All conversations
Where should the human rating live, so it isn’t hidden?
Same container as today’s real screen in every column — the thread header row, the list table, or the message bubble. One delta each. Click to grade — every control here writes real, feeling state.
A compact, always-visible rating pill in the thread header, beside the flag icon and the ⋯ trigger.
- Seen the instant the thread opens — no hover, no menu.
- Reuses the exact header icon-row slot the flag/kebab already live in.
- Only helps once you’ve opened a conversation — the list itself still shows nothing.
A small rating chip on every list row. The open-conversation view underneath is untouched — the whole-conversation verdict still lives only in the ⋯ menu once you’re inside.
- Visible before you even open anything — browsing the list is enough.
- One glance across 15 rows shows what still needs a look.
- A rollup the list has never shown before — today's rows reflect message state, not a whole-thread verdict.
The exact hover-only thumbs, at ~35% opacity at rest instead of invisible — full opacity on hover or once graded. Nothing else changes.
- Zero new UI — same icons, same position, same slot.
- 35% opacity still reads as background chrome most people scan past.
- Doesn’t touch the whole-conversation rating at all — still buried in the ⋯ menu.
Grading placement rules
Three zones, ground-truthed against what prod already renders. Every new grading affordance gets placed by finding its zone below — not by taste in the moment.
- What it is: content the message itself carries — sender, timestamp, read state, reactions.
- Examples today: the avatar + timestamp row under every bubble; SMS tapback reaction chips (a parallel workstream, same zone — they render at the bubble's bottom edge, same as this).
- Visible to: anyone who can see the conversation, PM or otherwise.
- What it is: a human judgment being recorded about the message or thread.
- Examples today: the hover thumbs / persistent verdict pill on a Clara message; "Rate this conversation" in the ⋯ menu.
- Visible to: graders only — gated the same way the app already gates grading (never a property_manager-only viewer).
- What it is: a finding the judge model produced on its own, before any human looked.
- Examples today: the Flag icon next to the ⋯ trigger — a quiet status icon, never its own banner or row.
- Visible to: graders only, same gate as review chrome.
| Size | Where | Example |
|---|---|---|
| 28px | Header icon buttons | Flag, ⋯ kebab, close, external-open — the thread header row |
| 28px | Message corner controls | Hover thumbs, persistent verdict pill — top-right of a Clara bubble |
| 26px | Verdict pill (compact) | Icon + grader initials, wherever the pill appears — corner, header, or list row |
| 16px | Inline glyphs | Avatar, channel icon, chevron — sit inline with text, never a tap target on their own |
header / corner
verdict pill
inline glyph
| Tier | Meaning | Example |
|---|---|---|
| Always | On screen at rest, no interaction needed | Topic pills, channel icon, persistent verdict pill once graded, the flag icon when a finding exists |
| Hover / focus | Appears on mouseover or keyboard focus, gone at rest | Per-message thumbs before grading — the exact tier Fede flagged as too hidden |
| Menu | Behind one extra click, grouped with related actions | Copy link, copy conversation ID, view agent trace, whole-conversation rate/change/remove |
4-judge placement panel
Four independent judges, each scoring the three variants through one lens only. Scores are shown in full — never a blind verdict.