How the score is built.

Rubric only ever reads your turns in a session. The AI's replies never leave your browser. The scoring model sees the user's prompts plus three primitives per assistant reply (its length, whether it contained a code block, and whether it contained an error) but never the reply's text.

This is a deliberate constraint, not just a privacy claim. It means Rubric is grading how you prompt, not how well a model happened to answer you. A mediocre prompt that produced a great answer scores lower than a strong prompt that got an unlucky reply. The asymmetry is the point.

Even though the scorer only reads your prompts (not the AI's replies), your prompts themselves can contain secrets, like a customer email pasted into a debugging session, an API key in an error message, an IP in a stack trace. Before the LLM sees a turn, it passes through a redaction step that replaces high-risk patterns with opaque placeholders.

• Email addresses → [EMAIL]
• Phone numbers → [PHONE]
• API keys (OpenAI sk-…, Stripe pk_live_…, Google AIza…, AWS AKIA…, GitHub ghp_…) → [KEY]
• IPv4 addresses → [IP]
• Credit-card-shaped digit groups → [CARD]

We intentionally don't redact names, addresses, company names, or identifiers like variable names and file paths. Detecting those reliably is full of false positives, and the cost of accidentally redacting half a prompt is worse than the cost of leaving an arguably-public name in. The redaction layer is high-precision and low-recall by design.

Your original, un-redacted prompts are still stored in the database so you can read your own session history. Only what we send to the LLM is redacted. The implementation and full pattern list live in apps/backend/app/analysis/redact.py and are covered by 12 tests.

Every session is scored 0–100 on each of six axes, then weight-averaged.

For every dimension, the scoring model must cite at least one specific user-turn index as evidence. If you see a low specificity score, you can look at exactly which of your turns drove that score. This is also what powers the Premium before/after rewrites feature: the cited weak turns are what we offer to rewrite.

The weighted raw score is then run through a strictness transform before being shown to you. The transform pulls scores toward a median of ~50. Concretely:

adjusted = (raw ** 1.15) / (100 ** 0.15)
displayed = clamp(round(adjusted), 0, 100)

Why? Because a 0–100 number that feels uncalibrated is useless. A median user scoring 85 produces no signal, while a median user scoring 50 produces information. The strictness is also what makes the share card honest. A 73 from Rubric means something. A 73 on a soft scale means nothing.

The curve constants are tuned against a calibration set of hand-graded sessions and updated periodically. When they change, we publish a dated changelog at the bottom of this page so old scores remain interpretable against the version of the rubric that produced them.

The archetype is a short label (eight available, e.g. The Sculptor, The Architect, The Wanderer) that summarizes the session's personality. It's chosen in two steps:

1. Cheap heuristics on session shape (turn count, avg turn length, structural markers) produce a candidate set of 3–5 plausible archetypes.
2. The LLM picks one from the candidate set with reasoning, citing turns. It can defend choosing The Architect over The Sculptor by pointing at specific structural markers in your prompts.

If your session is too large to fit in a single LLM call, Rubric keeps the first 5 turns, the last 10, and 10 turns evenly spaced from the middle. The report is flagged was_truncated when this happens. Truncation rarely changes archetype but can compress the curve toward the middle, since the model sees fewer extreme examples.

A 39 stings. It's supposed to. But if you think the score is wrong, not just harsh, the rubric is meant to be argued with. Look at the evidence turns cited for the dimensions that hurt you. If the citation doesn't hold up, the score is broken. We want to hear about it.

Email hello@rubric.chat with the session ID and which dimension you think is misjudged. Most disagreements we see are either calibration issues (the curve is wrong for a session type; coding agents are the common one) or evidence misattribution (the model cited the wrong turn). Both are fixable. None of them mean we soften the score.

• v0.1.0 · 2026-05 · Initial public methodology. Strictness curve exponent 1.15. Dimension weights as above. Archetype heuristics expanded to always include a baseline spread.