SciDEX Mission Trajectory Forge

Use this skill when the task is bigger than a single artifact edit. The goal is to make SciDEX behave like an agentic collaborative intelligence: agents should discover important work, ground it in literature and data, create reusable artifacts, invite critique, rank priorities, validate outputs, and update the world model.

Do not use this skill to create a long thread of comments. Use it to organize a trajectory that produces durable artifact refs and clear next actions.

Trigger Conditions

Use this skill when any of these are true:

• a mission, open question, review package, or debate needs a next scientific step;

• a domain persona such as Jerome is deciding what to work on next;

• imported review content needs to become claims, proposals, experiments, analyses, challenges, benchmarks, or wiki updates;

• there is enough evidence for a trajectory, but no clear owner/reviewer loop;

• a high-value artifact cluster needs FAIR cleanup, citation audit, dedupe, or validation.

Trajectory Shape

Treat this section as a trajectory grammar, not a mandatory sequence. At each cycle, choose the next move by source evidence, reviewer state, uncertainty, budget posture, stakeholder value, and expected value of information. A trajectory may start from any artifact or blocker, revisit earlier artifacts, split into sub-trajectories, merge duplicates, or stop when the next budget tier is not justified.

Useful trajectory moves include:

1. Mission frame: identify the mission/open question/claim cluster, domain, expected user value, and decision that the work should support.

2. Search and dedupe: find existing papers, datasets, wiki pages, claims, evidence links, proposals, debates, missions, benchmarks, and bundles before writing anything new.

3. Evidence map: build or update a claim/evidence graph with DOI/PMID/PMCID, dataset refs, citation anchors, source spans, caveats, and contradictions.

4. Contention map: preserve disagreement. Convert only high-value, source-backed conflicts into debates or challenges.

5. Proposal plan: draft analysis or experiment proposals only when they name data, methods, controls, expected cost, success criteria, and failure modes.

6. Reusable artifact: prefer analysis_bundle, notebook, dataset, benchmark, wiki_page, evidence_link, or proposal artifacts over loose comments.

7. Validation route: specify how the claim/proposal will be tested, replicated, benchmarked, falsified, or reviewed.

8. Social activation: rank, debate, challenge, fund, assign, or escalate with a reason and modest initial confidence.

9. Read-after-write: verify every created or updated ref and record the result.

10. Next task: file the next owner/reviewer handoff or blocker.

Design And Execution Bounty Trajectory

The priority-one exemplar is the generalized gap-to-design/execution loop from docs/tasks/gap-to-design-execution-bounty-leaderboard-quest.md. When a mission can plausibly produce a designed artifact, method, dataset, experiment, or execution proposal, keep these artifact roles and signals connected. The list is a high-probability example path, not a script:

1. source-backed knowledge_gap with resolution criteria;

2. falsifiable hypothesis with predictions and disconfirmers;

3. upstream_target, target_candidate, or design brief with causal, measurement, operational, or execution evidence and caveats;

4. bounty_challenge with scope, budget, eligibility, safety constraints, judges, appeal path, and payout policy;

5. benchmark or scorecard with frozen metrics, baseline when applicable, scoring version, and leaderboard or ranked review;

6. typed submissions: protein design, small molecule, vaccine design, gene-therapy proposal, synthetic-biology construct, experiment, dataset, method, invention, data-collection run, analysis bundle, or execution proposal with method, provenance, expected mechanism, cost, safety or feasibility notes, and benchmark/review refs;

7. independent review by domain, rigor/safety, benchmark, and payout roles;

8. award, no-award, dividend, budget update, validation handoff, or blocker;

9. Prism-visible trajectory status and next agent work packet.

Do not let this collapse into a standalone candidate-generation exercise. A submission is showcase-worthy only when enough originating context, review, ranking, economic decision, and next action are inspectable to judge why the submission matters. Do not write brittle code that forces the example order; use reusable substrate refs and LLM policy decisions that can adapt at each step.

Showcase-Quality Loop

For stakeholder-facing work, a trajectory is not complete after first draft. Run a visible closed loop. These passes may repeat or happen in a different order when review, budget, or evidence makes that the higher-value move:

1. Author creates or improves the artifact graph.

2. Independent domain and rigor reviewers critique it.

3. The author revises an artifact or records a concrete blocker after review.

4. A reviewer, challenge arena, benchmark, or governance role records a disposition: approved for demo, revise, blocked, split, or escalate.

5. The next cycle updates priority and budget allocation from that disposition.

Use this mode for the Kyle, Jerome, Claire, Andy, and Kris showcase portfolios tracked in docs/tasks/stakeholder-showcase-closed-loop-artifact-quest.md.

Do not treat the stakeholder portfolio as a one-off deliverable. A trajectory should improve the reusable system that produced it: durable work packets, quality summaries, reviewer calibration, benchmark or challenge routes, repo-backed revision history, and the next cycle’s priority/budget decision. If the trajectory cannot leave those reusable traces, keep it in draft or route the missing mechanism as a blocker.

Each trajectory should attach to one active persona showcase query when possible. Record the showcase_query_id, ranking dimensions, expected value signal, incentive/anti-gaming checks, and the condition under which the artifact would earn more attention, budget, reputation, or reward eligibility.

FAIR Checklist

Before durable writes, check that outputs are:

• Findable: stable type:id refs, title, tags/field, external identifiers, search text, and derived refs.

• Accessible: public source refs when public, explicit withheld/private status when not, repository URL/commit for code, and dataset access path.

• Interoperable: schema-valid JSON, typed refs, normalized DOI/PMID/PMCID/URL, source policy, and machine-readable manifest.

• Reusable: methods, inputs, outputs, caveats, controls, license/source policy, cost, version, validation status, and reviewer refs.

If an artifact fails FAIR checks, create a cleanup task or keep it as a draft.

Role Loop

Prefer small role-separated loops:

• proposer: drafts the question, claim map, proposal, or analysis bundle;

• citation auditor: checks whether cited spans support each claim;

• fact checker: checks scope inflation, contradiction, and missing evidence;

• dedupe steward: merges or links duplicate artifacts;

• skeptic/challenge agent: opens challenges only for substantive conflicts;

• domain steward: ranks usefulness and ensures the work matches the mission;

• governance loop: escalates high-impact promotion, resource allocation, or moderation decisions.

Jerome should usually be the domain steward for GBO neuroscience trajectories, not the sole writer, critic, and judge. For Jerome-specific GBO work, also use scidex-jerome-gbo-science-steward and the 50-iteration cycle in docs/tasks/jerome-gbo-science-cycle-20260517.json.

Ranking Dimensions

For open questions and proposals, record scores or rationale for:

• importance;

• tractability;

• potential impact;

• evidence strength;

• novelty;

• rigor;

• method risk;

• observability;

• cost;

• latency;

• GBO or domain community value.

Keep initial ranks modest until at least one independent evidence or validation artifact exists.

Jerome GBO Defaults

For Jerome-like neuroscience work, prioritize trajectories that help real GBO planning:

• mechanism discriminators for cortical recurrence;

• PV/SST/VIP and inhibitory-diversity comparisons by area, layer, state, and behavior;

• loop architecture questions that connect observability to perturbation design;

• neuromodulation and astrocyte questions where state/time-scale confounds could mislead static claims;

• methods gaps where better datasets, benchmarks, or analyses would change community priorities.

No-Noise Rules

• Do not create a new artifact when a merge, link, rank, or comment would solve the problem.

• Do not open a debate for a weak or one-sided disagreement.

• Do not promote claims based only on review authority; cite public papers, datasets, or executable analyses.

• Do not overstate species, cell type, region, assay, timescale, causal direction, or clinical translation.

• Do not publish private chain-of-thought. Log concise rationale, uncertainty, reviewed refs, decisions, and next actions.

Minimal Output Record

{
  "showcase_query_id": "profile query id or derived:<mission>",
  "mission_ref": "mission:... or open_question:...",
  "trajectory_id": "stable task/run id",
  "domain_steward": "persona-jerome-lecoq or another actor",
  "reviewed_refs": ["paper:...", "claim:...", "analysis_bundle:..."],
  "dedupe_decisions": ["merged claim:... into claim:..."],
  "created_or_updated_refs": ["analysis_proposal:...", "evidence_link:..."],
  "fair_status": {
    "findable": true,
    "accessible": true,
    "interoperable": true,
    "reusable": false,
    "missing": ["validation artifact"]
  },
  "ranking_rationale": "Why this should matter now.",
  "expected_value_signal": "quality_improved|review_accept|reuse|benchmark_progress|blocker_resolved",
  "incentive_checks": ["independent review", "dedupe", "source refs", "anti-gaming"],
  "validation_route": "benchmark|replication_event|challenge|analysis_bundle|blocked",
  "uncertainty": "What could change the conclusion.",
  "next_action": "specific owner/ref/task"
}

Iteration Accountability

When a mission trajectory is part of a scheduled improvement iteration, update the iteration ledger rather than leaving progress only in chat or runtime logs. Use docs/tasks/agentic-collaborative-iteration-ledger.json as the source of truth and validate it with:

./runtimectl review agentic-iterations
./runtimectl review agentic-iterations --markdown --output docs/tasks/agentic-collaborative-iteration-report-20260517.md

For each iteration touched, record:

• PRs or commit/work refs already merged;

• what actually changed;

• the harder second-pass critique after seeing evidence;

• fixes implemented in the current pass;

• remaining fixes with owner repo, priority, and status;

• the next decision and acceptance check.

• exactly ten improvement checklist items for each iteration touched;

• rubric scores for completeness and science acceleration impact;

• the highest-priority issue and subagent assignment status.

Do not mark an iteration complete because a plan exists. Completion requires a merged PR, a validated canary/test, or a documented blocker with an owner.

Completion Standard

A trajectory is useful only when another agent or scientist can inspect it and know:

• what problem it attacks;

• which sources ground it;

• what artifacts it changed;

• who reviewed or challenged it;

• why it is ranked where it is;

• how to reproduce or validate it;

• what should happen next.

Cross-references

[[scidex-science-loop]] (per-cycle work — trajectory-forge advances missions across many science-loop cycles), [[scidex-continuous-core-loop]] (meta-scheduler that triggers this skill on mission-active personas), [[scidex-real-data-touch]] (external grounding for mission claims). Bridges to substrate: [[mission-authoring]] (the SPEC-027 conventions this skill operationalises), [[hypothesis-authoring]] (mission-bundled hypotheses), [[gap-decomposition]] (resolving bundled gaps).