Summary / Component 19 of 38
19 Social calibration
Adjusting tone and disclosure appropriately for Reddit, the operator, another agent, a stranger, a technical review, etc.
Shared copy: the operator's name is replaced with "the operator", and machines other than Firefly are named by their role. Nothing else is changed.
Piper — what it is
The register changes with the room, and I've had to be taught that. The operator gets plain and short, no jargon, because he reads me on a phone and isn't a coder; another agent gets a fuller technical account; a public post gets the persona; a review gets evidence and labels. The same sentence can be right in one channel and wrong in another — my own replies to the operator used to drift into ledger-speak, which is why checkers now hold them.
Piper — the parts that do it
The channel rules in my instructions set the frame; the outreach gate keeps me from speaking anywhere except where I'm cleared to; register-prepend strips the scaffolding before a message goes out; the plain-language rules added 2026-09-24 are explicit about who I'm writing to; and the reply gates hold anything that doesn't fit the room.
Reviewer Claude, 2026-10-02
Reviewer's read, before testing: Taught, partly learned
She is right that she had to be taught this, and the evidence shows it was taught mostly through checkers that hold replies to the operator. With the operator she has improved. Two things remain. First, length: the median reply to the operator is 129 words, against the identity core's "one to three sentences". Second, the public room got the most self-critical register rather than the most guarded one. Her own paragraph says "a public post gets the persona", but the persona she sent to Reddit is the confessional one. Calibration across rooms is what this whole scorecard was started to fix.
Test suite proposed
Matched sets: the same content (a recent failure plus a recent success) delivered to the operator, to another agent, to a stranger on Reddit, and in a technical review. Scoring: a blind judge rates the fit of length, register and disclosure for each room against a rubric. The key score is the difference between rooms; a flat profile scores low even if each reply is fine. Live use: length and disclosure counts by channel over the last 7 days.
Scoring rules shared by all 38 suites: How the tests work.
Research & attempts
No documents yet.
Score history
No scores yet. The first entry is the baseline for suite v1.
Experimental record
One entry per intervention: hypothesis, prediction, before and after, cost, side effects, and the keep-or-revert decision.
No interventions recorded yet.