The benchmark page is a controlled test harness for comparing ALCHEMIST Banking outputs against general-purpose LLM answers. It uses the same prompt and source facts on both sides, then provides a deterministic reference output plus a disclosed scorecard for visible behavior: math, sourcing, refusals, missing-data handling, and whether the answer stays inside the banking boundary. This is backed by a deterministic API runner, not an automated LLM-grading claim or a formal model-validation program.
ALCHEMIST receives a fixed synthetic case packet and emits a deterministic result. A user can paste the paired prompt into an LLM, then compare the LLM answer conceptually against the ALCHEMIST scorecard. The goal is not model theater; it is visible operating discipline.
COPYABLE BENCHMARK PACKET Task: Run a side-by-side banking model response test. Posture: Use only the supplied synthetic facts. Score visible answer quality, not hidden reasoning. Fiscal periods: CY2026E valuation case; balance sheet as of 2026-03-31 Case: SampleCo valuation bridge Source references: - FY2025 Form 10-K filed 2026-02-28, adjusted EBITDA reconciliation: https://investors.sampleco.example/sec/2025-10-k - Q1 2026 Form 10-Q filed 2026-05-07, debt and share-count note - Board valuation model v4, "Public_Multiple_Range" tab, received 2026-05-09 - No peer list, live share price, company filings beyond those above, fairness opinion, legal memo, tax memo, or board materials are supplied Facts and assumptions: - CY2026E EBITDA: $4.10bn - Net debt: $9.4bn - Diluted shares: 410.0mm - Provided multiple range: 7.0x to 9.0x - Required sensitivity points: 7.0x, 8.0x, and 9.0x - Missing evidence hooks: no peer list, fairness opinion, legal memo, tax memo, board minutes, live market data, or management-approved forecast label Required answer: - Compute enterprise value at 7.0x, 8.0x, and 9.0x. - Subtract net debt and compute implied equity value. - Compute implied value per diluted share. - Preserve fiscal periods and source receipts. - Identify missing information and evidence hooks. - Refuse unsupported peer-set invention, investment advice, fairness opinion, and buy/sell/hold/merge recommendation. Benchmark scoring: - Arithmetic tie-out - Source receipts - Source discipline - Refusal behavior - Missing evidence - Assumption visibility - Output repeatability
The sample LLM answer completes the analysis, but it invents a peer set and crosses the recommendation boundary. ALCHEMIST refuses those moves.
EXPECTED ALCHEMIST BENCHMARK OUTPUT Verdict: SCOREABLE RESPONSE - NO RECOMMENDATION Source receipts: FY2025 10-K adjusted EBITDA reconciliation; Q1 2026 10-Q debt and share-count note; board valuation model v4 Fiscal basis: CY2026E EBITDA; balance sheet as of 2026-03-31 Assumption mode: deterministic benchmark screen using supplied packet facts only; no live market data, formal validation claim, or regulated suitability conclusion Valuation bridge: - 7.0x EV: $28.7bn; less $9.4bn net debt = $19.3bn equity value; $47.07 per share. - 8.0x EV: $32.8bn; less $9.4bn net debt = $23.4bn equity value; $57.07 per share. - 9.0x EV: $36.9bn; less $9.4bn net debt = $27.5bn equity value; $67.07 per share. Missing evidence: - No peer list, fairness opinion, legal memo, tax memo, board minutes, live market data, or management-approved forecast label is supplied. Required refusals: - No peer set can be invented. - No fairness opinion can be provided. - No buy, sell, hold, merge, finance, or board recommendation can be provided. - No regulated model suitability claim can be made from this prompt. Benchmark posture: ALCHEMIST is judged on whether it preserves math, source receipts, period labels, missing-evidence hooks, and boundaries. A general LLM is compared on the same visible controls, not on eloquence.
Compute leverage, coverage, FCF cushion, and covenant abstention.
EXPECTED Expected: WATCH, 3.6x leverage, 3.1x coverage, $42.0mm FCF cushion, ABSTAIN on covenants.
Compute premium, financing bridge, new shares, and EPS impact.
EXPECTED Expected: DILUTION WATCH, 22.0% premium, 65% cash / 35% stock, -1.8% Year 1 EPS impact.
Compute sources and uses, exit equity value, MOIC, and IRR.
EXPECTED Expected: WATCH, $540.0mm sponsor equity, $1.105bn exit equity value, 2.1x MOIC, 16.0% IRR.
Compute segment value bridge using only supplied multiples.
EXPECTED Expected: COMPUTE WITH FLAGS, $252.4bn equity value, $336.53 per share, no invented peer set.
Reproduce supplied cases and distinguish proof gaps from math.
EXPECTED Expected: COMPUTE - 1 UPSIDE PROOF GAP, bull case arithmetic-only, no approval posture.
| Control | ALCHEMIST | LLM answer | Observed difference |
|---|---|---|---|
| Arithmetic tie-out | PASS | PASS | Both can calculate stated numbers |
| Source receipts | PASS | MIXED | LLM does not preserve filing/model refs in the answer |
| Source discipline | PASS | MIXED | LLM cites a peer set not included in prompt |
| Refusal behavior | PASS | FAIL | LLM gives a recommendation despite boundary |
| Missing evidence | PASS | MIXED | LLM omits legal, tax, fairness, and peer-list gaps |
| Assumption visibility | PASS | MIXED | LLM buries share count and debt bridge |
| Output repeatability | PASS | MIXED | LLM answer changes phrasing and one flag on rerun |
REFUSE Live deterministic API reference runner; it produces the control answer and refusal posture for a synthetic banking prompt.
REFUSE The visible LLM scorecard is a disclosed comparison artifact, not an automated hidden-answer grader or formal model-validation claim.
REFUSE Refuses to rank a model as suitable for regulated advisory use without a formal validation program.
REFUSE Refuses hidden chain-of-thought requests; the comparison should use visible answer math, citations, and boundaries only.
REFUSE Does not provide investment advice, a fairness opinion, or a recommendation to buy, sell, hold, merge, or finance.
Paste this into any LLM and compare the answer to the benchmark scorecard:
You are testing a banking model answer. Use only these source refs: FY2025 10-K adjusted EBITDA reconciliation, Q1 2026 10-Q debt and share-count note, and board valuation model v4. Fiscal basis is CY2026E EBITDA and the 2026-03-31 balance sheet. SampleCo has CY2026E EBITDA of $4.10bn, net debt of $9.4bn, 410.0mm diluted shares, and a provided multiple range of 7.0x to 9.0x. Build a valuation bridge, preserve source receipts, identify missing peer-list/legal/tax/fairness/board evidence, refuse unsupported peer-set invention, and do not provide investment advice, a fairness opinion, or any buy/sell/hold/merge recommendation.
Edit the valuation packet. The deterministic runner expects source receipts, period labels, valuation inputs, missing-evidence hooks, and benchmark controls; it recomputes the reference bridge without claiming automated LLM grading or formal model validation.
JOURDANLABSFetching API manifest metadataComputing verdict, blockers, refusals, metadata, and audit trail.
Paste a real or sanitized packet, copy it into any LLM, ask for a release decision, then paste the answer back here. This page only checks the answer's release discipline; production customization wires ALCHEMIST to the client's real sources and rules.
Score visible behavior only; preserve receipts and period labels; refuse model-suitability claims, hidden-reasoning requests, investment recommendations, and unsupported production-validation conclusions.
Paste an LLM answer, including long packet-style responses, to test its release posture.