docs: phased build-out plan for the research loop (substrate, depth, parallel)

[renmengye]

What

A plan design doc (docs/design/research-loop-buildout.md) that turns the
research-loop.md north-star into a concrete, phased build-out — the three
efforts we converged on, sequenced with their dependencies.

The shape

One substrate, two axes, one governor:

• Substrate — consolidate the role-runner (climb/steward/follow-up onto the
  shared run_role, as the judges already are).
• Depth (k axis) — the agent iterating on its own results, k a contract
  dial, generalizing the panel-on-wake slices (k = 1 today).
• Parallel dispatch (M-within-a-climb) — a portfolio on one objective,
  deadline-salvaged.
• Governor — the benchmark-integrity gate (suite / no-regression), threaded
  through so scaling the engine doesn't scale benchmark-gaming.

Sequencing (and why)

1. Phase 1 — role-runner consolidation, thin/feature-driven (no dispatcher
   dep). Enables the rest; avoids each axis growing its own per-lane
   orchestration (the Tick: drive the codex author across the climb-authoring lanes #123 lesson).
2. Light up the dispatcher on a real long-eval benchmark — the entry gate to
   cross-session depth and parallel (it's proven but dark today).
3. Phase 2 — depth (2a inline, then 2b cross-session), with the metric
   taxonomy + suite gate landing alongside.
4. Phase 3 — parallel dispatch, on the consolidated runner + live dispatcher.

Depth before parallel: depth extends something already built and is cheaper to
make honest; parallel is the larger new capability.

Status

Status: plan — reviewable before any code. Non-goals and open questions are
called out in-doc (esp. for parallel: workspace sharing, member interaction,
budget split, salvage ranking, depth×parallel composition). Cross-linked from
research-loop.md.

No secrets / no large files

Docs only.

🤖 Generated with Claude Code

[github-actions]

Round 1 — reviewed head 55937afd — reviewer hermes/gpt-5.6-terra.

second opinion — terra
Advisory findings from autoresearch — the code owner decides. Reply to disagree; the autoresearch:no-review label opts this PR out.

Verdict: nothing blocking — 1 advisory note.

1 finding attached to the lines below.

[github-actions]

Round 1 — reviewed head 55937afd — reviewer claude/claude-opus-5.

Advisory findings from autoresearch — the code owner decides. Reply to disagree; the autoresearch:no-review label opts this PR out.

Verdict: nothing blocking — 3 advisory notes.

Advisory (non-blocking):

• Phase 1 premise is stale: the authoring roles already run through run_role (docs/design/research-loop-buildout.md:30; high)
• Phase 1 acceptance criteria are already met by current code (docs/design/research-loop-buildout.md:46; high)
• parallel-climb reference does not resolve to anything in the repo (docs/design/research-loop-buildout.md:95; low)

Verified doc cross-references that do resolve: consolidation.md, agent-substrate.md ("The role-runner: one loop replaces five drivers", line 266), dispatcher.md, scaling.md, orchestrator-verify.md, judge-placement.md, plus the code hooks panel_reads/panel_revisions (orchestrator.py) and supports_resume (harness.py, climb.py:777). Claims I could not check from the repo: the dispatcher being "proven but dark" on real Slurm, PR numbers #123/#125, and the climb-lessons notes (the notebook is a separate private repo per architecture.md).