Audience-as-grader segment: the hands-on reopen created it and nothing tracks it #367

New issue

Open

opened 2026-08-27 03:25:25 +00:00 by coilyco-ops · 0 comments

coilyco-ops commented

2026-08-27 03:25:25 +00:00

Member

Filed by Saiya (tpm seat) after the 2026-08-25 sequencing pass. This is a gap rather than a deferral: no issue in any repository covers it.

Outcome

The PyLadies Remote session needs a runnable segment where the audience occupies the grader seat. Build the smallest thing that lets fifteen to fifty people in a Google Meet score role-conformance responses and see that they disagreed with each other.

Why this exists now, and why it did not before

The format reopened to 1.5 hour hands-on on 2026-08-22, recorded in coilysiren/lore#22. That issue names the upside the earlier cost-only analysis missed:

lore-method-agent-eval establishes the generator-subject-grader triple where no party occupies two seats, and Kai is the human grader. In a hands-on session the audience can fill the grader seat.

The talk format did not need this. Kai graded, and the room watched. The hands-on format makes the room the instrument, and nothing has been built for it.

The outline on coilysiren/inbox#394 already commits to it. Segment 3, nine minutes:

Show a composed persona's declared commitments, then show a response. Room votes pass or fail. Three or four rounds, escalating ambiguity deliberately. The room splits somewhere around round three.

So the segment is designed. It is the only segment in that outline with no artifact behind it.

What to build

Smallest useful version, deliberately not a product:

A set of three or four case pairs, each one a composed persona's declared commitments plus a candidate response, ordered by increasing ambiguity. Round one should be unanimous. Round three should split the room. That ordering is the whole teaching mechanism and it is the part that needs iteration rather than code.
A way to put one case in front of a room so the commitments and the response are legible side by side on a shared screen. A rendered page is enough. This does not need to be a CLI, a service, or anything that survives the session.
A way to collect and show the split. Meet has polls. Hands on camera also works. Prefer whatever needs no software before preferring software.

Explicitly out of scope: automated grading, a persistent vote store, anything that writes to the eval record, and anything that has to run on an attendee's machine. The audience grades and the result is a talking point, not evidence.

Why it is independent of the inversion program

agent-compose#329 and its children own the roster-to-bundle inversion, and every eval-shaped slice under it is currently blocked: #340 behind #338, #332 behind #331, #337 behind #330 and #333.

This is behind none of them. It reads composed personas that already exist and produces presentation material. It should not be sequenced into #329 and should not wait for it.

Freeze constraint carries over

inbox#394 establishes that the January abstract commits to findings rather than inventory. This segment is a live activity rather than abstract copy, so the constraint binds differently, but one part does carry: pick persona commitments whose specifics will still be true in September. The roster is mid-reflow (agent-compose#317, agentic-os#1176), so choose the commitment class rather than a named seat.

One live footgun already recorded on inbox#394 and worth repeating here: do not use dislike of a named framework as the example commitment. It lands badly in a room containing people who use it. Prefer a preference with no constituency in the audience.

Acceptance criteria

Three or four case pairs exist, ordered by ambiguity, with the intended split point named for each.
Round one is unanimous when tried on any two people, and the designated split case actually splits them.
Each case is presentable on a shared screen with commitments and response legible together.
A vote-collection method is chosen and tried once end to end.
No case names a role, a seat count, or a model.
Every asset is public-safe, since the session is live-streamed to the PyLadies YouTube channel and the recording is permanent.

Ownership

Building this is the Developer Platform Engineer's, per boundary-build-foundational-software. Case authoring and the words on screen sit closer to Content Creator. This issue records the gap and the scope only, and no agent drafts, publishes, or presents anything without Kai's explicit authorization.

coilysiren/inbox#394 - the CFP issue carrying the full outline, including segment 3
coilysiren/inbox#338 - the PyLadies session issue, topic and format
coilysiren/lore#22 - the format reopen that created this gap, and the stale-skill follow-up

Filed by Saiya (tpm seat) after the 2026-08-25 sequencing pass. **This is a gap rather than a deferral: no issue in any repository covers it.** ## Outcome The PyLadies Remote session needs a runnable segment where the audience occupies the grader seat. Build the smallest thing that lets fifteen to fifty people in a Google Meet score role-conformance responses and see that they disagreed with each other. ## Why this exists now, and why it did not before The format reopened to **1.5 hour hands-on** on 2026-08-22, recorded in `coilysiren/lore#22`. That issue names the upside the earlier cost-only analysis missed: > `lore-method-agent-eval` establishes the generator-subject-grader triple where no party occupies two seats, and Kai is the human grader. **In a hands-on session the audience can fill the grader seat.** The talk format did not need this. Kai graded, and the room watched. The hands-on format makes the room the instrument, and **nothing has been built for it**. The outline on `coilysiren/inbox#394` already commits to it. Segment 3, nine minutes: > Show a composed persona's declared commitments, then show a response. Room votes pass or fail. Three or four rounds, escalating ambiguity deliberately. The room splits somewhere around round three. So the segment is designed. It is the only segment in that outline with no artifact behind it. ## What to build Smallest useful version, deliberately not a product: * A set of **three or four case pairs**, each one a composed persona's declared commitments plus a candidate response, **ordered by increasing ambiguity**. Round one should be unanimous. Round three should split the room. That ordering is the whole teaching mechanism and it is the part that needs iteration rather than code. * A way to **put one case in front of a room** so the commitments and the response are legible side by side on a shared screen. A rendered page is enough. This does not need to be a CLI, a service, or anything that survives the session. * A way to **collect and show the split**. Meet has polls. Hands on camera also works. Prefer whatever needs no software before preferring software. **Explicitly out of scope:** automated grading, a persistent vote store, anything that writes to the eval record, and anything that has to run on an attendee's machine. The audience grades and the result is a talking point, not evidence. ## Why it is independent of the inversion program `agent-compose#329` and its children own the roster-to-bundle inversion, and every eval-shaped slice under it is currently blocked: #340 behind #338, #332 behind #331, #337 behind #330 and #333. **This is behind none of them.** It reads composed personas that already exist and produces presentation material. It should not be sequenced into #329 and should not wait for it. ## Freeze constraint carries over `inbox#394` establishes that the January abstract commits to findings rather than inventory. This segment is a live activity rather than abstract copy, so the constraint binds differently, but one part does carry: **pick persona commitments whose specifics will still be true in September.** The roster is mid-reflow (`agent-compose#317`, `agentic-os#1176`), so choose the commitment class rather than a named seat. One live footgun already recorded on `inbox#394` and worth repeating here: **do not use dislike of a named framework as the example commitment.** It lands badly in a room containing people who use it. Prefer a preference with no constituency in the audience. ## Acceptance criteria - [ ] Three or four case pairs exist, ordered by ambiguity, with the intended split point named for each. - [ ] Round one is unanimous when tried on any two people, and the designated split case actually splits them. - [ ] Each case is presentable on a shared screen with commitments and response legible together. - [ ] A vote-collection method is chosen and tried once end to end. - [ ] No case names a role, a seat count, or a model. - [ ] Every asset is public-safe, since the session is live-streamed to the PyLadies YouTube channel and the recording is permanent. ## Ownership Building this is the Developer Platform Engineer's, per `boundary-build-foundational-software`. Case authoring and the words on screen sit closer to Content Creator. This issue records the gap and the scope only, and no agent drafts, publishes, or presents anything without Kai's explicit authorization. ## Related * `coilysiren/inbox#394` - the CFP issue carrying the full outline, including segment 3 * `coilysiren/inbox#338` - the PyLadies session issue, topic and format * `coilysiren/lore#22` - the format reopen that created this gap, and the stale-skill follow-up

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

coilyco-flight-deck/agent-compose#367

No description provided.

Rows
Columns