Pineal Capital
Deep-value research · v2.0.4

Pineal Capital's AI Equity Analyst v2.0

Built by Grand Canal Capital Partners · grounded in SEC EDGAR filings · 11-section IC memoranda across the Peter Furlong test set.
Tier 1 Complete Internal Demo

A production-testable deep-value equity research system.

Real SEC 8-K catalyst discovery live. Qualitative synthesis grounded in 10-K §1 / §1A / §7 footnotes. Peter Furlong test set — TDOC / PRKS / FIVN — externally shareable. Click any ticker card below to read the full IC memo.
PETER / CONOR READY
Tier 1 complete · 14 commits
14
Commits today
~1 every 75 min
56/56
Tests passing
+1 gracefully skipped
9/10
Stress-test clean
Up from 4/6 baseline
13.6k
Lines of Python
12 operational components
10
Tickers validated
US-listed via SEC EDGAR
~3 s
Per-ticker chain
Full /research workflow

The journey so far — v0 → v2.0.4

v0.5
Full chain
v0.9
Usable prototype
v0.95
TDOC unblocker
v0.96
Analyst-grade
PF set
TDOC/PRKS/FIVN
v2.0
LLM layer
Skill
Max 20x path
v2.0.1/2
Cowork review
v2.0.3
Stress bugs
v2.0.4
Real 8-K + FIVN

Sample output — click any ticker to read the full IC memo

TDOCMONITOR
Teladoc Health · healthtech · Band C
Price: $5.69 · SOTP FV/share: $29.57
Entry: $23.65 · MoS: +316%
Owner earnings: +$285M · Yield: 4.8%
Conviction: 6/10 (analyst override)
Real catalysts: Mgmt refresh Mar 2026, Earnings Feb 2026
Thesis: Two-segment SOTP mispricing — Integrated Care (102M members, +9% YoY) alone is worth more than current EV. MONITOR pending BetterHelp paying-user stabilisation + no-new-impairment quarter.
▸ Open full IC memo
PRKSINVEST
United Parks & Resorts · hotels_hospitality · Band C
Price: $36.07 · SOTP FV/share: $104.07
Entry: $83.26 · MoS: +131%
Expected 3y IRR: +35% (prob-weighted)
Conviction: 6/10
Real catalysts: Mgmt refresh Jan 2026, Earnings Feb 2026
Thesis: Last independent mid-cap theme-park operator post Cedar Fair / Six Flags merger. Zoological moat + consolidation optionality. Leverage risk flagged at 2.7x interest coverage FY2025.
▸ Open full IC memo
FIVNPASS
Five9 · tech_saas · Band D
Price: $15.78 · SOTP FV/share: $10.39
Entry: $7.80 · MoS: −51%
Expected 3y IRR: −27% (prob-weighted)
Conviction in call: 8/10
Real catalysts: Mgmt refresh Jan 2026, Earnings Feb 2026
Thesis: AI-cannibalisation risk unpriced — Item 1A flags need to replace seat revenue with AI modules. Market paying premium for inverted AI optionality. Revisit <$10.
▸ Open full IC memo

What's built — component architecture

FPR Module
LIVE · 9 sub-agents · 3,800 LOC
Orchestrator
LIVE · /research chain
Thesis Builder
LIVE · 3 U's, checklist
Catalyst Tracker
LIVE · real SEC 8-K (v2.0.4)
Valuation Screener
LIVE · SOTP + scenarios
Risk Monitor
LIVE · sizing + exposure
Compliance Checker
LIVE · 3-track QIAIF
IC Paper
LIVE · 11 sections
LLM Layer
LIVE · Claude SDK + fallback
10-K Text Fetcher
LIVE · Item 1/1A/7 extract
8-K Fetcher
LIVE · 13 item codes (v2.0.4)
Pineal-Research Skill
LIVE · Max 20x path
Pineal Bootstrap
LIVE · task classifier
Morning Briefing
SCAFFOLD · needs API key
Catalyst Sentinel
SCAFFOLD · hourly scan
Thesis Challenger
SCAFFOLD · weekly bear case

Two LLM integration paths — Max 20x for research, API for automation

Interactive Research (primary) — Max 20x subscription
User types /pineal-research TDOC in Claude Code or Cowork. Skill runs pipeline + fetches 10-K text + Claude Code itself synthesises §3/§4/§7 using the user's existing Max 20x quota.
Per-ticker cost$0included in Max 20x
Latency~30 sresearch + synthesis
ModelOpus 4.7latest available
Unattended Automation (future) — Anthropic API
Scheduled agents run without a human: morning brief 07:00, catalyst sentinel hourly, thesis challenger weekly. Code in place; activates when ANTHROPIC_API_KEY is set.
Cold cache~$0.60per ticker run
Warm cache~$0.15per ticker run
10 tickers/day~$1.50cache warm

Capabilities matrix — what's real vs what's next

CapabilityStatusEvidence / Notes
Deterministic FPR scoringLIVE9 sub-agents, 8 sector adapters. TDOC Band C, MSFT Band A, JACK Band F — all defensible.
SEC EDGAR financials ingestionLIVEXBRL primary tier; 10 tickers validated; 90% clean-run rate.
10-K footnote extractionLIVEItem 1 / 1A / 7 sections, 24h disk cache.
Real SEC 8-K catalyst discoveryLIVEv2.0.4 · 13 item codes mapped to Pineal categories. JACK 5 real catalysts incl. April 2026 mgmt refresh.
Sector-routed SOTP valuationLIVEPer-ticker overrides: TDOC (IC + BetterHelp), PRKS (Admissions/Spend/Season-Pass), FIVN (Cloud + Services + AI).
Current-price fetch + MoSLIVEYahoo Finance endpoint, 5s timeout, silent fallback.
Scenario-weighted valuationLIVEBull / base / bear with band-keyed probabilities + 3y IRR + expected return.
Owner earnings (Buffett)LIVEOCF − maintenance capex with year-alignment.
Exit discipline (§9)LIVETarget price + stop-loss + re-rate triggers.
LLM-synthesized §3 / §4 / §7LLMVia Max 20x skill for interactive; via API for automation.
Pineal-config parameter layerLIVE15 parameters env-var overridable. Conor call = one PR edit.
Compliance 3-track (QIAIF)LIVEFund structure + restrictions + ESG placeholder.
3 U's with real analyst / ownershipNEXTHeuristic today; yfinance wire is ~2h (Tier 2).
Scoring anchor calibrationNEXTMSFT/AAPL at 100/A a bit too generous.
Business-description DB expansionNEXT10 curated today; target 30-50.
Memory knowledge graphNEXTEach /research starts fresh; ~6h.
Eval harness (returns backtest)NEXT50-ticker labelled dataset to prove alpha.
Background agents liveNEXTMorning brief + sentinel + challenger — requires API key.

Methodology deep dive

Methodology matrix — what makes v2.0.4 different 12 analytical dimensions ViewHide ▸
Analytical dimension v0.0 spec v0.5 pipeline v2.0.4 (today)
Data source ChatGPT training memory XBRL / stockanalysis.com XBRL + SEC EDGAR 10-K text + 8-K filings + Yahoo Finance live price
Financial statements None — AI guesses numbers Reported only Reported + adjusted (SBC stripped) + normalised
Qualitative analysis AI writes from memory Python template strings Claude reads 10-K §1 / §1A / §7 and writes grounded synthesis
Valuation Not computed Single-point DCF Scenario-weighted (bull / base / bear) + per-ticker SOTP + band-keyed bear multipliers
Conviction / confidence Not addressed Band score only Dual axis: Financial Quality 0–100 × Conviction 0–10. Analytical vs invest conviction split.
Owner earnings Not addressed FCF only Buffett OCF formula: OCF − maintenance capex. Maintenance vs growth capex split.
Catalyst framework Not addressed 5 generic placeholders Real 8-K filings parsed — 13 item codes → dated events from SEC EDGAR
Falsification Not addressed Not addressed Kill-the-thesis: 3 falsifiable tests per ticker with monitoring triggers
Exit discipline Not addressed Not addressed Entry price + target + stop-loss + re-rate triggers
Sector awareness Generic template Default adapter only 7 adapters: healthtech, hotels, SaaS, banks, staples, industrial, healthcare services
Red-flag detection Not addressed Basic threshold checks 9 numeric triggers with severity bands (SBC load, cash conversion, ROIC < WACC, leverage spike, etc.)
Output format Chat message JSON + basic markdown 11-section IC memorandum with methodology hash, source log, pre-output checklist
Design lineage — Buffett · Klarman · Burry · Dalio framework grounding + implementation status ViewHide ▸
Buffett / Munger

Owner earnings & moat

OCF-based owner earnings (not FCF, not net income). Maintenance vs growth capex split. Durable competitive-advantage assessment in §4 investment thesis. Circle-of-competence check via sector-adapter routing.

Implemented · v2.0.1
Klarman

Margin of safety & catalysts

Source-keyed unknowns discount (not a flat 15%). Scenario-weighted downside analysis with band-keyed bear multipliers. Re-rating catalyst framework built on real 8-K events from SEC EDGAR. Explicit admission of unknowns.

Implemented · v2.0.4
Burry

Forensic filing analysis

10-K text extraction (Item 1, 1A, 7). SBC adjustment bridge (reported → adjusted → normalised). Accrual-ratio monitoring. Revenue-recognition pattern detection via LLM synthesis on filing text.

Partial · deeper forensics on roadmap
Dalio

Scenarios & risk architecture

Bull / base / bear scenario-weighted fair value. Band-keyed probability ladders (higher-quality bands get higher base-case weighting). 9 numeric red-flag triggers with severity bands. Regime classification and risk-parity sizing on Tier 3 roadmap.

Partial · regime / parity Tier 3
Honest assessment — what's real today vs what still needs work pipeline vs skill-synth · strengths · gaps ViewHide ▸
Pipeline only (no LLM synthesis)
~30% genuinely analytical70% structural template
Skill-synth path (Claude reads 10-K + synthesises)
~65% genuinely analytical35% structural

What an IC would trust today

  • Grounded qualitative analysis from actual SEC filings
  • Per-ticker SOTP with reported segment overrides
  • OCF-based owner earnings with maintenance-capex split
  • Band-keyed bear scenarios with realistic wipeout tails
  • Real 8-K catalyst events dated from EDGAR
  • Analyst override capability (TDOC INVEST → MONITOR)
  • Kill-the-thesis with 3 falsifiable tests per ticker
  • Conviction split: analytical vs investment confidence
  • Exit discipline: entry, target, stop-loss, re-rate

What still needs work

  • Scoring anchors a touch generous (MSFT / AAPL / NFLX all hit 100 / A)
  • 3 U's evidence partly fabricated — PASS / FAIL without hard data
  • No eval harness — can't prove methodology predicts returns
  • No memory across runs — each ticker starts cold
  • Business-description DB covers 10 tickers; target 30–50
  • No insider transaction tracking (Form 4 / 13F)
  • No regime classification or risk-parity sizing
  • Pipeline conviction scores remain band-keyed heuristics

Next steps — what Pineal should drive

Immediate — with Peter

30-min alignment call

  • Fill 15 open Pineal-config parameters (3 U's, unknowns discount, risk caps, fund structure, ESG source)
  • Confirm Bloomberg + AlphaSense credential path
  • Decide API key provisioning timing
  • Watchlist expansion priority
This week — development

Tier 2 technical work

  • Scoring anchor calibration (~2h)
  • yfinance 3 U's data wiring (~2h)
  • Expand business_descriptions.json to 30+ tickers (~1h)
  • Per-ticker SOTP overrides for watchlist (~1h)
Next sprint — strategic

v3.0 scope

  • Memory knowledge graph (ticker state)
  • Eval harness against 50-ticker labelled returns
  • Background agents live
  • MCP data layer: Bloomberg + AlphaSense

Bottom line for Peter

The system produces genuinely different, evidence-grounded calls for different tickers — TDOC MONITOR (deep-value +316% MoS, held back by BetterHelp decline), PRKS INVEST (scarcity-value theme-park operator at +131% MoS), FIVN PASS (AI-cannibalisation risk unpriced). Three papers, three reasons, all anchored in actual SEC filings including real 8-K catalyst events (dates, items, accession numbers). Peter sits at a Cowork session, types /pineal-research TDOC, and gets an analyst-grade IC memo in ~30 seconds using his existing Max 20x subscription — zero additional infrastructure cost until automation is needed.