How Three·Lens reads a company

Section 1

How a reading is built

Each "scan" runs a fixed sequence. The expensive, AI-written part comes last — only after the facts have been gathered and computed:

1. Resolve the company. Your input is matched to a real listing — the SEC's master ticker file for the U.S., a market-data lookup for India and Canada. Light misspellings are tolerated, and when the match is ambiguous a short list of close names is offered so you can confirm the right company.
2. Pull market data. Current price, market capitalisation, and shares from the market-data provider.
3. Compute fundamentals. Revenue, margins, returns, cash flow, and balance-sheet figures, on a trailing-twelve-month basis, with multi-year trends.
4. Read the latest annual filing. For U.S. companies, the most recent 10-K is parsed — business, risks, management discussion, properties — into a grounded, paraphrased summary.
5. Gather recent news. A short, cited summary of substantive recent developments (routine noise is filtered out).
6. Map operations. For physical-footprint sectors, a state-level location map drawn from the filing.
7. Find peers. A few comparable companies, each given a metrics-only sub-read to populate the comparison table.
8. Synthesise the reading. One AI pass turns the vetted facts into the Value / Growth / Quality reading, plus the bull and bear cases and forward-looking notes.

For U.S. companies, a separate pass produces the Results Read — a reading of the latest reported quarter drawn only from that quarter’s filing: its headline metrics, the key statements management made in the earnings release, and the full quarter financials. It is kept deliberately distinct from the trailing-twelve-month figures above, so a single-quarter number is never mistaken for the annual one.

After a reading is written, a second, low-cost model re-checks every figure in it against the exact facts it was built from and flags any it can’t match — a guard against a confident number that isn’t grounded. Flagged figures are surfaced on the reading for you to verify; nothing is removed.

Section 2

The three lenses

Every company is read three ways. Most look different through each lens — cheap on one, weak on another — and those gaps are exactly what the framework is designed to surface. Two careful investors can weigh the same three readings and still reach different conclusions, and that's expected: Three·Lens presents the perspectives, not a single “correct” answer.

• Value — Is it cheap relative to what you get? The reading anchors on what the current price implies about the business, comparing valuation multiples (such as price-to-earnings) against the fundamentals. It never leans on analysts' price targets or a synthesised "consensus."
• Growth — How fast, and how durably, is it expanding? Revenue growth and multi-year trends, with the durability question kept in front.
• Quality — How profitable, well-run, and financially sound is it? Returns on capital, margins, cash generation, and balance-sheet strength.

Around the three lenses, a reading also names the tension between them (say, high quality at a demanding price), what would have to be true for the optimistic case, a bull and a bear case, and what to watch next.

Section 3

The figures we use

All figures are computed deterministically in code — not by the AI — on a trailing-twelve-month basis. Multi-year growth rates (CAGR) are calculated in code as well, and are never fabricated across a swing from loss to profit.

A reading draws on the standard set an analyst would expect across four areas: valuation (price, market capitalisation, and the usual multiples such as price-to-earnings and EV/EBITDA), growth (revenue and its multi-year trend), profitability and returns (margins, and returns on equity and on capital), and balance-sheet strength (debt, liquidity, cash, and free cash flow) — plus a few quality signals such as shareholder yield. Where a figure is approximate or rests on assumptions — an enterprise value, a cash runway — the reading says so.

The AI is required to cite these verbatim and is barred from recomputing, re-deriving, or approximating them — a rule enforced both in the prompt and by a code check, so a number you read in a lens is the number we computed.

Section 4

Data sources

• SEC EDGAR — U.S. fundamentals (from structured XBRL data) and filings. Free and primary-source.
• Twelve Data — market prices and shares for the non-U.S. markets, plus fundamentals for India and Canada and as a U.S. fallback.
• Perplexity — recent-news summaries, with citations.

Coverage. Three live markets — the U.S. (NYSE / NASDAQ), India (NSE / BSE), and Canada (TSX).

Timeliness. Prices may be delayed by roughly 15 minutes in some markets and are effectively end-of-day in others — appropriate for long-term reading, not for trading.

Provenance. When a fundamental comes from a vendor feed rather than an audited filing, the reading flags it — and caps its own confidence accordingly (see below).

Section 5

How the AI is used — and constrained

AI (Anthropic's Claude) turns the vetted facts into plain-language teaching. It is held to a small set of non-negotiable rules:

• Ground everything in the provided facts. If a figure isn't in the inputs, the reading says so rather than guessing — and a gap is never treated as if it were meaningful about the company.
• No advice, no prediction. Never buy / sell / hold; never a price or return forecast. Analysis and tensions only.
• Cite figures verbatim; never recompute.
• Editorial, not promotional. A reading is written like a skeptical internal memo — willing to say "this looks expensive" or "we're not sure" — not a sales pitch.
• Paraphrase, never reproduce. News headlines and filing passages are summarised in the AI's own words.

We use a stronger model for the synthesis and the tutor chat, and a lighter, faster one for the per-metric teaching, quizzes, and peer lookups — each task matched to the right tool.

Section 6

Confidence & evidence

Each lens carries a verdict and a confidence level, rated on its own evidence:

• High — the verdict follows directly from hard metrics.
• Medium — it rests on interpretation.
• Low — the facts barely cover it.

If the underlying fundamentals are vendor-sourced or unaudited, no lens may be rated High — it is capped at Medium and the limitation is named. That cap isn't left to the model's goodwill; the system applies it independently.

Every lens shows its evidence as discrete facts, each tagged by its role in that verdict — supporting, against, mixed, or context — so you can see what's driving the read, not just the conclusion. And the same judgment is written at three depths — plain-language, investor, and advanced — where the depth toggle changes the wording and detail, never the verdict itself.

Section 7

Freshness & caching

Generating a reading is costly, so completed readings — and the market and filing data behind them — are cached and shared across users in a de-identified form, keyed by company and filing, never tied to you. The first person to scan a company pays the cost; everyone after reuses it until it goes stale.

A reading is refreshed whenever the company's situation moves materially — a new annual filing, substantive news, or a meaningful price move — and whenever we improve the method itself. So what you read reflects both the latest facts and our current approach.

Section 8

Limitations & honest caveats

We'd rather you know the edges of the method than discover them later.

• AI can be wrong. It can omit, misread, or sound confident while mistaken. Verify anything important against the primary filing before relying on it.
• It's only as current as its inputs. Fundamentals depend on the latest filed data; between filings, the picture can lag reality.
• Coverage isn't uniform. U.S. companies get a deeper read because SEC filings are richly structured and free; Indian coverage leans more on market data and may carry less business-model context.
• Vendor data is second-best. Unaudited feed data is flagged and rated down relative to audited filings.
• Prices aren't real-time. See Data sources.
• A lens is not a verdict. The framework organises how to think about a business; it doesn't tell you whether it's a good or bad investment — and nothing here is advice.

Questions about the method? Contact us. The full disclaimer is in our legal & privacy page.