Co-Scientist L4 — control-plane ownership: who may write, rank, route

This is a recorded boundary, not a roadmap item.

Anchor

• Paper: The Context Memory component (paper L1349–1354) "stores and retrieves states of the agents and the system during the course of the computation"; the Meta-review agent (paper L1498–1503) appends feedback into other agents' prompts; the Supervisor agent routes work across the agent coalition.
• ARS: the Material Passport, with state_tracker_agent as the writer of pipeline state; per-stage checkpoint records; pipeline_orchestrator_agent-led routing under user confirmation. POSITIONING.md — "the human decides at every gate."

Problem

Co-Scientist and ARS both have memory, both propagate information across agents, and both route work between stages. The mechanisms look similar from the outside, but they answer different questions:

• Co-Scientist's memory is a shared mutable context that agents read from and write to. The system optimizes for collective learning across the agent coalition.
• ARS's Material Passport is an authority ledger. Single-writer discipline (state_tracker_agent), append-only semantics, and audit-grade restart safety are load-bearing properties.

Importing a paper-derived mechanism without making the control-plane ownership explicit silently shifts ARS from authority-ledger to shared-mutable-context. The shift is invisible at the surface — the API looks the same — but the integrity guarantees change.

The three questions

Every paper-derived mechanism proposal must explicitly answer these. This document is their canonical reference.

1. Who may WRITE to state?

ARS state writers:

• Pipeline state (Material Passport): state_tracker_agent is the sole writer of pipeline_state, current_stage, blocking flags, and materials. One narrow, controlled exception exists and proves the rule: collaboration_depth_agent may append only to its own collaboration_depth_history[] through a dedicated append-only interface, and never touches pipeline state. The discipline is "one writer per state region", not "literally one agent writes the whole passport".
• Checkpoint records: pipeline_orchestrator_agent after user confirmation.
• INSIGHT Collection: produced in the RESEARCH stage (deep-research, socratic mode — socratic_mentor_agent with the research-question and devil's-advocate agents), under the Stage-1 user-confirmed scope. It is a research-stage deliverable, not a write performed by synthesis_agent.

A new mechanism must either route its writes through these existing writers, or declare a new state region with its own single writer and the same discipline (audit log, append-only semantics, restart-safety guarantees). It may not add a second writer to an existing region.

2. Who may RANK candidates?

Permitted shapes:

• AI enumerates without ranking (green line — see L1).
• AI surfaces scores to the user as advisory information; the user picks (yellow line, with L1's three conditions).

Rejected shape:

• AI ranks and the system selects top-K without showing the user the full set (red line — hidden culling).

3. Who may ROUTE to the next stage?

ARS routing writers:

• pipeline_orchestrator_agent computes the next stage from the pipeline configuration and the Material Passport state.
• The user confirms each transition at a checkpoint.
• No agent unilaterally decides what stage runs next.

A new mechanism must either route through the orchestrator, or document why a new routing primitive does not violate the user-confirmation invariant.

Why this distinction matters

The Co-Scientist paper presents shared mutable context as a strength. Within its paradigm that is correct: hypothesis generation benefits from continuous information flow across agents.

ARS's paradigm is different. The Material Passport's value comes precisely from its single-writer discipline. The same property the paper considers limiting (state cannot be freely mutated by every agent) is the property that makes ARS auditable.

Without the three questions above, a paper-derived mechanism can preserve the surface API while silently undermining the discipline. The questions force the design to be explicit about control-plane ownership before implementation begins.

Application — when this is invoked

Any mechanism that writes state, ranks candidates, or routes between stages must answer the three control-plane questions above. This document is their canonical reference: a design that cannot answer them has not established its control-plane ownership, and the boundary is unmet until it can.

See L1 (who may rank) and L2 (who may mutate a downstream prompt).

Non-goals

• Not a prohibition on cross-agent information flow. Information flow is fine; the question is which agent has authority to mutate state, which to rank, which to route.
• Not a claim that Co-Scientist's shared-mutable-context design is wrong. It is the right design for hypothesis generation at the volume the paper operates.

Verification note

Verified against the codebase 2026-06-02. Two corrections relative to the originating issue:

1. Naming. The writer agent is state_tracker_agent (the bare state_tracker is its schema/role field name); the orchestrator is pipeline_orchestrator_agent. The single-writer discipline itself is verified (state_tracker_agent is sole writer of Material Passport state), as are orchestrator-written checkpoint records and orchestrator-led routing.
2. INSIGHT Collection attribution. The issue listed it as a synthesis_agent write "after Stage 1". Verified incorrect — INSIGHT Collection is a deep-research RESEARCH-stage artifact, and synthesis_agent is a Phase-3 integration agent that does not produce it. The "who may WRITE" list above reflects the corrected attribution.

The substance of the boundary (single-writer authority ledger, orchestrator-led routing under user confirmation) stands; only these labels were tightened.