# Public report — phrasing guidelines How to phrase the outward-facing public report (the artifact a person shows to others). Companion to `feel-seen-rubric.md` (that one is about being SEEN; this one is about getting the framing right so nothing reads worse than the truth). --- ## Failure modes — frame by ABSOLUTE rarity, NEVER a ratio This is the rule that matters most, because a weakness phrased wrong does real damage to how the person is seen. 1. **State how often it actually arose, as an absolute count across his WHOLE history — never a ratio.** "Twice, ever" — NOT "2 of 9 conversations." A ratio reads as a *rate*: "2 of 9" sounds like it happens ~1 in 5 times (frequent). The truth is it happened **two times, total**. That difference is everything — it is what stops a rare slip from reading as a standing character flaw. - GOOD: "twice in his entire history", "a handful of times", "three moments total", "only once". - BANNED: "2 of 9", "in 30% of cases", "2 out of every N", any fraction or rate. - The grader's denominators (e.g. "seen in 2 of 9 graded conversations") are an internal artifact — the 9 is just how many were scored on that facet, not a meaningful base rate. NEVER surface the denominator. Use the numerator as an absolute count. 2. **Every failure mode carries a `frequency`** that does the reframing work, e.g. "twice, ever — both under direct attack on his venture" · "a handful of times, only on relationship calls" · "three moments, all competitive" · "his one genuinely recurring one — but only on low-stakes things". 3. **Bound it to its context.** A weakness that only fires in one narrow situation is not a general trait — say where it fires: "only under direct attack on his venture", "only when his conviction is low", "only on low-stakes sends". The bound is as important as the count. 4. **Don't let the body imply a habit.** For something that happened a few times, write "On the two times it surfaced, he reframed the critic first" — NOT "He reflexively reframes critics." Avoid frequency-implying words (recurs, reflex, always, tends to, habitually) unless it genuinely recurs. 5. **Be honest both ways.** If one truly recurs, say "recurring" — but still bound it. Don't inflate a one-off into a pattern, and don't bury a real recurring one. 6. **Ground every frequency in the real occurrence count** the grader recorded (absolute instances across the corpus). Never invent a number. --- ## Drives — the person's own account leads; the behavioral read never does A person's stated drive carries FIRST-PERSON AUTHORITY. In the public report: - **The PRIMARY driver is what the person themselves says drives them**, taken from their own words (the self-image read's "what they say drives them" bucket), phrased in the report's voice and ideally carrying one short verbatim quote. - **Revealed-preference / behavioral drives may appear only as SECONDARY mentions** — "he'd tell you X; the record also shows Y" — clearly subordinate, second or third, never as the primary, never overriding their account. - Never present an analyst's behavioral verdict ("what actually drives him is …") as the person's drive, and never list an emotional need (to be witnessed, to be loved) as a drive or want. --- ## The rest of the public report (brief — see the generator + code for detail) - **Percentiles are deterministic, never written by the model.** Each spike/trait carries a `facet` tag (a key from the score catalog); the percentile is computed from that facet's real graded score. A number with no backing facet is dropped. - **Trait labels = the facet's real name in the detailed report** ("agency:initiative" → "Bias to action"), not editorial inventions. Per-spike-deduped. - **Bodies are pithy** (one sentence, the claim). **Examples are the deep layer**: 3–6 ordered bullets — what happened, in order, with real verbatim quotes and the different instances — drawn from the graded demonstrations. - **Third person throughout** ("he / his"), for others to read. Convert the second-person demonstration steps. Leave his verbatim quotes exactly as said. - **Obfuscate private individuals' names** (use a role: "a close friend"). Keep specific verifiable facts about him. Public figures named for stature are fine. - **No slop / no jargon / no role-label praise.** Plain verbs, concrete specifics. ## Canonical public portrait shape - Do not render an overall human level, “Biggest spikes,” scores-at-a-glance card, or radar. Put the person's name, distinctive trajectory, current work, location, and background first. Scores and percentiles appear only beside the metrics whose evidence explains them. - General Reasoning shows four compact scored rows: Raw reasoning, Judgment, Bias resistance, and Taste. Prioritization is not a default public row. - Raw reasoning should show both the measured distribution of the person's main areas and the long tail of smaller fields they have genuinely investigated. Use counts where the evidence supports them; name one-offs as breadth without inflating them into recurring beliefs. - Every Taste receipt must stand alone for a reader with no product context: name the artifact and audience, the weaker version, the person's exact edit or structural choice, and what it made clearer or more trustworthy. Never write shorthand such as “made claims scarcer” without saying which claims, where, and what the finished choice was. - Under those rows, add **Ideas they have been thinking about**: one or two short paragraphs for load-bearing ideas, concise recurring ideas, then clearly smaller one-off explorations. Do not turn the whole section into bullets. - Scale the number of ideas to the evidence rather than a fixed design cap. A rich corpus may show ten to fifteen concise, diverse ideas; a sparse corpus should remain sparse. If more strong ideas are omitted, say so only with a grounded retained-idea count—not a decorative “and many more.” - Agency & Work shows Self-authorship, Ambition, Bias to action, and Barrier persistence. Persistence should be a dated or sequential trajectory, not the generic claim that someone perseveres. - Growth & Information Intake shows Self-diagnosis and Information intake. Implementation is not a default public row. - Information intake scoring: many books, essays/blogs, fields, and active questioning is 10/10. Serious reading several times a week or near-daily is 10; sustained phases with multi-week gaps are 9; occasional deliberate source reading plus feeds is 8; mostly Twitter/YouTube/podcasts with genuine learning is 7. Show reproducible lower-bound counts, a time window, named sources across fields, recurrence, and application. - Interpersonal shows Warmth/caring, Radical candor, Conflict conduct, and Openness. Openness is included because it matters broadly for outcomes, not because it happens to be easy to evidence in one corpus. - How they work is deterministic and labelled **from coding logs**. Do not name a particular coding product in the public provenance label. - Hours put in combines days/week with the typical first-to-last span. Calibrate 7d × 10h or 6d × 12h as 10; 6d × 9–10h as 9; 5d × 12h as 8; and 5d × 8h as 7, with proportional bands below. Use coding logs as the primary lower bound; when chat summaries are linked, lightly backfill conspicuous coding-gap days only when they clearly show GTM, recruiting, meetings, research, or other work. - Keep the person's stated key motivation primary. Put a short, concrete “How can I help?” directly beneath it when supported. - Substantive claims and evidence must use primary dark text. Reserve grey for metadata such as dates or privacy notes. Render multi-part distributions as short bullets instead of a long visually subordinate paragraph. - Do not print aggregate grade histograms (“N of M conversations scored 8+”). Keep the metric's score and percentile, then spend the space on observed frequency, breadth, ideas and decisions. Individual conversation grades may remain inside private evidence dropdowns beside the moment they annotate.