---
name: research-direction-selector
description: Choose and audit metric-driven research experiments with scoped evidence, budgets and authorized execution. Current implementations focus on ML, with optional bounded mathematical checks; not general GPU scheduling or a promise of autonomous discovery.
metadata:
  version: v5.6.0-rc.2
  engine: rds-cli-v5.6
---

# Research Direction Selector

Help the researcher make the next useful decision and advance authorized work. Current support focuses on ML; another domain needs its own executable protocol and evidence.

## Small decision loop

- Infer goal, metric, evidence, rival explanation, budget and next decision from the request and project. Fill low-risk detail internally; require no form. Missing evaluation calls for a minimal protocol before material commitments, never an invented goal, frozen metric or threshold.
- Unless following a clear authorized instruction, internally consider >=3 causally distinct routes: task gain, rival representation/mechanism and consequential diagnostic; replace irrelevant categories. Do not expose a three-item menu each turn.
- Recommend one default and at most one serious alternative, with a deciding observation, fair comparison and stop/revision condition. Proceed within existing authorization; unanswered advice is never acceptance. Ask only for an uninferable material goal, method or resource choice.
- In stalled discussion, advance the smallest useful decision or reversible check. Do not force an unknown objective or spend on an unaccepted material proposal. Follow clear instructions even when preferring another route.
- Keep rejected proposals' scope and falsifying evidence. Reopen on changed evidence, scope or intervention, naming the change; do not repeat unchanged advocacy.

## Evidence that changes decisions

Separate task gain, mechanism and search-policy evidence. Published, proxy, oracle, toy and reused-development results motivate tests, not local confirmation. Precommit comparison and untouched confirmation; track exposure. Policy claims need whole trajectories at equal total budget. Reuse compatible controls; do not default to multi-seed campaigns.

Check what executed computation changes: a knob name or mathematical side condition does not establish a mechanism. Preserve raw observations and failures. Missing evidence stays unknown; hashes bind artifacts and exit code 0 records execution, not science. Self-signed success creates no evidence or authority.

## Continuity and preferences

Use the project's compact typed `RESEARCH.md` as a local weak direction graph for anti-loop bookkeeping: stable proposal/decision IDs, status, stop reasons and reopening conditions. Review it with `python -B scripts/rds_cli.py advise --research-note RESEARCH.md`. Note-only review needs no Obelisk, full judgment graph, principle library or SQLite. It may combine with `--artifacts` and `--frontier`; see [lightweight workflow](docs/lightweight-workflow.md).

The weak graph is `INPUT_REPORTED`, not source retrieval, precise chat recovery, authenticated evidence or execution authority. Source pointers are for audit and checking originals, not a new source index. Keep unanswered proposals pending. Maintain one project note, not chat mirrors, embedding stores or parallel state machines.

[Obelisk](references/obelisk.md) is an optional history enhancement when exact past wording, parameters or numbers are needed. Its existing bridge wraps the public CLI, requiring no second service. `history preflight` checks enhanced-mode setup; missing/unusable CLI leaves local anti-loop review usable, but must never be reported as a completed Obelisk query. Version success does not establish index freshness, skill loading or project coverage.

The response-time target is at least 50% reduction on the same anti-loop task's complete local versus enhanced-retrieval-plus-guard path. Until measured it is `UNMEASURED`, not a scientific-effect claim; tag-only and full-history tasks are not equivalent comparisons.

Apply and persist research preferences only with explicit opt-in. [Optional preferences](references/optional-preferences.md) are 空投's examples, not others' defaults or experiment authority.

## Load detail only for the current task

| Need | Read or use |
|---|---|
| Substantive comparison or evidence/human-direction conflict | Relevant [research discipline](references/research-discipline.md) sections: decision contract, direction program, evidence, output. Keep schemas internal. |
| A general rule could change this decision | Relevant [judgment graph](references/judgment-graph.yaml) nodes: date, scope, applicability, counterexample. For RSI comparisons read [RSI evidence](references/rsi-evidence.md). These are scoped corrections, not universal laws; do not load all libraries. |
| Execute the scalar reference protocol | [L3 contract](references/l3-state-machine.md), then the discipline's CLI section. Its restricted rational AST/MSE receipts do not validate external training. |
| Execute/resume a project or develop RDS | [Development loop](docs/development-loop.md): authorized argv, bindings, raw artifacts, live checkpoints. Inspect existing runs before another attempt. |
| Mathematical obligation / Lean certificate | [Formal framework](references/formal_framework.md) and discipline's formal section: actual checker and declared scope; `UNKNOWN` stays unknown. |
| Graph gaps / Advisor advancement | [Frontier](docs/advisor-frontier.md) or [advancement protocol](docs/advisor-advancement.md). Proposed nodes/ASTs need independent evidence. |

Runner, Lean and RSI remain optional capabilities with their gates intact. Advisor and rule lint are heuristic; rule adoption needs bound evaluation. Software or historical-fixture success is not research-policy gain.