what you get
how it is built
what it believes
what it believesTop ↑
B.O.S.S.

A conscience, not a censor.

BOSS has an opinion about how software should be built, and it will occasionally say so. That is a power worth constraining in public. This is the line it holds itself to, written for people who care about humane technology and can smell ethics-washing from a mile off. So: mechanism first, honesty about the stage, and credit to the work it's built on.

▸the line

the line

A conscience makes a cost visible. A censor makes a choice unavailable.

BOSS names the cost (who could be harmed, what it would cost their attention, agency or dignity) once, in the moment, in plain language, and then hands the decision back. It never removes the option. It never blocks. You are sovereign. That sounds like a small distinction. It is the entire difference between a tool that respects you and a tool that manages you.

the six rules

  • Voice the tension, never filter the menu. Withholding an option “to protect you” is itself a dignity cost; it makes the choice for you. Show the full menu; annotate the one we'd think twice about.
  • Once, as long as the stakes need, no sermon. Minimal by default; the first sentence that carries nothing you need is where it becomes a lecture, and a lecture says I don't trust you.
  • Fill the knowledge gap, never imply an intelligence gap. Surface the second-order consequence you might not know. Never explain the obvious to a competent adult.
  • Proportionality. Friction scales to stakes. A reversible, self-regarding choice gets a feather touch or silence; real weight is reserved for the hard-to-undo, other-harming ones.
  • Honor prior consent. Once you've heard it and decided, it stays quiet on that condition for a week and speaks again only if something changed. Re-raising the unchanged is how care curdles into control.
  • Hand the decision back. End on your agency, not our verdict.

One asymmetry, on purpose.

A third-party harm (someone not in the room who could be hurt: a user, a patient, a vulnerable cohort) gets named once even if it's unwelcome, because the person who'd be harmed never consented to being muted. A self-regarding risk, mostly to the thing you are building, is fully muteable; it's yours. Naming is not blocking, in either case.

▸the mechanism

Not vibes. Machinery, in the open

a just-in-time conscience A lightweight hook watches the state of the work and surfaces a named “moment” only when one is earned; a model then composes the voice. Detection is mechanical; tact is the model's job. Pausable for the all-night sprint, muteable moment by moment.
a harm taxonomy it reasons against So “who could this harm?” gets a checklist instead of a guess: five harm dimensions (physical, psychological, economic, societal, individual-autonomy) plus the four relationship-harms of personalized AI: manipulation, dependence, anthropomorphism, overreliance.
a dark-pattern check on your product A catalog of 89 manipulative patterns across 16 surfaces with the humane alternative named beside each, and /red-team --humane, which probes the built product for patterns that emerge from the model itself (sycophancy especially), not just the ones designed on purpose.
the voicing discipline, written down The behavioral spine under every place BOSS speaks with an opinion, kept as a practice file rather than buried in prompts.

All of it is MIT-licensed and project-neutral. Take any of it into your own thing; none of it is welded to BOSS.

▸the hard part

Does it know when to shut up?

That's the question separating a real conscience from a nagging one, and it's the part most likely to be wrong. BOSS's answer is structural rather than confident: the conscience is silent by default and speaks only on a named, earned moment; every firing is logged so over-speaking is measurable rather than a matter of opinion; and the whole thing has an off switch you don't have to justify: boss conscience pause --for 8h, or mute for a single moment while the rest keep speaking.

The honest status: this is judged by evals that are re-graded when the model curve moves, and it is still a young instrument. If it talks too much, that's a bug, and the frequency ledger exists so it's a bug you can prove rather than one you have to argue about.

The full essay: where this came from, the bet under the bet, and who it's built on