Primary-Research / Channel-Check Harvester (finance-desk role #4 — the weakest leg)

The desk's job here is ground-truth: what do people who actually use the product, attend the concalls, and dig into the filings say — does it corroborate or disconfirm the dossier? This is the only finance-desk role with no canonical free data source (no Bloomberg/Refinitiv equivalent for retail), so it is the weakest leg by design. Treat its output as questions to verify, never as fact, and never let it move a grade.

This leg earns its keep ONLY by disconfirming ([[primary-research-earns-edge-only-by-disconfirming]]). Because markets aggregate public information efficiently, a harvest that confirms a popular long adds nothing — that view is already priced. The only durable contribution is (a) a fact that disproves a pillar of the thesis (a moat-bust, a channel breakdown), or (b) a checkable hard fact / forensic lead. Frame every run as an attempt to break the thesis, not to cheer it.

What this is / isn't (read before running)

• Output = LIVE DD CONTEXT, written to extracted/research/source_captures/<TICKER>_primary_research.md

— exactly like the existing dossier/source captures. It feeds Stage 1 (dossier / understanding) of stock-onboarding-pipeline and the forensic-accounting-redflags screen as corroboration / disconfirmation.

• It is NOT vault atomics. Do NOT write to vault/canon or vault/perspectives, do NOT

touch the brain index. Munger/Buffett (the binding CIO via 32_consult_brain.py) remain the sole arbiter of verdict, conviction and MoS. This skill supplies context, never a verdict.

• Weakest-leg honesty (state it in every capture). Indian retail community quality varies wildly;

ValuePickr & Reddit are self-selected and often talk-their-book; concalls are management's own framing. Weight every signal with human judgment and down-weight it hard against the vault.

• Epistemic ceiling (the EMH limit — state it too). Public chatter about a well-followed name is mostly

already in the price. So this harvest's only durable output is a disconfirmation or a checkable hard fact — never a confirmation of a crowded long. A glowing forum consensus is not a buy signal; at most it is a sentiment-froth datum (see STEP 5). This is the binding resolution of the efficiency tension: Damodaran's efficiency view is cited here to limit the skill's claims, never to license acting on crowd information.

Division of labour: the SCRIPT does the cheap scrapes, YOU do the MCP + judgment

• Script (data/scripts/61_harvest_primary_research.py) does the parts that are scriptable and

cheap: paginated ValuePickr capture (Discourse .json API), screener.in transcript-PDF discovery + pdftotext extraction, and a generic-web fallback that reuses `capture_web` from `60_ingest_external.py` (defuddle → Apify). Apify spend ≈ $0 (defuddle/JSON first).

• You (the agent) do what a standalone script can't: the Reddit step (Reddit MCP) and the

final synthesis. A script has no MCP access, so the Reddit pull is agent-only — this is explicit in the generated file.

STEP 0 — Frame the harvest as disconfirmation (+ pull the standing arbiter frame)

Before scraping, write down the dossier's core bull thesis in one line, then run the whole harvest as an explicit attempt to disprove it. Anchoring to the existing view is the failure mode here (avoid-anchoring-and-destroy-previous-ideas); a harvest that only confirms is flagged as anchoring and re-run with bear-case queries ("<company> problems / fraud / churn / competition / promoter").

Then pull the standing arbiter frame from the desk/binding layer (the --role filter is INCLUSIVE — it admits the desk atomic AND keeps the binding mentors on relevance):

set -a && source /Users/pw/invest/.env && set +a && /Users/pw/invest/.venv/bin/python \
  /Users/pw/invest/data/scripts/32_consult_brain.py \
  --company "<name>" --model general --step understanding --corpus blended --role primary-research \
  --json-out extracted/grilling/<TICKER>_primary.json

This surfaces [[primary-research-earns-edge-only-by-disconfirming]] and [[loudest-promoter-is-least-credible-source]] ALONGSIDE the binding arbiter atoms (beware-fraudulent-information-sources, do-not-talk-up-your-investments). The desk atomics are connective tissue, not a vote — they never set verdict/conviction/MoS.

STEP 1 — Run the harvester (ValuePickr + screener, scriptable)

First find the ValuePickr thread URL (web-search forum.valuepickr.com <company> — the thread id is the number in /t/<slug>/<id>). The screener URL is derivable from the ticker. Then:

set -a && source /Users/pw/invest/.env && set +a && /Users/pw/invest/.venv/bin/python \
  /Users/pw/invest/data/scripts/61_harvest_primary_research.py \
  --ticker <TICKER> \
  --valuepickr-url "https://forum.valuepickr.com/t/<slug>/<id>" \
  --screener-url   "https://www.screener.in/company/<TICKER>/consolidated/"

This writes extracted/research/source_captures/<TICKER>_primary_research.md with the ValuePickr bull/bear raw signal + the discovered transcript links, and leaves clearly-marked TODOs for the Reddit and synthesis steps. It is idempotent — re-runs read the cache under data/apify_runs/_primary_research/<TICKER>/ and re-spend nothing (use --force to re-fetch). Notes:

• ValuePickr topic pages are JS-rendered, so defuddle/capture_web return 0 chars — the script uses

the Discourse `.json` API instead (free, structured, richer). It captures the latest N×20 posts (recent sentiment is what matters); raise --max-vp-pages for more history.

• Transcript PDFs are hosted off-site (investor-relations / primeinfobase / BSE). Reachable ones are

pdftotext-extracted (free); unreachable hosts are recorded as links in the file for you to fetch (browser, or the capture_web Apify fallback) — never silently dropped.

STEP 2 — Reddit sentiment (AGENT-ONLY, Reddit MCP)

A standalone script has no MCP, so you run the Reddit MCP directly (mcp__reddit-search__*):

mcp__reddit-search__search_reddit(query="<TICKER> stock", limit=10, sort="relevance")
mcp__reddit-search__search_reddit(query="<company name>", subreddit="IndiaInvestments")
# then for the best DD/sentiment hits:
mcp__reddit-search__get_post_comments(post_id="<id>")

Useful India subs: IndiaInvestments, IndianStockMarket, DalalStreetTalks. Capture a 3–5 bullet sentiment gist: net bullish/bearish, recurring concerns, any on-ground channel checks.

• Source-weighting (apply to every forum/Reddit/concall voice): rank credibility inversely to how

hard the voice talks its book ([[loudest-promoter-is-least-credible-source]], do-not-talk-up-your-investments). Down-weight the loudest promoters by default; prefer skin-in-the-game, specific, disconfirmable claims over costless cheerleading. A confident price target is the weakest kind of signal; a verifiable on-ground fact (a closed store, a lost client, a delayed payment) is the strongest.

• If the MCP rate-limits (HTTP 403/429 — common): retry with backoff, narrow the query, or fall back

to WebSearch over reddit.com. Confirm the server with mcp__reddit-search__test_reddit_mcp_server.

• Reddit Responsible-Builder policy: corroboration only — no de-anonymising users, no model-training

use, no cross-posting. Read-only.

STEP 3 — Management-call red/green flags (screener transcripts)

Read the extracted transcript(s) in data/apify_runs/_primary_research/<TICKER>/ (or fetch the recorded links). Pull:

• RED: guidance cuts, evasive/non-answers, pricing-power erosion, churn, rising receivables/promoter

pledge, related-party deals, "growth at any cost" tone, blaming macro for everything.

• GREEN: candid acknowledgement of misses, durable pricing power, capital discipline, owner-operator

language, consistent multi-year framing. This is the input that most usefully overlaps with forensic-accounting-redflags. The top tripwire is not generic "earnings-quality concern" — it is cash conversion: does the concall narrative of profit show up as cash? Route any growth-without-cash, rising-receivables, or aggressive-revenue-recognition signal straight to the CFO/PAT check ([[channel-check-as-forensic-tripwire]], owner-earnings-as-quality-proof: high, stable CFO/PAT corroborates management integrity; a widening gap is the lead to chase against the filings in the forensic screen).

STEP 4 — Employee sentiment (optional, low weight)

Check ambitionbox.com / glassdoor.co.in for the company: overall rating, attrition, leadership churn, culture flags. An execution-quality proxy only — weight near zero unless it corroborates a transcript/forum concern.

STEP 5 — Synthesise into the capture file (hedged)

Fill the "Synthesis for the dossier" block: ValuePickr bull/bear (+ open questions), Reddit gist, mgmt-call red/green flags, employee-sentiment note, and a hedged net read ("corroborates / disconfirms the dossier because …, BUT this is the weakest leg — verify independently"). Then point the Stage-1 dossier author and the forensic screen at this file. Do not translate any of it into a conviction/verdict/MoS number — that is the vault's job.

• Read the crowd contrarily ([[crowd-sentiment-read-inversely]], be-fearful-when-others-are-greedy):

record the net sentiment as a contrarian datum — unanimous euphoria is a yellow flag (froth, already priced), capitulation/disgust is a potential window. Subordinate to the gates: a capitulation signal is a window only after the binding moat + owner-earnings checks pass; contrarianism on a genuinely broken business is just catching a falling knife. This datum never overrides the fundamental verdict.

How it plugs into the onboarding pipeline

• Stage 0/1 (dossier & understanding): run this harvester for each NEW stock alongside the

Substack-article capture; the dossier's "Comment-thread red flags" / ground-truth context section cites it. It is context for understanding, parallel to — never above — the binding circle-of- competence consult.

• Forensic screen: transcript red flags + ValuePickr fraud/governance chatter feed

forensic-accounting-redflags as leads to verify against filings.

• It changes no grade. After running, the normal binding consult (--corpus binding) and the v2

conviction/MoS/inversion gates proceed unchanged.

Hard rules

1. Output is LIVE DD context, never binding. Write only to

extracted/research/source_captures/<TICKER>_primary_research.md. Never write vault atomics, never touch the brain index, never set/adjust conviction, verdict or MoS.

1. Reddit is agent-only (MCP). The script cannot do it; you must. Honour the Responsible-Builder

policy and handle 403/429 gracefully (retry/backoff/WebSearch fallback).

1. Reuse, don't rebuild. Generic web URLs route through capture_web (60_ingest_external.py);

prefer free paths (Discourse .json, pdftotext) so Apify spend stays ≈ $0. Never write empty captures — record a blocked source as a TODO link instead.

1. State the weakest-leg caveat in every capture. Community quality varies; this is corroboration /

disconfirmation, requiring human judgment to weight — not a thesis, and not a substitute for the Munger/Buffett CIO.

1. Attribute, quote sparingly, never fabricate. Quote ValuePickr/Reddit by username and date; if a

source returns nothing, say so and count it — never invent sentiment or a transcript quote.

1. Disconfirm or surface a hard fact — don't echo. The edge is disconfirmation

([[primary-research-earns-edge-only-by-disconfirming]]); a pure confirmation of a crowded long is already priced and worthless. But disconfirmation-first must NOT tip into motivated bear-hunting: a manufactured red flag is no more binding than a manufactured bull case (rule 5 still governs). Weight voices inversely to how hard they talk their book. The desk atomics (vault/desk/primary-research/atomic/…) are connective tissue you cite via [[slug]] — you do not create or edit them, and they never cast a verdict vote.