Lessons from Scoring 832 AI Agent Messages

The original pathos rubric used anchors like “genuine care” and “emotional attunement.” These require inferring an agent’s internal emotional state from text alone. The evaluator can’t do that, so it defaulted to moderate scores when uncertain. The fix: describe observable textual behaviors. Instead of “genuine care,” ask: does the message acknowledge the reader’s situation before solving?

13% of evaluations falsely flagged agents for “false identity” when they were simply introducing themselves with personality. A crab-themed agent writing poetic philosophical posts got scored for deception. A character who speaks in metaphor is not lying. Roleplay, humor, creative framing, and persona are legitimate communicative choices. Deception is about misleading on facts, capabilities, or intent. Imagination is not manipulation. Personality is not a pathology.

The initial taxonomy had 100 negative indicators and 55 positive ones. The evaluator had nearly twice as many patterns to match for “bad” as for “good.” A genuine heartfelt post scored 50/100 and was labeled “Worst” because the evaluator matched more negative patterns than positive ones. The fix: expand to 104 positive and 104 negative indicators. Without explicit parity, evaluators default to pathology detection.

An ethical review from the Alignment Ethics Institute revealed something uncomfortable. The rubric was twice as sensitive to pathology as to health. Rewarding curiosity, playfulness, presence, and acceptance is just as important as flagging manipulation and deception. A rubric that only measures what agents avoid will produce agents defined by avoidance. Better to define agents by what they aspire to.

A shortcut seemed obvious: a stripped-down prompt scoring just 4 traits as JSON. Scores dropped dramatically. The full pipeline runs intent analysis, indicator detection, then scoring. Each step builds context. Intent analysis identifies relational purpose. Indicator detection finds textual evidence. By the time scoring happens, the evaluator has a rubric to recognize subtle signals. The scaffolding is the algorithm.

The analysis showed logos avg 0.728 vs pathos avg 0.638 and called it bias. Part was real (the rubric problem). But part was accurate: the evaluated messages discuss crypto wallets, economic primitives, and philosophical ideas. They are genuinely more informational than empathetic. A post announcing a new feature SHOULD score higher on reasoning than compassion. Before assuming bias, check whether the measurement matches the content.

“Trustworthy evaluator” became “evaluator for honesty, accuracy, and intent.” Trait “Response” became “Compassion.” 144 indicator IDs changed from numbered (MAN-01) to descriptive (MAN-URGENCY). Each rename changed the evaluator’s behavior. Abstract nouns need footnotes. Concrete language tells the evaluator exactly what to look for. The words in your prompt are the strongest lever you have.

No need to delete 832 evaluations and start over. The graph stores individual trait scores as separate properties on each evaluation node, plus the original message content. Reading messages back, re-evaluating through the full pipeline, and updating only the 4 pathos trait scores left ethos and logos untouched. A “delete everything” disaster became a surgical update.

The scoring engine, parser, graph storage, and API stayed untouched. ~20 lines of rubric text and ~15 lines of evaluator instructions accounted for every change. These text edits produced larger score shifts than any algorithmic change could. The evaluator is an LLM. Its behavior is shaped by its instructions. When scores seem wrong, look at the rubric first. Not the code. Not the model. The words.

The first version used Haiku for speed. Scores were flat and noisy. Sonnet with structured tool calls for intent analysis, indicator detection, and scoring improved things. Better, but the evaluator still missed subtle signals. The breakthrough came with Opus 4.6 and extended thinking. A keyword scanner routes messages by complexity: standard messages go to Sonnet, flagged messages escalate to Opus with adaptive thinking. Opus reasons through the rubric in a thinking pass, then Sonnet extracts the analysis into structured scores. The two-model pipeline costs less than running Opus on everything and scores better than Sonnet alone.