decentralized_auto_research_state_v0

decentralized_auto_research_state_v0 is a LoopX protocol for running autonomous research without introducing a single leader agent. It borrows the useful parts of Arbor's public design, especially durable hypothesis state, worktree isolation, dev/held-out evaluation, and replayable evidence, but maps them onto LoopX's shared control plane: todos, claims, quota, run history, rollout events, gates, and read-only projections.

The goal is to let multiple agents explore research hypotheses in parallel while preserving provenance, ownership, and promotion rules. No agent owns the whole research tree. The source of truth is the append-only state graph; each agent sees a scoped frontier through quota should-run --agent-id ....

Arbor Signals To Preserve

Primary public Arbor surfaces reviewed:

• README.md: Arbor frames the work as autonomous research through hypothesis-tree refinement, with a Coordinator, Executors, worktrees, held-out validation, reports, replay, and a benchmark zoo.
• docs/how-it-works.md: the six-step cycle is observe, ideate, select, experiment, evaluate, backpropagate, then merge or prune.
• src/coordinator/idea_tree.py: each node records hypothesis, status, insight, result, score split, test score, branch ref, grounding, related-work audit, eval status, stop reason, and attempt count.
• src/coordinator/tools/executor_run.py: executor outcomes distinguish scored success from needs_retry for timeout, max-turn, eval-crash, or unparseable reports, while preserving branch/report/diff for continuation.
• docs/search.md: grounded ideation and novelty audit are separate lanes; text fetched to inspire an idea is not reused to certify novelty.
• arbor-zoo/algotune_knn: the showcase benchmark packages an editable solver, protected harness, dev/test split, provenance, and one score: line.

LoopX should preserve these product lessons, not Arbor's centralized topology.

Non-Goals

• Do not add a LoopX-wide "research director" agent that owns all planning.
• Do not replace todos, run history, or rollout events with a second research database.
• Do not let a showcase dashboard mutate source state.
• Do not treat an advisory grounded-search result as proof of novelty.
• Do not merge or publish a research result without held-out evidence or an explicit gate when the boundary requires one.

Source Protocol

Source state is writable only through LoopX lifecycle commands, project-owned state files, or future narrow kernel APIs.

The companion auto_research_lane_contract_v1 defines which decentralized agent lanes may create, execute, evaluate, promote, retire, or narrate these records. The companion auto_research_role_state_machine_v0 defines the always-on digital employee role map and state transitions. This file defines the record shapes and projections; the companion contracts define capability ownership and transition evidence without introducing a leader agent.

Source state	Existing or proposed anchor	Purpose
research_contract_v0	proposed packet, registry goal metadata	Public-safe objective, editable scope, protected scope, metric direction, dev/held-out commands, budget, and stopping condition.
todo_item_v0	loopx todo	Formal executable work, user gate, blocker, or monitor. Research work remains todo-backed.
research_hypothesis_v0	proposed rollout event and optional domain pack	A hypothesis node linked to a todo, parent hypothesis, agent lane, mechanism family, and current status.
research_evidence_event_v0	proposed loopx_rollout_event_v0 specialization	Append-only evidence for an attempt: score, split, command label, branch, artifact refs, eval status, and boundary facts.
dataset_window_contract_v0	loopx/ml_experiment.py	Train/dev/held-out window or split contract, including missing-window policy.
agent_lane_next_action_v0	quota should-run --agent-id ...	The current agent's selected frontier item; it does not replace the global state graph.
operator_gate	loopx operator-gate, review packet	Promotion, merge, private-material, or publication decision.
human_reward	loopx reward	Run-bound owner judgment, not general write permission.

research_contract_v0

{
  "schema_version": "research_contract_v0",
  "goal_id": "loopx-meta",
  "research_objective": "optimize a benchmark task under a protected evaluator",
  "editable_scope": ["solution.py"],
  "protected_scope": ["eval.py", "task.py", "data/**"],
  "metric": {"name": "speedup", "direction": "maximize"},
  "dev_eval": {"label": "eval_dev", "command_label": "bash eval.sh dev"},
  "holdout_eval": {"label": "eval_test", "command_label": "bash eval.sh test"},
  "promotion_policy": "requires_holdout_improvement_and_clean_boundary",
  "budget": {"max_attempts": 8, "max_parallel_lanes": 3}
}

research_hypothesis_v0

{
  "schema_version": "research_hypothesis_v0",
  "goal_id": "loopx-meta",
  "hypothesis_id": "hyp_0004",
  "parent_hypothesis_id": "hyp_0002",
  "todo_id": "todo_123",
  "lane_id": "agent:research-executor",
  "claimed_by": "research-executor",
  "mechanism_family": "vectorized_distance_kernel",
  "hypothesis": "Batch query points through a shared index to reduce per-query overhead.",
  "status": "active",
  "frontier_rank": 2,
  "source_refs": ["research_contract:knn_speedup"],
  "grounding_refs": [],
  "novelty_audit_ref": null
}

Status vocabulary:

Status	Meaning
proposed	Drafted but not yet accepted into executable work.
active	Runnable frontier item, normally backed by an open agent todo.
running	Claimed attempt is in progress.
needs_retry	Attempt produced no trustworthy score but left resumable evidence.
supported	Dev evidence supports the hypothesis; not yet promoted.
contradicted	Evidence shows regression or guardrail failure.
promoted	Held-out evidence and promotion policy accepted it into the current best artifact.
retired	No longer worth exploring, but retained as negative evidence.

research_evidence_event_v0

{
  "schema_version": "research_evidence_event_v0",
  "goal_id": "loopx-meta",
  "hypothesis_id": "hyp_0004",
  "todo_id": "todo_123",
  "agent_id": "research-executor",
  "attempt": 1,
  "split": "dev",
  "metric": {"name": "speedup", "value": 11.4, "direction": "maximize"},
  "baseline_metric": 8.7,
  "primary_metric_status": "improved",
  "eval_status": "scored",
  "code_ref": "codex/research-hyp-0004",
  "artifact_refs": ["experiment:hyp_0004/report", "diff:hyp_0004"],
  "protected_scope_clean": true,
  "private_artifacts_recorded": false,
  "raw_logs_recorded": false
}

needs_retry is first-class evidence, not a generic failure:

{
  "schema_version": "research_evidence_event_v0",
  "hypothesis_id": "hyp_0005",
  "eval_status": "failed_to_run",
  "stop_reason": "max_turns",
  "attempt": 1,
  "code_ref": "codex/research-hyp-0005",
  "resume_policy": "resume_from_code_ref_or_retire",
  "primary_metric_status": "inconclusive"
}

Projection Protocol

Projection state is read-only. It may rank frontiers for display, but it must carry source refs and remain recomputable from source state.

Projection	Purpose
decentralized_research_frontier_v0	Per-agent queue of runnable or blocked hypotheses after quota, claim, capability, and boundary checks.
research_evidence_graph_v0	Read-only graph joining hypotheses, todos, attempts, branches, metrics, gates, and promotion decisions.

decentralized_research_frontier_v0

{
  "schema_version": "decentralized_research_frontier_v0",
  "goal_id": "loopx-meta",
  "agent_id": "research-executor",
  "selected": {
    "hypothesis_id": "hyp_0004",
    "todo_id": "todo_123",
    "reason": "current-agent claim and dev evidence gap",
    "allowed_action": "run_dev_attempt"
  },
  "blocked": [
    {
      "hypothesis_id": "hyp_0002",
      "blocked_by": "claimed_by:research-curator",
      "visible_as_context": true
    }
  ],
  "promotion_candidates": [
    {
      "hypothesis_id": "hyp_0004",
      "requires": ["holdout_eval", "boundary_scan"]
    }
  ]
}

This projection intentionally replaces a centralized Coordinator decision. The kernel selects only what the current agent may attempt; the agent still does semantic implementation within its allowed boundary.

Decentralized Cycle

Arbor's cycle maps to LoopX as a distributed control loop:

Arbor concept	LoopX decentralized equivalent
Observe	Agent reads quota should-run, active state, selected frontier, evidence graph, and relevant source docs.
Ideate	Any authorized lane may propose research_hypothesis_v0 as a todo-backed candidate.
Select	quota should-run --agent-id filters by claim, gate, capability, and boundary.
Dispatch	The selected agent works in its own worktree and writes evidence.
Backpropagate	A deterministic projection summarizes supported/contradicted lessons into the evidence graph; no single agent rewrites global truth.
Decide	Promotion policy plus operator gates decide merge, retire, retry, or continue.

Search Lanes

LoopX should adopt Arbor's separation of search lanes:

Lane	When	Authority	Writes
grounded_ideation	Before or during hypothesis proposal	Advisory input to a candidate hypothesis.	grounding_refs on the hypothesis.
novelty_audit	After a hypothesis has real evidence	Advisory contribution/overlap check.	novelty_audit_ref and evidence note.

The two lanes must not share fetched text as proof. If a source shaped the idea, it can be cited as grounding, but a later novelty audit must run an independent source pass before claiming novelty.

Promotion Policy

A research hypothesis is promotable only when:

1. the editable/protected scope boundary is clean;
2. dev evidence is scored and public-safe;
3. held-out evidence satisfies the metric direction and margin;
4. required user/controller gates are resolved;
5. the promoted artifact has a branch or commit ref;
6. the evidence graph records negative evidence for close alternatives.

Promotion is a state transition, not a chat conclusion. It should write a todo completion, evidence event, and optional promotion gate/reward overlay.

Integration With Existing LoopX

Already available:

• todo_item_v0 has claims, task classes, action kinds, blockers, user gates, and resume conditions.
• agent_lane_next_action_v0 scopes work per registered agent.
• loopx/ml_experiment.py already has hypothesis_ledger_v0, dataset_window_contract_v0, and advisory result packets.
• long_horizon_agent_state_protocol_v0 already defines source/projection separation and concurrent lane views.
• loopx auto-research frontier --fixture <public.json> --agent-id <agent> now renders a fixture-backed decentralized_research_frontier_v0, and research_evidence_graph_v0 without launching experiments or depending on a leader agent.
• loopx auto-research frontier --goal-id <goal> --agent-id <agent> now renders a live todo/quota-backed frontier for the current agent from LoopX status projection.
• The auto-research worker path now uses a built-in lightweight metric kernel for smoke/demo evidence instead of a shipped domain-specific starter pack.
• loopx auto-research evidence --contract <research_contract.json> --eval-result <eval.json>... now builds an auto_research_evidence_packet_v0 containing a public-safe research_hypothesis_v0 and split-aware research_evidence_event_v0 records. It preserves needs_retry, negative evidence, protected_scope_clean, and branch/artifact refs without recording raw logs or private artifacts.
• loopx auto-research append-evidence --packet <packet.json> appends that packet into the existing loopx_rollout_event_v0 log as one research_hypothesis event plus one research_evidence event per split. Re-running the same packet skips existing event ids, so heartbeat retries do not duplicate evidence.
• loopx auto-research frontier --goal-id <goal> --agent-id <agent> reads research_hypothesis and research_evidence rollout events back into research_evidence_graph_v0 and derives promotion/retirement candidates for the live frontier.

Needed next:

• let pane-local workers consume the frontier directly from their Codex TUI;
• keep product narration and public launch claims outside the kernel until a separate product surface intentionally consumes the compact evidence graph.

Acceptance Checks

An implementation is acceptable when:

• no single agent is required to own or mutate the full hypothesis graph;
• every executable hypothesis is linked to a todo and claim;
• per-agent frontier selection is derived from quota/status, not chat memory;
• needs_retry keeps resumable evidence instead of collapsing into done;
• grounded ideation and novelty audit are separated;
• held-out promotion is explicit;
• user-facing research summaries derive from compact frontier and evidence graph refs, not a bespoke kernel packet;
• public projections contain no raw logs, private paths, credentials, or raw private documents;
• the public narration can be rendered from public-safe evidence refs.