Deterministically compile ResearchOS Pre-T5 research artifacts into a source-traceable, schema-valid external_executor/handoff_pack.json. This is the contract used by T5-REBOOST-GATE when it re-boosts project, literature, hypothesis, novelty, risk, and experiment-plan artifacts into executable context, or when an existing handoff must be reconciled, repaired, or validated before project-specific Skill compilation or external execution. Do not ask an LLM to recreate the handoff, run experiments, implement the method, compile the executor Skill Suite, or produce the final realized method.
Instrucciones de origen · Vista previa de solo lectura
name
research-reboost
execution_scope
state_machine
execution_owner
T5-REBOOST-GATE
description
Deterministically compile ResearchOS Pre-T5 research artifacts into a source-traceable, schema-valid external_executor/handoff_pack.json. This is the contract used by T5-REBOOST-GATE when it re-boosts project, literature, hypothesis, novelty, risk, and experiment-plan artifacts into executable context, or when an existing handoff must be reconciled, repaired, or validated before project-specific Skill compilation or external execution. Do not ask an LLM to recreate the handoff, run experiments, implement the method, compile the executor Skill Suite, or produce the final realized method.
Compile research intent into one auditable executor contract. Treat reboost as deterministic semantic recompilation, not LLM summarization. Produce method_intent in the same compiler pass as context_reboost; never run it as a parallel, independent drafting path.
Required result
Write external_executor/handoff_pack.json conforming to references/handoff_pack.schema.json. Keep all substantive statements traceable through stable source references. Mark unresolved conflicts; do not silently choose a convenient interpretation.
The pack is an execution-time source of truth for T5 external work, but method_intent is only an implementation constraint. It is never the final Method source for T8.
Workflow
1. Locate inputs and establish the output boundary
Find the project root and the required Pre-T5 files listed in references/reboost-protocol.md. Use project-relative paths in the pack. Do not copy whole source documents into the handoff.
In standalone Codex-style environments, run the deterministic inventory first:
Treat the inventory as discovery evidence, not semantic interpretation. The current T4.5 flow supplies the selected Candidate dossier and kill_criteria.yaml; idea_scorecard.yaml and risks.md are legacy fallbacks, not extra files a current workspace must recreate. When present, research_dossier.json, validation_map.yaml, contribution_hypothesis_map.yaml, and ideation/proposal/research_proposal.md preserve the post-novelty research meaning that a short hypothesis page cannot safely carry. The Proposal must be read together with proposal_manifest.json, recorded in source_manifest, and carried in context_reboost.research_context.proposal_context. It is a planning source, not empirical evidence, not a substitute for exp_plan.yaml, and not a final writer fact source. If neither member of a required source role is available, continue only far enough to produce a blocked diagnosis; do not invent its content.
When this Skill is executed inside the ResearchOS T5 state machine, the Runtime invokes the registered compile_research_reboost_handoff() tool directly, without an LLM-authored handoff_pack argument. The tool is the deterministic derivation, validation, and publication boundary: it reads current sources, writes a private candidate, runs the bundled validator, then atomically publishes the pretty-printed JSON, T5 executor control files, and diagnostics under . A failed rebuild leaves a diagnostic rather than a completed pack or stale executor controls. remains owned by workspace initialization. An explicitly supplied is legacy compatibility input only and is audited before any deterministic repair; it is not the normal T5 path.
external_executor/report/
blocked
external_executor/expr/
handoff_pack
2. Read sources by decision relevance
Read all available required sources. Backtrack into optional paper notes, Cross-domain catalogs, the literature resource catalog, resources, or user seeds only when a required source leaves a concrete decision unsupported or ambiguous. When literature/resource_catalog.jsonl exists, read it together with literature/resource_catalog_summary.json before declaring a baseline, dataset, benchmark, model, code, or supplement unavailable. The catalog is a paper-associated discovery ledger, not an approved asset list: it may identify an official-source lead and a Phase B requirement, but it cannot certify access, licensing, revision, security, protocol equivalence, or empirical performance. Record every actually consulted file in source_manifest; set used=false for discovered but unread optional files. Generated workspace navigation files named _DIR_GUIDE.md are never literature, resource, or evidence sources and must not enter the manifest; legacy entries may remain visible for audit but their changing text does not invalidate a handoff. A Cross-domain catalog is retrieval context, not direct experiment evidence: it can guide a baseline or follow-up reading decision but cannot certify a mechanism, implementation detail, baseline equivalence, or result.
Load references/reboost-protocol.md before resolving conflicts or deciding whether optional backtracking is required.
3. Build an evidence ledger before synthesis
Assign stable IDs to sources, hypotheses, mechanism invariants, modules, baselines, claims, experiments, and gates. For each decision-bearing statement, record one or more source_refs with a locator and a short relevance note.
Distinguish:
source fact: explicitly supported by a Pre-T5 artifact;
reconciliation: a choice made under the precedence policy;
inference: a conservative connection not stated verbatim;
unresolved item: a missing or conflicting decision that cannot be settled safely.
Never present inference as source fact. Put material unresolved decisions in unresolved_items and set generation_status accordingly.
4. Reconcile the research contract
Compile context_reboost from the ledger:
project goal, research question, central hypothesis, and falsification conditions;
research_context: research problem, scholarly stakes, conditional practical and commercial implications, affected stakeholders or processes, contribution intent, evidence status, and source references. Preserve unknown or proposed_not_verified; never translate this context into observed business or scientific results;
target setting and explicit exclusions;
core mechanism, mechanism invariants, and contribution intent;
novelty-audit resolution and required-baseline consequences;
execution priorities, risks, and known context mismatches.
when present, resource_discovery_context: catalog path, counts, types, and the Phase B verification boundary. Do not promote a discovered link to an acquired resource or an experiment-ready baseline.
Use the conflict rules in references/reboost-protocol.md. Novelty constraints control required baselines and must-not-claim boundaries; the experiment plan controls protocol details only where it does not weaken those constraints. Preserve every material mismatch even after applying precedence.
5. Draft method intent
Derive method_intent from the reconciled context, not directly from one source file. Keep both required constants:
Specify candidate modules, intended interfaces, algorithm flow, refinement permissions, silent-change prohibitions, mechanism-to-ablation mappings, and an initial framework sketch. Label core, candidate, and supporting modules explicitly. Do not turn an implementation convenience into a contribution or settle an unsupported design choice.
6. Compile executable evidence contracts
Build the following as peer fields of the pack, using IDs rather than duplicated prose:
baseline_matrix: requirement, role, availability, fairness, reproduction target, and substitution policy;
Fix all schema and semantic errors. A completed pack must return exit code 0. A structurally valid pack may remain needs_review or blocked, but must not be presented as ready for external execution.
Status gates
completed: required sources are available, no blocking mismatch remains, cross-references resolve, and validation passes.
needs_review: the pack is usable for human resolution but contains a material non-blocking ambiguity, substitution, or inference requiring approval.
blocked: a required source, central scope decision, novelty constraint, or minimum executable contract is unavailable or contradictory.
Do not invoke project-specific skill compilation or the external executor unless status is completed.
Output discipline
Return the validated handoff artifact first, followed by a concise status summary listing source coverage, mismatches, unresolved items, and validation result. Do not emit hidden reasoning, fabricate research content, run experiments, implement code, or generate realized_method_package.
Resources
references/handoff_pack.schema.json: normative JSON Schema for the output.
references/reboost-protocol.md: source roles, precedence, reconciliation, provenance, and quality gates.
scripts/inventory_sources.py: deterministic required/optional input discovery and hashing.
scripts/validate_handoff.py: dependency-free schema-subset and ResearchOS semantic validation.
