E5: a blind test and stepwise optimisation of the Jirgenson Model operating prompt on ten small-business crisis scenarios

A blind, pre-registered test of whether a language model answers small-business crisis messages better when it is given the canon of the Jirgenson Model and its operating prompt, and the complete record of the rounds of improving that prompt (rounds 0, 1, 2, 3 and 5) while the canon stayed fixed. Design. Ten frozen scenarios, each with an answer key (gold arithmetic, mandatory elements, forbidden items, word and question budgets). Conditions: A, no system prompt; B, a generic structured-adviser prompt; C, canon + operating prompt. Generator Claude Sonnet 5.5, three answers per scenario and condition; blind pairwise judge Claude Opus 5.5, every pair in both orders; score 0–100; noise band ±1.51 points. Results. The untuned first version v1.0 scored 85.9 against 55.3 without a prompt (+30.6, 95% CI +26.7 to +34.4) and 67.7 with the generic prompt. The final version v1.5-c scored 88.8 against 56.2 (+32.5, +26.6 to +38.5), better in 10 of 10 scenarios and preferred in 30 of 30 pairs. Answers with at least one judge-flagged invented fact: 70% without a prompt, 0.8% with v1.5-c; answers with a violation: 53% and 0%. The total-score gain of v1.5-c over the previous version v1.3-c is not robust (+1.71 primary, +0.86 within shared judge calls); a judging-context effect that explains the difference is documented. Two of the scenarios were posed in Latvian and Russian to test that the answer keeps the user's language; in this record they and the answers to them are given in English translation, so the whole record is in English. Limits. The prompt was tuned on the same ten scenarios; a sealed hold-out set (SHA-256 published) has not been measured; one model family generated and judged; the answer key was written by the Model's author; no human raters. Contents. Every operating-prompt version tested, the canon, scenarios and answer key, judge prompt, the exact generator and judge inputs, all 240 raw answers, all 600 judge records (plus 9 re-judged), frozen score tables, hard checks, figures, analysis scripts, judge calibration and a second-judge spot check, and the working notes (protocol, report, run log, decisions). Start with STUDY.md. The record is split into two archives: E5_record_v1.0.zip (everything except inputs/) and E5_inputs_v1.0.zip (inputs/); unpack both into the same folder. MANIFEST_sha256.csv lists a SHA-256 for every file. Licences. Canon and operating prompts: CC BY-ND 4.0. Data, documentation and other text: CC BY 4.0. Scripts: MIT. See LICENSE.md. Companion to the book Martin Jirgenson, Steer, Don't Guess. The Jirgenson Model (2026), ISBN 979-8-1648-3374-9. Not financial, legal or investment advice.