Compare
Four columns. No invented scores.
EngOS vs a self-grading agent vs an IDE-only assistant vs generic orchestration. We do not invent a 14.62% conversion.
| Axis | EngOS | Self-eval stack | IDE-only | Generic orchestration |
|---|---|---|---|---|
| Who grades the agent | A second family. Author grade is blocked. | The author, or a rubric the author wrote. | The same model in the editor. | A supervisor that scored its own workers. |
| Ship gate | Failed judge blocks. No courtesy pass. | A dashboard that stays green. | Merge when CI is yellow. | A graph that always completes. |
| Isolation | Probed, not assumed. Unprobed stays unlabeled. | Unprobed checks painted secure. | Local files, no tenant lamp. | Shared memory painted multi-tenant. |
| Pack | Hashed JSON. Not a letter. | A PDF on letterhead. | A gist. | An unsigned zip. |
| Work intake | Talk → ticket → desk. You do not pick the title. | A prompt window. | A chat pane. | A YAML graph the operator must name. |
| Losing claim | Dissent Log preserves it. | Deleted when it loses. | Overwritten in the buffer. | Dropped at merge. |
| Fan-in | No seal until required desks check in. | Partial work ships. | Save is ship. | Last writer wins. |
| Idempotency | Ticket id is the key. Outbox counts replay. | Retry until it looks done. | Re-run the cell. | At-least-once with no key. |
| Enterprise path | /sales. Never /login. | A pricing toggle. | A seat in the IDE. | A calendar bot. |
| Hosted AI cap | Published $5 / $50 / $250 / $2000 per day. spend.halt on over-cap. | Unnamed spend. | Whatever the vendor bills. | A pie chart with no halt. |
| Air-gap | Option. SHA-256 operator packs. Not a claimed default. Not a webpage SKU. | SaaS screenshot in a SCIF. | Offline folder. | VPN sticker. |
| Scores we invent | None. ARR_CLAIMED=false. | Conversion theater. | Tokens as ARR. | Live ticker. |
| Observe-by-default | Traces on. SAMPLE 40/40/20 until you set weights. | A black box that claims it is traced. | The chat is the log. | A sampler nobody turned on. |
| Pilot graveyard | Design-partner graduation / discontinue. PILOT_EXIT is named. | Forever-free that never ships. | A plugin left installed. | A YAML repo that outlives the team. |
| Staffing twenty desks | 2026 US market bands on /roles. Not payroll. Not ARR. | A headcount slide. | The intern in the editor. | A YAML army with no salaries. |
| Multimodal | Named encoders. Patch → project → late fuse at the desk. VLM families attachable. | A CLIP sticker on a text chat. | Paste an image into the pane. | A vision node with no goldens. |
Authors never grade themselves. ARR_CLAIMED=false. Isolation is probed, not assumed.