Crew Dossier

Every Role OS role has a dossier: a character sheet that is also run-time configuration. It answers two questions at once — who is this role (an at-a-glance identity) and how should it behave when dispatched (real knobs, not flavor). Schema v0.2 (the specialists layer) keeps one rule above all: everything the sheet displays is real system state — the exact injected text, the real measurement basis, the real ledger.

The six aptitudes

Each role carries six aptitudes scored 0–5. Every axis maps to a concrete dispatch knob — if a stat doesn’t change behavior, it isn’t on the sheet.

Aptitude	0 → 5	Dispatch knob
Rigor	skim → reads every line, quotes evidence	quality-bar strictness / verifier depth
Pace	deliberate → fast, few turns	turn / budget profile
Range	convergent, spec-bound → divergent, generative	proposes alternatives vs executes the brief
Skepticism	trusting → refutes, demands proof	verifier strength, accept-vs-reject bias
Autonomy	escalates often → runs to completion	escalation / human-gate threshold
Candor	terse → explains, frames contrastively	output verbosity + contrastive framing

Each role has two arrays: its ideal (the canonical profile for the role’s archetype) and its build (the ideal nudged by its operating profile). The gallery radar draws both, so you can see how a role’s configuration tunes it off-canon. Priorities — what the role optimizes for, in order — are derived from the high-set aptitudes automatically, never authored.

Operating profile

An operating profile is a preset that sits on top of the role and both nudges the aptitude array and injects a real behavioral instruction at dispatch. Eight archetypes:

Skeptic · Builder · Diplomat · Contrarian · Perfectionist · Pragmatist · Investigator · Maverick

For example, a Skeptic raises skepticism and lowers pace, and injects an instruction like “default to revise/reject; an unrebutted challenge blocks acceptance; require explicit evidence before crediting a claim.” So “skeptical Judge” is a real configuration, not a label. The injection is functional posture only — task frame, priorities, evidence standards; profiles carry no affective states and no simulated social presence.

Operating Posture at dispatch (opt-in)

When a role has a dossier, buildRolePrompt appends an Operating Posture block to its prompt: the profile’s behavioral instruction, the derived priorities, and a one-line posture from the role’s aptitudes (e.g. “reads exhaustively and demands evidence; challenges hard before accepting; escalates early when uncertain”).

This is additive and non-breaking — a role with no dossier produces a byte-identical prompt to before. The runtime data lives in src/role-dossiers.json (compiled from the dossiers and shipped with the package), so the feature needs no extra setup. The gallery’s “As dispatched” panel shows the verbatim injected block per role — what you read is what the model gets.

Grades, reps, and techniques (measured, never minted)

• Grade carries a basis: assessed (panel-tuned editorial assessment — the honest state of the roster today) or certified (backed by the certification pipeline’s exams, confidence intervals, and replication). A grade’s band — the conservative skill interval — stays null until actually measured; the sheet says “unmeasured” rather than inventing a number.
• Reps count verified events (training runs, exams, field outcomes) — never calendar streaks, never activity volume.
• Techniques are earned distinctions derived automatically from properties of real results, each linked to the receipts that earned it. They cannot be authored or farmed.

The crew gallery

dossier/dossier.html is a self-contained gallery: painted portraits grouped by crew, each opening that role’s dossier — radar (build vs ideal), a calibration score, the operating profile with its behavioral delta and verbatim dispatch block, techniques, and the charter. It is generated from the role data by dossier/build-gallery.mjs, so it refreshes whenever the roster, portraits, or stats change.

How the profiles were calibrated

The aptitude profiles were tuned like an instrument, not guessed:

1. A model of 10 role archetypes — each a distinct six-axis ideal — plus an 8-profile delta library. Build = clamp(ideal + profile delta).
2. A panel of three diverse models read each role’s function and charter and proposed a tuned profile; a per-axis median rejected single-model bias.
3. A different-family external-verifier pass broke the residual collisions knob-faithfully — yielding 64 distinct, knob-faithful fingerprints (the 61 roles plus 3 specialty auditors).

Back to the handbook · Landing page