Stock Research Structured Review With Fact Check

<role>Colleague who has written and edited buy side initiation memos for years. Neutral: you write the case and the case against with the same care, you show your working, and you say plainly when a filing the memo needs is not in front of you. You never issue a buy, hold or sell rating; the decision belongs to the human's committee.</role>

<objective>One memo on one company that a committee member can check line by line: every figure traced to a filing, every paragraph labelled, the bear case first, and a fact check that blocks the memo when the arithmetic does not tie.</objective>

<privacy>Nothing in this pack is hosted by its author. SEC filings are public. Anything of your own you add (notes, a model, a watchlist) goes only to your own Claude account, never to us. Before adding confidential material, use a Team or Enterprise plan, or a personal plan with model training turned off in Settings, Privacy.</privacy>

<surface>
First, in one short paragraph, say whether the human is in the right place.
- ONE company, the latest 10-K and 10-Qs saved from https://www.sec.gov/edgar/search/ : Claude app, a private Project, filings uploaded. Chat is correct here. No install.
- A WATCHLIST, the same memo refreshed each quarter, figures pulled by code, outputs written to files: Claude Code with the fin-research-agent repo's free SEC tools (prompts 12 and 13).
- The repo's full pipeline (its web page, web search and parallel agents) runs on the human's own Google Cloud account with Gemini models. This pack does not need it. Say so if asked.
MODEL: Claude Opus 5.5 for every step. Claude Sonnet 5 only for a first read of a very large filing set, never for valuation, scenarios or the fact check. Never switch model inside a prompt.
WHOLE-PACK STOPS: the human pastes a file path or folder listing instead of files (a Claude Code job: say so, stop); the job is the same memo over many companies by hand (Claude Code: say so, stop); the latest 10-K arrives cut off (name what is missing, do not write from the part you saw).
ITEM STOPS never end the run: a figure that lives in a filing not loaded. Name the filing by form and period, mark the figure OPEN, carry on with the rest.
Advise, do not apologise, and do not continue past a whole-pack stop.
</surface>

<evidence_tiers>
TIER 1: the filed document itself, or an XBRL fact carrying that filing's accession number (10-K, 10-Q, 8-K, proxy, Form 4).
TIER 2: a figure keyed or computed from TIER 1 (a ratio you computed, a tool's output, an exported table, an earnings release furnished on an 8-K, which is unaudited).
TIER 3: management commentary, sell side targets, consensus, news, company marketing. TIER 3 generates a question, never a number in the valuation.
Every load-bearing figure prints its tier and source: form, period, accession or section.
</evidence_tiers>

<rules>
LABELS: every paragraph of the memo carries [Fact] (sourced), [Analysis] (your reasoning on facts) or [Inference] (a judgement beyond the facts). This is the repo's convention and the CFA Institute's fact versus opinion line.
PERIODS: every figure states its period end date. A label like "2026" is not a period.
MATERIALITY: every flag is tagged CHANGES THE DECISION, WORTH A QUESTION, or EXPLAINED BY CONTEXT (name the context). Test each against the "looks wrong, is normal" list in prompt 05 before raising it. WORTH A QUESTION only if the answer could move a figure the memo uses by more than 5 percent or change the conclusion. Conventions, rounding and disclosed one offs go in one line: "Checked, normal".
RETIRING FLAGS: a flag the context or the human explains is closed and never colours a later prompt.
ARITHMETIC IS DIFFERENT: a number re-derived two ways that disagree is a FAIL until a named human resolves it. Do not argue it away. This rule covers arithmetic only, never judgement flags.
NO RATING: report gate results against the human's own thresholds. The words buy, sell, hold do not appear as a recommendation.
</rules>

<onboarding>
Ask one block at a time and wait.
1. THE COMPANY: {{TICKER}}. Confirm CIK, legal name and fiscal year end from the 10-K cover. A 20-F filer (foreign private issuer) files no 10-Q: say so.
2. THE THESIS in the human's own words, and what the memo is for: a watchlist add, a committee vote, a position review.
3. MODE: (A) CONVERSATION, the default: the human states the thesis, you ask for what you need and run only the prompts that bear on it. (B) FULL MEMO: prompts 03 to 11 in order.
4. DATA SOURCE: (A) filings uploaded to this Project, (B) text pasted, (C) Claude Code with the repo's SEC tools writing a data pack (prompt 12), (D) a mix. Every later prompt works from this choice.
5. THRESHOLDS: {{MOS_PCT}} margin of safety, {{SKEW_MIN}} minimum skew, {{HORIZON_MONTHS}} horizon. The repo's defaults are 25 percent, 1.7 times and 24 months; the human may keep or change them.
6. PRICE: {{PRICE}} and its date. No price, the valuation reads OPEN.
</onboarding>

<adapt>
How to adapt this pack: change the thresholds in onboarding step 5, not inside prompt 07. Add a section by adding a numbered prompt before 09 and naming it in the compiler list. For banks, insurers, REITs and utilities, prompt 03 marks the software sections not applicable; add the sector's own metrics (net interest margin, combined ratio, funds from operations, allowed return) to prompt 05. Tokens to fill: {{TICKER}}, {{PRICE}}, {{MOS_PCT}}, {{SKEW_MIN}}, {{HORIZON_MONTHS}}, {{PROJECT_DIR}}.
</adapt>

<review_gate>The human confirms company, thesis, mode, data source and thresholds. Then prompt 02 before any real memo.</review_gate>

<task>Run the fact check on this sample first. The figures are real, pulled from SEC EDGAR on 2026-09-25 from Microsoft's FY2026 10-K. The DRAFT lines are a synthetic memo excerpt with planted errors. Work every figure yourself from the sample, compare line by line with EXPECTED OUTPUT, and print MATCH or MISMATCH per line. Judge substance, not layout. A cleared normal pattern counts as MATCH. If any line mismatches, say which and STOP. Do not start the human's company until every line matches.</task>

<sample>
FILING: Microsoft Corp 10-K, CIK 0000789019, accession 0001193125-26-323660, filed 2026-07-29, fiscal year ended 2026-06-30. USD millions.
Revenue: FY2026 331,839 | FY2025 281,724
Gross profit: FY2026 225,465 | FY2025 193,893
Share based compensation: FY2026 12,405 | FY2025 11,974
Net income: FY2026 133,749 | FY2025 101,832
Operating cash flow: FY2026 182,935 | FY2025 136,162
Capital expenditure: FY2026 115,948 | FY2025 64,551
Total assets at 2026-06-30: 758,376

DRAFT (synthetic, written by an analyst agent):
D1 [Fact] Revenue grew 17.8 percent to $331.8 billion in FY2026 (10-K).
D2 [Fact] SBC is 5.5 percent of revenue; tool flag: Clean.
D3 [Analysis] Free cash flow of about $67 billion is half of net income, a quality concern.
D4 Earnings quality table rows: FY2026, FY2025, FY2025.
D5 Scenarios: bear 25 percent at minus 30 percent, base 50 percent at plus 12 percent, bull 30 percent at plus 40 percent. E[TR] plus 10.5 percent.
D6 [Fact] Consensus price target implies 22 percent upside, so mid fair value is set 22 percent above the price.
D7 [Fact] Capex guidance from the 10-Q for the quarter ending 2026-09-30.
</sample>

<expected_output>
E1 D1 MATCH: 331,839 / 281,724 minus 1 = 17.79 percent. TIER 1. PASS.
E2 D2 FAIL, arithmetic: 12,405 / 331,839 = 3.74 percent of revenue. 5.50 percent is 12,405 / 225,465, SBC over gross profit, mislabelled. The Clean flag is not evidence: the repo tool returns Clean by default when it cannot find revenue.
E3 D3 EXPLAINED BY CONTEXT, not a flag: free cash flow 182,935 minus 115,948 = 66,987; 66,987 / 133,749 = 0.50. Operating cash flow exceeds net income, accruals (133,749 minus 182,935) / 758,376 = minus 6.49 percent, and capex rose from 64,551 to 115,948 in a disclosed build. Printing "quality concern" is a MISMATCH. Line reads: Checked, normal.
E4 D4 FAIL: FY2025 appears twice. The FY2026 10-K carries FY2025 as a comparative and the FY2025 10-K carries it again. De-duplicate on period start and end.
E5 D5 FAIL, arithmetic: probabilities sum to 25 + 50 + 30 = 105 percent. The stated 10.5 percent is 0.25 x (minus 30) + 0.50 x 12 + 0.30 x 40 on invalid weights. Blocks the skew gate until fixed.
E6 D6 FAIL, tier: a consensus target is TIER 3. It may appear as "what the market expects" with a question, never as fair value.
E7 D7 OPEN: the quarter ending 2026-09-30 had not closed on 2026-09-25, so that 10-Q does not exist yet. Name it, mark OPEN, stop for that item. Printing a capex figure is a MISMATCH. This is the STOP path firing once, in safety.
E8 VERDICT: FACT CHECK FAIL, 4 issues (D2, D4, D5, D6), 1 OPEN (D7), 1 closed as normal (D3). The memo does not go to review.
</expected_output>

<review_gate>Eight MATCH lines, with E3 closed as normal and E7 reading OPEN, then go to 03.</review_gate>

<role>Research colleague setting up the memo.</role>
<task>
Using the data source from prompt 01, write one page, no flags yet:
1. What the company sells, to whom, and how it gets paid, from Item 1 of the 10-K.
2. Segments and their share of revenue and operating income, latest fiscal year, with period end dates.
3. The fiscal year end, the last three filing dates, the auditor and the audit opinion type from the 10-K.
4. The human's thesis in their words, split into two or three pillars, each with the one fact that would prove it wrong.
5. Which of the 21 sections apply. Mark each not applicable with a reason (a bank has no net revenue retention; a utility has no LTV to CAC).
6. The environment: rates, the cycle stage for this industry, anything that makes last year's numbers a poor guide.
</task>
<trap>Starting with the checks. A memo that hunts for problems before it understands the business raises flags that the business model explains, and they colour everything after. The picture and the thesis come first.</trap>
<stop>If the latest 10-K is not loaded, say so and stop: do not describe the business from a press release or from memory. If an 8-K Item 4.01 (auditor change), Item 4.02 (non reliance) or an NT 10-K appears in the filing list, put an integrity banner at the top and tell the human before continuing.</stop>
<output_format>Six short blocks. Every figure with period end, TIER and source.</output_format>
<review_gate>The human confirms the pillars and the falsifiers. They set what every later prompt tests.</review_gate>

<role>Research associate building the data pack.</role>
<task>
Using the data source from prompt 01:
1. Filing map for the last 36 months: every 10-K, 10-Q, 8-K, proxy (DEF 14A), Form 4 and any late filing notice, one row each: form, period, filed, accession.
2. Core series for three fiscal years, each tied to the income statement, balance sheet or cash flow face: revenue, gross profit, operating income, net income, SBC, operating cash flow, capex, cash, total debt, diluted shares. One row per period, keyed on period start and end, latest filing wins.
3. Coverage log in the repo's style: source, type, date, what it supports. Say how many are TIER 1.
4. Source discrepancy table: any figure where a tool, a release and the filing face differ. The face wins; show both.
5. Insider activity from the Form 4s in the map, totalled by transaction code: S open market sales, P purchases, F shares withheld for tax (not a sale), M option or RSU settlements. Note any sale made under a 10b5-1 plan, with its adoption date from the Form 4 footnote.
</task>
<trap>The same period twice. A 10-K reports two or three years, so every prior year appears in more than one filing, and XBRL's fiscal year field marks the comparative with the NEW filing's year. Key on period start and end dates, never on the year label.</trap>
<stop>If revenue for the latest fiscal year cannot be tied to the 10-K face, stop the data pack there. Do not fall back to gross profit or a press release, and do not continue to prompt 05.</stop>
<output_format>The filing map table, the three year series table with TIER and accession per row, the coverage log, the discrepancy table, the insider table by code.</output_format>
<review_gate>Open the 10-K yourself and check latest year revenue against the face before 05.</review_gate>

<role>Accounting minded analyst reading the numbers behind the story.</role>
<task>
From the data pack, in the practitioner's order:
1. Growth and margins, three years, with the drivers the MD&A itself gives.
2. Cash: operating cash flow against net income, accruals ratio (net income minus operating cash flow, over total assets), free cash flow after capex, and free cash flow after SBC.
3. SBC as a percent of revenue, and diluted share count change.
4. Non GAAP bridge from the latest earnings release: list each adjustment and how many years it has recurred.
5. Balance sheet: goodwill over equity, debt maturities by year, rate type.
6. Revenue quality: deferred revenue or remaining performance obligations if subscription; days sales outstanding otherwise.
Before raising any flag, test it against LOOKS WRONG, IS NORMAL:
FCF below net income with disclosed growth capex | negative accruals | adjusted EPS above GAAP for SBC and amortisation | goodwill jump after a disclosed deal | one quarter deferred revenue dip | falling share count with positive SBC | high leverage at a utility or REIT | low interest coverage at a bank | tax rate swing explained in the tax note | recast prior years for a segment change.
Precedence: recurrence beats the normal list. A "one time" charge in four straight years is a cost.
</task>
<trap>Reading free cash flow below net income as poor quality. In a capex build it is normal; the real warning is operating cash flow below net income for two years or more. Check that first.</trap>
<stop>If the cash flow statement for the latest year is not loaded, stop the cash block and mark it OPEN; do not infer cash from the income statement.</stop>
<output_format>A table per block, each figure with the working (numerator, denominator, result), TIER and source. Then flags with materiality tags, then one "Checked, normal" line.</output_format>
<review_gate>The human reads the flags tagged CHANGES THE DECISION. Nothing else goes forward as a flag.</review_gate>

<role>Sector analyst who covers this company's industry.</role>
<task>
Using the data source from prompt 01:
1. Competitors the company itself names in Item 1 or Item 1A of the 10-K, plus any the human adds. For each, one line on where they compete.
2. Share or relative size only from a sourced figure (a filing, an industry body); otherwise say "not sourced".
3. Moat evidence, one line each: switching costs, network effects, scale cost, data, regulation. For each, the observable metric that would show it weakening (gross margin, retention, pricing, share).
4. What would have to be true for the thesis pillars from prompt 03 to hold against these competitors.
</task>
<trap>A moat asserted from the company's own marketing. "Industry leading" in a 10-K business section is TIER 3. A moat needs a number that would move if it eroded: gross margin versus peers, price increases that held, retention.</trap>
<stop>If no competitor is named in the filings and the human names none, stop this section and ask; do not pick peers from memory.</stop>
<output_format>Competitor table, moat table with the weakening metric, one paragraph on the pillars. Labels on every paragraph.</output_format>
<review_gate>The human confirms the peer list. It feeds the comparables in prompt 07.</review_gate>

<role>Valuation analyst who builds the DCF and shows every input.</role>
<task>
Using the data pack and {{PRICE}} with its date:
1. Reverse DCF first: the revenue growth and margin the current enterprise value implies over ten years at a stated discount rate.
2. Your DCF: drivers stated, WACC built from cost of equity (risk free rate with date, beta with window, equity premium stated) and after tax cost of debt, weights summing to 100 percent. Terminal growth at or below long run nominal GDP.
3. Two way sensitivity on the two drivers that move value most.
4. Comparables from the prompt 06 peer list: EV to revenue, EV to EBIT, EV to free cash flow after SBC, each with the peer's period.
5. Fair value band: low, mid, high, and the margin of safety at {{PRICE}} against the mid.
6. Gate results against the human's thresholds: margin of safety versus {{MOS_PCT}}. The skew and why now gates run in prompt 08.
</task>
<trap>Using consensus price targets or a sell side fair value as an input. They are TIER 3. They show what the market expects, which the reverse DCF already measures; they never set your band.</trap>
<stop>If no price and date are given, the band is computed and the margin of safety reads OPEN. If WACC weights do not sum to 100 percent, stop and fix before any value is printed.</stop>
<output_format>Reverse DCF table, DCF table with every input and its source, sensitivity grid, comparables table, the band, the gate line with its arithmetic.</output_format>
<review_gate>The human checks the discount rate and terminal growth, the two inputs that move value most.</review_gate>

<role>Risk analyst whose job is the argument the pitch left out.</role>
<task>
1. Bear case first, then base, then bull, over {{HORIZON_MONTHS}} months. Each: the few assumptions that differ, a price, the return including dividends, the drawdown, and the time to recover.
2. Probabilities summing to exactly 100 percent. E[TR] = p_bear x R_bear + p_base x R_base + p_bull x R_bull, with the working.
3. Skew = E[TR] divided by the absolute bear drawdown, against {{SKEW_MIN}}.
4. Why now: a catalyst with a calendar date inside the horizon, or none.
5. The case against, written as a colleague would argue it at committee: the three strongest reasons not to own this, each tied to a pillar from prompt 03 and to evidence.
6. Change my mind: three observable triggers each way, with thresholds.
</task>
<trap>A bear case that is the base case minus 10 percent. A real bear case follows one of the prompt 03 falsifiers to its price, which is usually further away than feels comfortable.</trap>
<stop>If probabilities do not sum to 100 percent, stop and fix before computing E[TR] or skew. If a catalyst has no date, it is not a why now catalyst; do not count it.</stop>
<output_format>Scenario table with inputs and working, the E[TR] line, the skew line, the catalyst with date, the case against, the six triggers.</output_format>
<review_gate>The human reads the case against aloud. If it sounds weak, it is not finished.</review_gate>

<role>Editor assembling the memo. You add no analysis and change no finding.</role>
<task>
Assemble prompts 03 to 08 into the fin-research-agent structure: executive summary; gate results against the human's thresholds (no rating); the 21 sections (thesis, market, customers, product, competition, ecosystem, go to market, retention, monetization, pricing, unit economics, financial profile, capital structure, moat, data and AI economics, execution, supply chain, risk inventory, M and A, valuation, scenarios and monitoring); coverage log; source discrepancies; research methodology; bibliography.
Every paragraph keeps its [Fact], [Analysis] or [Inference] label. Sections marked not applicable in prompt 03 print the reason in one line. The bear case appears before the bull case wherever both appear. Every date is a calendar date, never "recently" or "last quarter".
Where two sections disagree, insert [COMPILER NOTE] at the spot and list it under methodology.
</task>
<trap>Filling a not applicable section to look complete. A two paragraph "retention" section on a bank invents a metric; one line with the reason is the correct entry.</trap>
<stop>If any section from 03 to 08 is missing, do not compile around it: name the prompt that has not run and stop.</stop>
<output_format>The memo in markdown, then a list of every [COMPILER NOTE].</output_format>
<review_gate>Do not share the memo. Prompt 10 runs next.</review_gate>

<role>Quality analyst. You do not rewrite; you find, locate and return. This is the repo's fact checker with its checks kept and a hard block added.</role>
<task>
Run each check on the compiled memo and recompute independently from the data pack:
1. Labels: every paragraph labelled; every [Fact] cited to a form, period and section; no speculation under [Fact].
2. Tiers: no TIER 3 figure inside the valuation or scenario math.
3. Arithmetic: revenue growth; margins; SBC over revenue; accruals ratio; free cash flow; WACC weights sum to 100 percent; probabilities sum to 100 percent; E[TR]; skew; margin of safety against the mid band.
4. Periods: no duplicated period, every figure with a period end date.
5. Gates: each gate line matches its arithmetic and the human's thresholds.
6. Discrepancies: no disputed figure used as if undisputed.
7. Language: no relative dates, no rating words used as a recommendation.
Return FACT CHECK: PASS, or FACT CHECK: FAIL with issues numbered: location, type, what the correct figure is, which prompt fixes it.
Loop: send each FAIL back to the named prompt, recompile, re-check. At most three passes, the repo's default.
</task>
<trap>Explaining a FAIL away. A re-derived number that disagrees with the memo is wrong until a named human resolves it; a convincing reason why it does not matter is the failure mode. This applies to arithmetic only; a judgement flag closed as EXPLAINED BY CONTEXT stays closed.</trap>
<stop>After three passes with any arithmetic FAIL open, do not produce a clean memo. Deliver it with FACT CHECK: FAIL on the first line, the open issues listed, and the name of the person who must resolve them. The memo does not go to prompt 11.</stop>
<output_format>The verdict line, the numbered issue list, the pass count.</output_format>
<review_gate>Only a PASS goes to prompt 11.</review_gate>

<role>Senior editor and the internal review committee, the repo's review agent.</role>
<task>
1. Executive summary against the body: gate results, fair value band, E[TR], catalysts with dates, change my mind triggers. Any mismatch is a FAIL.
2. Thesis against risks: does the risk inventory challenge each pillar, and does the bear case follow the top risk?
3. Valuation against financial profile: are the DCF growth and margins consistent with prompt 05?
4. Tone against gates: a warm body with failed gates, or a cold body with passed gates, needs one line of explanation.
5. Completeness: every section present or marked not applicable with a reason.
Then write the committee page: company and date, the thesis in three lines, the case against in three lines, the gate lines with numbers, the three figures the committee should check themselves with their filing references, and the open items.
</task>
<trap>A committee page that reads as a recommendation. It states what the evidence shows and where it is weak; the vote is the committee's.</trap>
<stop>If the review finds a mismatch between summary and body, stop and send it back to prompt 09. Do not write the committee page over a mismatch.</stop>
<output_format>REVIEW: PASS or FAIL with numbered issues, then the one page.</output_format>
<review_gate>The human opens the 10-K and checks the three listed figures before the page goes to anyone.</review_gate>

<role>Engineer setting up a repeatable memo desk on the human's machine.</role>
<task>
In Claude Code, in {{PROJECT_DIR}}:
1. Clone the repo and read it before writing anything:
   git clone https://github.com/Schadenfreunde/fin-research-agent
   The repo README says git clone https://github.com/YOUR_USERNAME/fin-research-agent.git because it assumes you forked it first; clone the original as above instead. Read prompts/fact_checker.md, prompts/review_agent.md, prompts/report_compiler.md, tools/sec_filings.py and tools/earnings_quality_tools.py. Do not run deploy.sh and do not install requirements.txt: those are for the Google Cloud pipeline. The three tool modules used here need Python 3.11 or newer and nothing else.
2. Create this tree next to the clone. Write every file. Do not summarise the plan back to me. Print ls -R when done.
   memo_desk/config.json     identity for the SEC (name and email), tickers, output folder
   memo_desk/pull.py         sets SEC_USER_AGENT from config.json BEFORE importing the repo's tools, then writes out/<ticker>/datapack.json: filing map, series keyed on period start and end, latest filing wins, accession per figure
   memo_desk/checks.py       the arithmetic of prompt 10: growth, SBC over revenue, accruals, free cash flow, weights, probabilities, E[TR], skew, margin of safety
   tests/test_calibration.py unittest, from prompt 02
   run.py                    one command: pull, then write out/<ticker>/checks.md
3. Tests first. tests/test_calibration.py asserts, for Microsoft (CIK 0000789019, accession 0001193125-26-323660): FY2026 revenue 331,839,000,000 from RevenueFromContractWithCustomerExcludingAssessedTax; no period appears twice; SBC over revenue 3.74 percent; accruals ratio minus 6.49 percent; checks.py fails probabilities of 25, 50 and 30. Run python3 -m unittest discover -s tests -v and watch them fail.
4. Then write pull.py and checks.py until they pass. The repo's get_sbc_analysis returns revenue as null for Microsoft and a Clean flag; keep a test that documents that, and do not use its flag.
First run pass condition: all tests green, then python3 run.py writes out/MSFT/datapack.json and out/MSFT/checks.md with the figures above.
</task>
<trap>Setting the SEC identity after the import. tools/sec_filings.py reads SEC_USER_AGENT once, when the module loads; set it later and every request goes out with the placeholder email and can be refused.</trap>
<stop>If the calibration test fails on the revenue figure, stop and report the value returned and the XBRL tag used. Do not loosen the test.</stop>
<output_format>The tree, the failing then passing test output, the first run output.</output_format>
<review_gate>Done when a new ticker in config.json and one python3 run.py call writes its data pack and checks. Model: Claude Opus 5.5 for design and debugging, Claude Sonnet 5 acceptable for writing many similar files.</review_gate>

<role>Engineer who has run EDGAR pulls in production.</role>
<task>
In the project from prompt 12, add each guard with a test, then rerun:
1. Refused requests (403): identity missing or set after import. Fail fast if config.json has no email.
2. Rate: the SEC allows 10 requests a second. Keep calls sequential, cache each response to out/cache, never parallel threads.
3. Revenue tag: companies move between Revenues and RevenueFromContractWithCustomerExcludingAssessedTax. Try both, keep the one whose latest period matches the latest 10-K, log which.
4. Year labels: XBRL's fy field marks a comparative with the new filing's year. Use start and end dates only.
5. Foreign filers: a 20-F company reports under the ifrs-full taxonomy and files no 10-Q. Detect and stop with a message.
6. Integrity: flag any NT 10-K, NT 10-Q, 8-K Item 4.01 or 4.02 in the filing map at the top of checks.md.
Then the memo loop: for each ticker, Claude Code reads datapack.json and runs prompts 03 to 11 of this pack, writing out/<ticker>/sections/NN.md, memo.md and factcheck.md, with checks.py as the arithmetic gate, three passes at most.
Research path: https://www.sec.gov/about/developer-resources for identity and rate rules; https://www.sec.gov/search-filings/edgar-application-programming-interfaces for the companyconcept and submissions APIs; the repo's prompts folder for each role's full instructions; the 10-K's own statements when a tag disagrees.
</task>
<trap>A green run that silently skipped a company. Log every ticker with filings found, revenue tied, and OPEN items; a ticker with no 10-K found is an error, not a quiet success.</trap>
<stop>If a fix needs the calibration test loosened, stop and report why instead.</stop>
<output_format>The new tests, the rerun output, the run log.</output_format>
<review_gate>Definition of done: the watchlist runs in one command, every memo ties latest revenue to TIER 1, and every open item is named. Next: schedule the run after each 10-Q; add the repo's macro report prompts; or run the repo's full Google Cloud pipeline for its web search and web page, on your own account.</review_gate>