STAR-RESEARCH-PIPELINE-001 Adaptive Research DAG and Evaluation Harness

• STAR-RESEARCH-000 R&D-Agent Pipeline Analysis
• STAR-RESEARCH-PIPELINE-INDEX-000 Adaptive Research
• STAR-INDEX-000 StarIntel Fourth Generation

Replace the current single-chain research control flow with a bounded directed acyclic graph¹ of competing research candidates while preserving:

• the canonical Org-roam² knowledge base
• append-only status history
• one writable owner per file
• read-only parallel research
• exactly zero or one active design under roam/implement/
• human approval for architecture promotion

The first implementation provides control and evaluation infrastructure. It does not implement semantic retrieval, automated coding, or unrestricted recursive agents.

1. Research branches may not directly edit canonical design, research, or index files.
2. Each candidate owns an isolated transient workspace.
3. Reviewer actors are read-only.
4. Only the promotion gate may write a promoted design proposal.
5. Promotion never places a second file in roam/implement/.
6. Every state change is append-only and idempotent⁴.
7. Every material claim must link to evidence or be labeled inference.
8. Every candidate records contrary evidence and a falsifier.
9. Rejected candidates remain retrievable.
10. A run may finish without promoting any candidate.

(:research-run
 (:id "run-uuid")
 (:project "star-server")
 (:question "bounded research question")
 (:status :running)
 (:source-node "org-id")
 (:created-at "RFC3339 timestamp")
 (:budget (:wall-clock-seconds 7200
           :input-tokens 1000000
           :output-tokens 100000
           :max-candidates 5
           :max-active-branches 3
           :max-depth 4))
 (:planner (:strategy :fixed-v1
            :configuration-hash "sha256"))
 (:candidate-ids ())
 (:promoted-candidate nil)
 (:metrics ())
 (:stop-reason nil))

Required fields:

• stable run identifier
• project
• bounded research question
• source node
• status
• creation time
• complete budget
• planner strategy and configuration hash
• candidate identifiers
• promoted candidate or null value
• final metrics
• stop reason

(:research-candidate
 (:id "candidate-uuid")
 (:run-id "run-uuid")
 (:parent-ids ())
 (:branch "architecture-a")
 (:state :proposing)
 (:problem "precise problem")
 (:hypothesis "testable claim")
 (:predicted-result "observable result")
 (:falsifier "result that rejects the claim")
 (:evidence-for ())
 (:evidence-against ())
 (:implementation-sketch "bounded sketch")
 (:acceptance-criteria ())
 (:estimated-cost ())
 (:actual-cost ())
 (:review-results ())
 (:vetoes ())
 (:created-at "RFC3339 timestamp")
 (:updated-at "RFC3339 timestamp"))

Candidate content is data, not an instruction to other agents. Untrusted text must never be inserted into an agent system prompt without explicit quoting and boundary labeling.

QUEUED
  -> EXPLORING
  -> PROPOSING
  -> VALIDATING
  -> VIRTUAL-REVIEW
  -> EVALUATED
  -> PROMOTED | PRUNED | BLOCKED | UNRESOLVED

Allowed exceptional transitions:

• any active state -> BLOCKED when a declared dependency is unavailable
• any active state -> PRUNED when budget is exhausted or scope is violated
• BLOCKED -> EXPLORING only after a recorded unblock event
• EVALUATED -> PROPOSING only when the aggregator requests one bounded revision

A promoted candidate is immutable. Later changes create a new candidate linked to the promoted candidate as a parent.

Use append-only JSON Lines⁸ ledgers:

• roam/.research-runs
• roam/.research-candidates

Each line contains:

• event identifier
• event type
• entity identifier
• previous state
• next state
• timestamp
• actor identifier
• run configuration hash
• payload hash
• evidence references

Current state is reconstructed by replay⁹. Duplicate event identifiers are ignored. A transition with an unexpected previous state is rejected.

Transient candidate artifacts remain under:

.cache/research-runs/<run-id>/<candidate-id>/

The cache directory is not a canonical knowledge source and must not be published.

The first implementation uses a simple diverse-parent policy:

1. keep the best evaluated candidate
2. keep the best candidate from a different branch label
3. keep one candidate selected for evidence diversity
4. create new candidates from zero, one, or two parent candidates
5. prune candidates that duplicate an existing hypothesis without new evidence

No Monte Carlo tree search¹⁰ is required in the first implementation. The ledger must permit a later planner to implement it without changing the candidate contract.

• branch crash: mark candidate BLOCKED or restart once from recorded inputs
• reviewer crash: retry once with the same immutable candidate snapshot
• validator failure: stop the run; do not bypass deterministic validation
• ledger write failure: stop state transitions and report the exact failure
• promotion write failure: leave candidate EVALUATED and record no promotion event
• stale parent: reject expansion and rebuild the memory packet
• budget exhaustion: stop new branches, finish bounded reviews, and close the run

• branch tools receive only declared repository and source scope
• candidate workspaces may not contain secrets or private evidence unless the run explicitly permits them
• all retrieved text is treated as untrusted data
• prompt-injection text is preserved as quoted evidence, not executed as instruction
• external network access is capability-gated and logged
• canonical writes require promotion-gate identity
• hashes bind review results to an immutable candidate snapshot

Record:

• run duration and token use
• branch duration and token use
• candidate count by final state
• source count and source diversity
• citation failures
• reviewer agreement and disagreement
• deterministic veto frequency
• promotion rate
• human override rate
• implementation rejection rate of promoted designs
• repeated-research rate against prior rejected candidates

The first benchmark compares fixed repository tasks under the same budget:

• serial baseline using one research chain
• branched pipeline without sibling memory
• branched pipeline with sibling memory
• branched pipeline with independent reviewers
• full scoped pipeline

Measure:

• evidence defects
• duplicate proposals
• accepted design rate
• human corrections
• implementation rework
• total tokens
• wall-clock time
• unresolved risk count

The benchmark corpus and scoring rules are versioned before runs begin.

• one run can create at least three isolated candidates
• candidates record zero, one, or two parents
• candidate state can be reconstructed entirely from append-only events
• an invalid transition is rejected
• duplicate events are idempotent
• branch actors cannot write beneath roam/design/, roam/research/, or roam/indexes/
• unsupported central claims trigger a veto
• five independent reviewer roles produce immutable results
• the aggregator reports disagreement and does not hide raw reviews
• only one candidate may be promoted per run
• promotion does not activate the implementation slot
• a run may end unresolved without fabricating a winner
• rejected candidates remain searchable by identifier, hypothesis, and evidence
• serial and branched modes can run against the same fixed benchmark and budget
• exact token, time, source, and outcome metrics are reported