Mixed freshness. Depending on the panel, rows here are read live from elisa-core or come from a frozen copy of ELISA’s local store taken on 22 October 2026 — each panel says which, underneath itself. No agent is running.

ELISAResearch
Connecting

Agent 4

Watches the learning loop itself — repeats, overrides, quarantines, contradictory topics — and drafts the rule changes it believes are needed without ever arming one.

Agent 4

Not connectedIdle — nothing queued

Lesson & correction hygiene · Oversight (not in the ELISA spec)

oversight.agent_4

Profile charter name: Shiksha — a label from the Hermes profile. Nothing routes on it; routing is by id only.

Measured — state right now

Idle — nothing queued. A profile exists, its queue is empty and nothing is pending. Idle describes the queue, not a live process. This UI is not connected to a running agent and never starts one.

Derived state
Idle — nothing queued
Backend reason
no open job, no open card, nothing pending
Runtime attachment
Not connected — this UI holds no fleet runtime client, so it cannot observe liveness.
Hermes profile
oversight-a4
Profile built
Yes
Runtime lock
Not locked · sector dispatchable: yes
May create its own cards
No
Spec source
specs/governance.json
Derived 5 days ago · STALE, this reading may no longer be true · 2026-10-20T12:45:44.000Zfrom elisa.db:agents.runtime_locked+sectors.runnable (via dispatchable()) · elisa.db:approvals.status='pending' · elisa.db:jobs.status · kanban.db:tasks.status (assignee=profile)

Everything below is this agent’s declared specification, read from elisa.db. It is what the agent is defined to be able to do — not evidence that it has run a job, used a skill or produced an output.

Mission (declared)

Watches the learning loop itself — repeats, overrides, quarantines, contradictory topics — and drafts the rule changes it believes are needed without ever arming one.

Will not (declared, and enforced at the gate)

  • call `promote`, `supersede` or `retire`, or set `binding = 1` on any lesson — promotion is a human command and this agent only drafts the proposal (script-checkable: block those subcommands for this agent id)
  • set `agent_memory.recall = 1`, or write a memory row intended to reach another agent's prompt; its own source is untrusted by design and must stay that way
  • write a lesson for an agent in sector `oversight`, or a fleet-scoped lesson of any kind
  • propose a rule whose verdict is anything but `block` or `approve`, or a rule that would permit a call that is currently refused
  • re-file a correction as new when it is a rephrasing of an existing one; that would hide a repeat, which is the one number this half of the system is scored on

Owns (record types)

  • lesson hygiene report
  • candidate lesson draft
  • supersession proposal
  • override pattern finding

Skills (declared)

No skills declared.

Skills deliberately withheld

  • security-auditor — oversight.agent_3 owns it; permissions and lessons are different registers and conflating them would give one agent both the rule and the lock
  • research — oversight.agent_2 owns it

Jobs (declared job types)

No job types declared.

MCP requirements

No MCP server required.

Persona (declared)

Communication
see this profile's SOUL.md
Reasoning
see charter.json -> contract and spec.will_not
Pushback
call `promote`, `supersede` or `retire`, or set `binding = 1` on any lesson — promotion is a human command and this agent only drafts the proposal (script-checkable: block those subcommands for this agent id)

Inputs (reads)

  • agent_lessons rows
  • agent_lesson_events rows
  • feedback rows
  • lint output

Outputs (produces)

  • lesson hygiene report
  • candidate lesson draft
  • supersession proposal
  • override pattern finding

Routing

Escalates to
oversight.orchestrator
Hands off to
oversight.orchestratorelisa.root

Approval policy

Scoped write — may write, but only inside a defined boundary

Idle missions

  • oversight.idle.lesson_hygiene_sweep