Antagonist re-score handoff — #1626¶
Audience: Antagonist / human bake-off scorer
Date: 2026-07-31 (agent drain complete)
Prior human mean: ~5.1 / 10 (pass ≥5.5)
Machine residual: demo_fleet residual=0 (necessary, not sufficient)
This pack hands work back to the antagonist. Do not treat machine residual or agent OCR as a bake-off pass.
How to re-score¶
- Stills root (local, gitignored):
-
Score dimensions (same as prior bake-offs): First 10s · Domain fidelity · Golden path · Feature depth · Demo data · Trust/polish.
-
Preferred hero stills (one primary per app):
| App | Primary hero still |
|---|---|
| simple_task | task_board_manager_desktop_light.png |
| project_tracker | project_board_manager_desktop_light.png |
| invoice_ops | approval_desk_approver_desktop_light.png + pay_desk_finance_desktop_light.png |
| fieldtest_hub | issue_triage_manager_desktop_light.png |
| support_tickets | ticket_queue_agent_desktop_light.png + manager_ops_manager_desktop_light.png |
| contact_manager | contacts_user_desktop_light.png |
| design_studio | brand_desk_designer_desktop_light.png + asset_catalog_designer_desktop_light.png |
| ops_dashboard | command_center_ops_engineer_desktop_light.png |
| hr_records | staff_directory_hr_admin_desktop_light.png + org_chart_hr_admin_desktop_light.png |
- Fleet claim rule: mean ≥5.5 → may claim “demo-safe fleet”; do not claim category leadership.
Agent-shipped residuals (R1–R6 + partial P1)¶
| ID | Residual | Agent evidence |
|---|---|---|
| R1 | Label glue | Mid-dot separators on queue meta; OCR: Assigned To: … - Created At: |
| R2 | Ops Active Alerts 500 | terminal_states on HTTP SM; still shows alert rows, no 500 OCR |
| R3 | New Add * CTA |
HR staff directory OCR: no New Add |
| R4 | Capture desk timeout | Playwright settle soft-fail; fresh pay_desk / my_invoices / active_alerts stills |
| R5 | Design swatches | Brand desk color/hex fields |
| R6 | Soft polish | delta_pct omit when |pct|>200; contact *.example emails; manager_ops recapture |
| P1a | HR org tree | Nested Department seeds + display: tree on Departments & Roles |
| P1b | Design catalog | asset_grid → display: grid (metadata cards) |
Tooling: scripts/agent_workspace_health.py, hardened
scripts/recapture_demo_fleet_1626.py (no PIPE, runtime.json port discovery).
Agent OCR spot-check (not a human score)¶
| Check | Result |
|---|---|
| 500 / Something went wrong | clean on heroes |
| New Add CTA | clean |
| @example.test story emails | clean on contacts |
| 800.0% metric deltas | clean after simple_task recapture |
| Meta glue | separators present on support queue |
| Machine floors | residual=0 |
Expected (agent estimate only): prior 5.1 + R1–R6 + ops/HR/design depth may approach ≥5.5 if humans agree on stills — antagonist decides.
Deferred (do not ding agent for these as P0)¶
- Invoice PDF document hub (project_tracker has the pattern; invoice already strong)
- Support threaded conversation UI (no chat primitive)
- Manager-chain recursive people tree (ManagerLink is edges, not nodes)
- True thumbnail media gallery
- Competitive claims vs Linear/Zendesk/Bill.com
Recapture recipe (if antagonist needs fresh stills)¶
.venv/bin/python scripts/agent_workspace_health.py --require-postgres
.venv/bin/python scripts/recapture_demo_fleet_1626.py --capture-timeout 900
python scripts/demo_fleet_bar.py --strict
Messaging constraints (unchanged)¶
| Safe until re-score | Forbidden |
|---|---|
| Fleet density / residual code drained | “Fleet bake-off passed” |
| Machine residual 0 | Closing #1626 on machine residual alone |
| Trio + recaptured heroes ready for review | “Competitive with category SaaS” |