Why your stock analysis should debate itself

Most stock-analysis tools tell you what to do. Quorum's 10-seat Council — five bull seats versus five bear seats, plus a synthesizer — argues with itself first, and refuses to issue a verdict unless the dissent survives into the briefing.

What a "chip" is

Every position on Quorum carries a one-word recommendation called a chip: BUY MORE, HOLD, TRIM, SELL, or DROP. Glanceable, not advisory. The chip is paired with a one-line reason and tagged with its source — the deterministic engine, a Council verdict, or a single-model deep dive.

The chip is the headline. The reasoning behind it is what you should actually read.

The problem with single-LLM stock analysis

Ask any LLM "should I trim AAPL?" and you get a confident answer. Same LLM, same data, different day — different confident answer. The model has no skin in the game. It will tell you what it thinks you want to hear, then a week later tell you the opposite, with the same conviction.

That's tolerable for brainstorming. It's catastrophic for capital allocation. The model isn't wrong — it's just missing the adversarial structure that real markets require. Every trade has a winner and a loser. Analysis that mirrors that asymmetry has to argue against itself before it commits.

How the Council works

The Council is a 10-seat adversarial roster: five bull seats arguing the upside and five bear seats arguing the downside, plus a Synthesizer Chair who composes the final briefing on top of the math. Quorum Lite, Quorum Pro, Quorum Max, and Quorum Ultra-Max share the canonical verdict and Executive Briefing contract. Research depth, seat backbone, prompts, daily limit, and latency vary by tier.

Five bull seats

• The visionary — non-linear upside, optionality, new-market TAM, narrative-driven multiple expansion.
• The moat seat — structural competitive advantage: switching costs, network effects, cornered resources, pricing power.
• The sector seat — structural tailwinds lifting the peer group and why this name is a prime beneficiary.
• The capital-allocation seat — constructive activist lens: M&A, ROIC trajectory, shareholder yield.
• The technical seat — tape-reading, breakout structure, relative-strength accumulation, volume confirmation.

Five bear seats

• The short-seller — accounting red flags, aggressive revenue recognition, executive departures, bull-traps.
• The geopolitical-risk seat — tariffs, sanctions, regulatory crackdowns, jurisdiction risk, supply-chain fragility.
• The macro seat — liquidity, interest-rate regime, credit cycle, demographic headwinds.
• The forensic-filings seat — language-drift in 10-K/10-Q risk factors, hidden liabilities, footnote discrepancies; one verbatim filing citation per call.
• The value-skeptic — overvaluation vs. historical means, margin compression, absent margin of safety.

The Synthesizer Chair

Each of the 10 seats emits a numeric conviction with self-reported uncertainty. Those 10 votes feed a bounded statistical mixture — the bell curve you see on every verdict surface — and the Synthesizer Chair composes the Executive Briefing on top of it: applies profile weighting and veto discipline, names the load-bearing dissents, writes the TL;DR. More on the bell curve here. The chip on the dashboard is the verdict zone with the most probability mass, not a majority vote — and on a contested ticker, the signal itself reads No Conclusion instead of forcing a winner.

Four research tiers, one canonical verdict contract:

• Quorum Lite — all 10 seats on our private GPU. Free tier.
• Quorum Pro — all 10 seats on a faster cloud syndicate with a 15-day shared-verdict cache.
• Quorum Max — frontier-model debate, unlocked by the Pro+ subscription.
• Quorum Ultra-Max — 7 Claude Opus and 3 Codex voting seats, invite only and paced for deliberate runs.

Why this is the right way to use LLM for stocks

Three reasons LLMs are uniquely useful here, and three reasons the naive way to use them fails.

What LLMs add that quant + humans miss

• They read the qualitative. Traditional quant misses CEO transitions, supply-chain narratives, regulatory commentary buried on page 47 of a 10-K. LLMs read the document.
• They scale across providers. A single human analyst has one perspective and one set of priors. Six adversarial models on six different training corpora surface six different blind spots.
• They cite. Every claim in a Council debate is required to cite a specific data slot or web URL. The audit trail is built in.

Why the naive way fails

• Sycophancy. Every frontier model is RLHF-trained to agree with the user's framing. "Should I trim AAPL?" implies you want to trim. The model leans toward yes. Adversarial seat assignment forces the model to argue against the user's prior.
• Single-perspective drift. One model's verdict is one model's bias. Three frontier models + two locals + a proprietary recursive seat — that's six different priors voting under different roles.
• Verdict-as-truth. The output of an LLM debate is not the verdict. It's the dissent that survived. A 4-1 chip with one well-cited dissent is more useful than a 5-0 unanimous, because it tells you the precise condition that flips the thesis.

Real example — BYRN (SELL, HIGH confidence)

BYRN is a public showcase. A deterministic screen flagged 5 red flags; the Council made the bearish case load-bearing.

• Multiple accounting + working-capital warnings aligned. Days-Sales-Outstanding stretching +258% over four quarters wasn't an isolated anomaly — the Council connected it to customer concentration disclosed in the 10-Q.
• The Bull conceded BUY → HOLD under rebuttal. Confronted with the customer-concentration evidence, the strongest BUY case couldn't survive its own named falsifier.
• The Quant flagged an arithmetic error. Another seat misread a percentage; the adjudicator caught it. That's exactly why the role exists.

See the full BYRN Council debate →

And then AAPL (HOLD)

Same system, opposite verdict on a quality compounder.

• The quality case stayed intact after the seats argued moat, durability, and cash-conversion.
• The valuation debate stayed balanced — neither rubber-stamped bullish nor bearish.
• The final chip favored patience. The old deterministic rules used to trim AAPL on every earnings release ("IV-crush risk"); the long-term Council framework HOLDs.

See the AAPL analysis →

What this is NOT

Important framing because the wrong expectation kills trust:

• Not a trading signal. Educational research. Every output ends with "your decision."
• Not better than chance at multi-year forecasting. No LLM is. The Council's value is in the quality of the reasoning trail, not the predictive accuracy of any single chip.
• Designed for selective use. A full Council debate is many model calls in sequence — meant for contested positions and load-bearing decisions, not every position every day. That's why the cloud-Council tiers carry a hard daily cap.
• Not unanimous-by-design. The system is built to surface disagreement. If everything agrees, that's a confidence downgrade, not an upgrade.

What you actually get

• Daily local-LLM baseline on every position (overnight, free, runs on local GPU).
• Weekly deep synthesis of SEC 10-K and 10-Q filings via proprietary engine (200k-token documents read in structured chunks).
• On-demand 10-seat Council (5 bull + 5 bear) for contested calls. Free tier: 20 runs/day on the local council.
• 27 fundamental ratios with sourced thresholds (Piotroski 2000, Beneish 1999, Damodaran, Lynch, Schilit).
• DSO + A/R proxies for customer-payment-health (caught BYRN before its drawdown).
• Daily macro snapshot — yen, dollar, oil, rates, BoJ/Fed/PBOC headlines — auto-injected into every analysis.
• Per-position chip framework configurable by user profile (long-term / swing / active; conservative / balanced / aggressive).
• Custom-weighted basket charts (S&P 5, future themes) vs SPY benchmark.
• Decision log: every chip on every day, source-tagged (Council / engine / re-analyze), fully auditable.

Beta is capped at 10 free seats

Not artificial scarcity. The system runs on a single Tesla P100 and the daily LLM cron is compute-heavy at the union of every user's portfolio + watchlist. Adding users at scale needs GPU we don't have yet.

Right now: free seats are filling. Beyond the cap, you join a waitlist. When the architecture catches up — paid tier, more compute, broader ticker caps — the waitlist gets first access.

Join the beta or get on the waitlist →

Educational research tool. Not investment advice. Output may be wrong. Read the dissent log, not just the verdict.