AI Disclosure

Effective 2026-04-20 · Last updated 2026-06-29

The central fact. RockTurners' analysis is generated by a large language model. LLMs make mistakes — they miss things, misinterpret footnotes, and occasionally invent numbers that sound plausible. Treat every verdict as a starting point for independent research, not a conclusion. You remain responsible for every investment decision you make.

1. What the AI does

For each ticker in a scan, RockTurners performs a two-stage evaluation:

Stage 1 — quantitative screen. Deterministic, code-only. Reads XBRL fundamentals from SEC EDGAR and Financial Modeling Prep and applies numeric filters (market cap, margins, growth, balance-sheet health, and so on) from the strategy definition. No AI involvement.
Stage 2 — qualitative evaluation. Sends the strategy's system instructions, a shortlist of tickers that cleared Stage 1, and source documents (10-K risk factors, MD&A excerpts, earnings-call transcripts, Form 4s) to an AI model. The model returns a structured JSON verdict, per-dimension scores, citations, and a summary.

The verdict is then mapped to a category (STRONG_MATCH, PARTIAL_MATCH, NO_MATCH — shown as Gemstone, Geode, and Sediment) per the strategy's thresholds and displayed with the evidence that drove it.

2. Models we use

You choose which AI engine runs an analysis — a sensible default, a per-run choice, or a saved default in your account. We offer only models a RockTurners operator has reviewed, priced, and enabled in our model catalog, on a provider we have configured. The provider and model that produced a result are stamped on its output.

Providers. Anthropic (Claude), OpenAI (GPT), xAI (Grok), Google (Gemini), DeepSeek, Mistral, and OpenRouter (a router that forwards a request to an upstream provider). Only enabled, configured providers appear in the engine selector.
Default engine. The lowest-cost model in the standard tier — we default to the cheapest option that meets the analysis spec, not the most expensive.
Web search. Anthropic's hosted web_search_20250305 tool is available only on the Anthropic path, and only where a strategy enables it. The other providers run on the supplied filings — no live web access. JSON return via structured output applies to every provider.
Token ceiling. 2,000 output tokens minimum when web search is enabled; 90-second per-attempt timeout (up to roughly 180 seconds including retries).
Concurrency. Sequential per ticker; calls are never parallelized inside a single analysis loop, which prevents partial-result loss on timeout.

We upgrade to newer model versions when they demonstrably improve output quality on our internal regression suite, and we select the cheapest model that meets the spec. We do not silently downgrade. The output stamp shows which provider and model produced it.

Some providers process data outside the United States — for example, Mistral in the EU and DeepSeek in China; OpenRouter forwards to upstream providers in a range of jurisdictions. See the Privacy Policy § International transfers for where your data may be processed and the safeguards that apply.

3. Grounding sources

We feed the model information from these sources, and nothing else:

SEC EDGAR — 10-Ks, 10-Qs, 8-Ks, DEF 14A proxies, Form 4 insider transactions. Primary filings, not summaries.
XBRL fundamentals — structured financial data tied to filing line items.
Earnings-call transcripts — where licensed.
Financial Modeling Prep — market quotes and derived fundamental ratios.
Web search results — on the Anthropic path, where the strategy explicitly enables that tool. Citations are captured and displayed alongside the analysis.

Model output includes citation markers (for example, <cite index="2-3"/>) that link back to the source span used. If a claim has no citation, it was inferred without a primary-source anchor and should be scrutinized.

4. What the AI is not doing

It is not trading, proposing trades, or acting as a fiduciary.
It does not have access to your brokerage, bank, or tax information.
It does not see non-public information. It reads only what you can read.
It does not track markets in real time. Quotes are delayed per the provider licenses, and analysis runs on the snapshot available at call time.
We do not use your custom strategies, portfolios, or analysis output to train models — ours or anyone else's.

5. Known failure modes

Hallucinated numbers. LLMs occasionally generate financial figures that look plausible but are not in the source filings. We reduce this risk by forcing citation markers on every numeric claim and by cross-checking against XBRL when possible — but it still happens. If a number looks wrong, it might be wrong.
Stale filings. Filings can lag real events by weeks or months. An analysis produced today may miss a material disclosure made yesterday. Scans show the analysis timestamp — if it is older than the strategy's staleness window, we flag it as stale in the interface.
Strategy misuse. Strategies have philosophical shape. A deep-value strategy will rank a momentum stock poorly and vice versa. If you run a Fisher-style scuttlebutt strategy against a commodity producer, the output will look unflattering regardless of merit. The strategy is the lens — choose the right lens.
Unknown unknowns. The model reports gaps via an explicit unknowns field in its output, but only the gaps it notices. There are blind spots it will not flag. We publish the honest-gaps section on the home page precisely because we would rather under-claim than over-claim.

6. Guardrails we enforce

Server-side only. Every AI provider API call is made from our backend. Provider API keys never reach the browser.
Structured output. Responses are constrained to a JSON schema; free-form prose is limited and checked against that schema.
Prompt-injection defenses. User-supplied content — a custom strategy name or notes, for example — is never concatenated into instructions that could alter analyzer behavior.
Per-call cost caps and timeouts. Runaway or stuck calls are cancelled at 120 seconds. Monthly caps stop uncontrolled spend; estimated cost is shown before any paid scan starts.
Audit logging. AI interactions are logged (tokens, cost, model, verdict) without logging sensitive user data.

7. How to tell when AI wrote something

Any content produced by Stage 2 is AI-generated — including the summary, thesis, risks, and catalysts fields, the per-dimension evidence notes, and the final verdict justification. These are visually framed in the interface so you know what you are reading. Stage 1 numeric thresholds, strategy metadata you authored, and your own holdings data are not AI-generated.

8. Your options

Run a scan with Stage 2 disabled to get only the deterministic Stage 1 shortlist.
Switch models per strategy (Sonnet for broad scans, Opus for deep dives, for example) to balance cost and depth.
Export your scan results and verify citations against the linked SEC filings. We make this easy on purpose.
Delete any AI output associated with your account at any time — scans are scoped per-user and removable.

9. Reporting an AI mistake

If you see an AI output that is factually wrong — a fabricated number, a misread filing, a miscategorised company — email [email protected] with a link to the scan result. We use these reports to improve prompts and add regression tests. We do not delete old outputs in response to market outcomes ("the verdict turned out wrong"); we only correct factual errors.

10. Changes to this Disclosure

We'll update this page whenever the underlying model, grounding sources, or guardrails change materially. The effective date above reflects the current version.

If you only remember one thing: the AI is a very fast, well-read intern with a confident tone. Brilliant at reading 10-Ks. Occasionally wrong. Always worth double-checking before you put money on it.