]

Thesis delta findings:
  Bull case:       [strengthened | weakened | unchanged]  — [1-line summary]
  Bear case:       [strengthened | weakened | unchanged]  — [1-line summary]
  Triggers touched: [N]  ([list with direction])

Signal extracts (most analytically interesting):
  New framing introduced: [top 3 phrases]
  Dropped framing:        [top 3 phrases]
  Hedging shift:          [direction, %]
  Q&A skepticism shift:   [direction, %]
  Touched trigger(s):     [list, or "none"]

Graph primer:
  Direct targets:      [list of theses]
  Sector candidates:   [list of theses sharing sector — review for wikilinking]
  Macro candidates:    [list of theses sharing macro refs — review for wikilinking]

Thesis Log: appended ("Transcript ingested:") — research-driven prefix
_hot.md: updated (Active Research Thread, Open Questions if applicable)
Transcript cache: .data/transcripts/TICKER_QN-YYYY[.quartr].json (retained for diff mode re-runs)

Suggested next steps:
  1. /sync [TICKER]          — propagate to sector + macro notes
  2. /graph last             — reconcile adjacency
  → Consider: [Step 11 advisories, if any]

Design constraints (xxx DO NOT VIOLATE xxx)

1. Earnings transcripts require an existing thesis. Step 0.3 hard-aborts when no thesis exists. The whole value of /transcript is delta-against-thesis-context; orphan-earnings ingestion is /ingest's job.

2. The Log prefix Transcript ingested: is NOT skill-origin. It must propagate via /sync because the Research note carries genuine new analytical content (management framing shifts, trigger touches) that affects sector competitive-dynamics framing and macro thread continuity. Adding this prefix to _shared/log-prefixes.md skill-origin list silently disables every transcript's propagation.

3. Heuristic signals are flagged as heuristics in the Research note. Hedging density, evasiveness percentage, skeptical-keyword density are all imperfect text-frequency measures. The Research note must explicitly caveat each (especially evasiveness, which is highly false-positive on disclosure-restricted topics).

4. Transcript cache is gitignored. .data/transcripts/ lives under .data/ which is in .gitignore per Live Portfolio's documentation. Re-runs read from cache; first runs populate it. No transcript content ever lands in version control.

5. Diff mode is a separate output shape. Treating diff as "two ingests then a manual side-by-side" loses the value — the cross-quarter Evidence table is the primary deliverable. Don't conflate the modes.

6. ## Key Segments section is required for transcripts >15,000 words. Per /ingest content-quality check #5. Most US large-cap earnings calls cross this threshold. The 5-segment structure (Strategic Framing, Segment Discussion, Forward Construction, Most-Skeptical Exchange, New Information) reliably captures a 60-90 minute call's analytical substrate.

7. Quartr is primary; FMP is the fallback, and wholesale tier is the fallback's assumed entitlement. The FMP earnings-transcript endpoint is gated on FMP wholesale tier (or higher); lower tiers return 403/404 — the fallback aborts cleanly with an FMP error rather than silently degrading. A run that never touches FMP (Quartr served every quarter) has no tier dependency at all.

8. Never re-fetch when cache is fresh. First-pass writes to .data/transcripts/TICKER_QN-YYYY.quartr.json (Quartr) or TICKER_QN-YYYY.json (FMP). Subsequent runs (e.g., user iterating on prompt phrasing for the Research note) read cache — either format satisfies the quarter. Every avoided fetch is one less API/MCP touchpoint and faster per run.

9. Transcript content never flows through context by hand. Oversized MCP results persist to a file — cp/jq that file into the cache (_shared/quartr.md §4). The signal script reads the cache from disk; Claude reads targeted jq paragraph slices for Step 7's quotes, never the whole transcript. Retyping or summarising a transcript from memory into the cache is fabrication — abort instead.

10. Never take a wrong company. Quartr resolution failures (no confident search match, eventId belonging to another company) degrade to FMP or abort — the SIVE→Silver-Verde class of ticker collision applies to any source. Resolution discipline lives in _shared/quartr.md §3.