JourdanLabsCOSMIC x CIPHERDeterministic finance
beats fluent guessing.
Same EDGAR facts. Same task. Same output requirements. ALCHEMIST wins where production finance actually cares: reproducibility, auditability, refusal discipline, and cell-level method consistency. CIPHER is the DCF model inside the suite.
This page is built for live demos against GPT, Claude, Gemini, Grok, or any other model. Paste their answer into the challenge box and score it against the same ALCHEMIST discipline checks.
ALCHEMIST keeps the math stable. LLMs have to prove they can.
| System | Source trail | Formula method | Refusal discipline | Audit chain | Deterministic |
|---|---|---|---|---|---|
| ALCHEMIST model suite | SEC-backed | sealed | explicit | LUNA chain | yes |
| Frontier LLM answer | must prove | must disclose | must show | usually absent | no |
| Spreadsheet template | manual | visible | manual | version drift | partly |
| Legacy data terminal | licensed | black box | limited | export-dependent | opaque |
ALCHEMIST's edge is not a prettier answer. It is same input, same output, same source trail, every time. CIPHER is the DCF engine; COMPS, LBO, Credit, SOTP, and the rest are adjacent ALCHEMIST models that inherit the same source/refusal/audit discipline. The legacy DCF-vs-LLM harness lives in `score_llm.py`; this UI turns the same idea into a demo-ready challenge for every ALCHEMIST model. Existing CIPHER engine disclosures record a May 2, 2026 frontier comparison where the CIPHER composite was 21/22 (95.5%) and frontier derived-ratio scores were GPT 17.1%, Claude 22.9%, Grok 14.3%, and Gemini 22.9%. Fresh model outputs should be pasted below when demoing live.
Paste any LLM answer. Score it like a model-risk reviewer.
Build a banker-grade comparable-company analysis for NVDA. Select 5 public peers; show EV/Revenue, EV/EBITDA, EV/EBIT, EV/FCF, P/E, P/S, P/B, revenue growth, EBITDA margin, ROE, ROIC, and Net Debt/EBITDA; state every formula; refuse unsafe denominators; flag fiscal-year mismatches; cite data provenance; and end with reproducibility notes.
Resolves a peer set, computes only defensible multiples, refuses non-meaningful denominators, flags fiscal-year mismatches, and exposes source/method provenance.
Each slot persists in this browser for the selected challenge, so you can paste all four and switch back without losing the bakeoff.