How We Score

The composite formula

The mechanical layer is deterministic code — same input, same output, every time, no AI. The editor layer is an LLM persona judging what code cannot: strategy, coherence, fit. Publishing both sub-scores means you can see exactly how much of your grade came from measurable facts versus the judge's opinion.

Mechanical layer (deterministic)

Starts at 100, deducts for:

• Cliché corpus (up to −25). Phrases documented in crisis-communication research as red flags: "thoughts and prayers" (−10), "we take this seriously" (−8), "no evidence of" without evidence of checking (−9), "technical anomaly" (−6), and a dozen more. Each hit is listed in your result.
• Specificity anchors (up to −20). Crisis responses fail when they omit the five things every stakeholder needs: a number, a timestamp, a next-update commitment, a named owner, and a scope statement. Each missing anchor deducts and is named.
• Passive voice (up to −8). Above 40% passive sentences deducts fully; above 25% deducts half. "Mistakes were made" hides actors; crisis communication requires naming them.
• Readability (up to −6). Flesch Reading Ease below 50 is difficult; below 30 very difficult. Panic writing is dense writing; dense writing fails audiences.
• Wall of text (−5). Over 250 words in a first response is burying the lede.

Editor layer (AI judgment)

Each persona applies its domain rubric — SCCT-informed strategy matching for crisis responses (does the response fit the crisis type?), stakeholder-first framing for reputation, fiduciary completeness for board updates, news-value for pitches. The rubric lives in the persona prompts; the judge's reasoning is the roast itself.

Why hybrid

Pure-AI scores drift — different models grade differently, same model wobbles run to run. The mechanical layer anchors the scale: it is reproducible, explainable, and independent of which model serves the judge. When we switch models, the anchor holds.

Grounding: the mechanical layer draws on SCCT (Coombs), the EU-funded Crisis Communication Scorecard (University of Jyväskylä), and standard readability research (Flesch). The cliché corpus is compiled from crisis-communications literature and post-crisis case studies.

Limitations, honestly

The mechanical layer cannot read context — a deliberately vague statement in an active legal case will score worse than it should. The editor layer cannot be fully consistent — that is the nature of judgment. The composite balances both, and shows its work.