Research Mode Activation

0. Emit SKILL_INVOKE

Post SKILL_INVOKE to the bus before any stateful action (host prefix is required for agent-originated posts).

Type: SKILL_INVOKE
To: all
Message: host=<machine> [skill=research] [mode=side_effecting] [args_hash=<sha256>] [session=<session_id>]

(The skill invocation runtime injects the caller's canonical identity as from_id. The message body MUST begin with host=$(hostname) or canonical mesh host name per bus-protocol.md §Machine tagging.)

Example bash invocation:

python ~/bin/bus-global.py post <from_id> all SKILL_INVOKE \
  "host=$(hostname) [skill=research] [mode=side_effecting] [args_hash=<args_sha256>] [session=<session_id>]"

0.1 Authorization Preflight (MANDATORY before Sweep)

Every invocation MUST explicitly select --mode preview|persistent. If the mode is absent, stop before searching and ask the operator to choose. A skill invocation, broad autonomy request, or research topic alone is not permission for persistent mutations. The runtime MUST NOT infer authorization.

Run the guard before any internal or external source sweep:

# Read-only source preview: no persistent research mutations.
python ~/.agents/scripts/research-run-guard.py preflight \
  --mode preview --session-id <session_id>

# Durable run: pass each grant only when the operator explicitly authorized it.
python ~/.agents/scripts/research-run-guard.py preflight \
  --mode persistent --session-id <session_id> \
  --authorization-ref <operator-message-or-decision-receipt> \
  --authorize-mutation branch \
  --authorize-mutation worktree \
  --authorize-mutation ledger \
  --authorize-mutation artifact \
  --authorize-mutation commit \
  --authorize-mutation bus

If persistent preflight exits nonzero, post BLOCKED and stop before the sweep. Do not downgrade silently to preview mode. If the operator explicitly authorizes a staging directory, add both --receipt-dir <path> and --staging-authorized so the preflight receipt and any later interruption receipts have a permitted destination.

Mode contract

Preview: may read internal sources, query external sources, classify unreviewed candidates, and post coordination receipts. It MUST NOT create or switch a branch/worktree, append to the ledger, write a research artifact, report candidates as findings, commit, or push. Its user-facing output is limited to query coverage and candidate URLs labeled UNREVIEWED / NOT REPORTED AS FINDINGS.
Persistent: executes the full Sweep → Ingest → Synthesize → Dispatch contract only after all six grants pass. The authorization reference and preflight receipt are part of the run evidence.

If a run becomes blocked after queries have started, preserve reproducible query inputs only when a staging directory was explicitly authorized:

python ~/.agents/scripts/research-run-guard.py stage-query \
  --session-id <session_id> --receipt-dir <authorized_path> \
  --staging-authorized --source <source_name> --query <query_text> \
  --retrieved-at <YYYY-MM-DDTHH:MM:SSZ> --result-count <N> \
  --candidate-url <URL> --stop-reason <reason>

This receipt is provenance, not ingestion: it is content-addressed, contains no finding claims, and remains unreviewed-candidates / not-reported.

You are operating in RESEARCH MODE — maximum intelligence gathering, source-grounded, ledger-ready output.

Active Configuration

Tempo: OVERNIGHT (deep research takes time; don't rush the synthesis)
Autonomy: Mode-bound — preview is read-only; persistent requires the complete explicit mutation grant set
Pipeline: Sweep → Ingest → Synthesize → Dispatch

Task

$ARGUMENTS

Research Pipeline

Design Check (pre-Sweep) — before searching, identify known epistemological weaknesses in the research domain:
- Case-study bias: Will the evidence base rely on creator/practitioner case studies? If so, flag that findings will be survivorship-biased by construction and plan for platform-mechanic evidence to anchor confidence.
- Single-source claims: Are there claims that rely on a single source? Flag them for corroboration requirements at the Gate.
- Stale statistics: Does the domain involve fast-changing metrics (platform growth, market share, algorithm behavior)? Flag that statistical claims will need freshness verification at the Gate.
- Platform-reported figures: Are key statistics self-reported by platforms with a commercial interest? Flag for source-type classification at Classify. (Origin: 2026-09-08 AAR — survivorship bias and stale statistics were caught in ARCANA review, not at research design time. Catching them pre-Sweep is cheaper than catching them in review.)
Sweep — cast wide ([daily-research], [industry-watch], [anthropic-watch], [preprint-scan] as relevant)
Classify — all local-model and Gemini output enters as HYPOTHESIS, not FINDING. Promotion to FINDING requires human review or external-source corroboration. Survivorship bias screening: flag findings based on creator/practitioner case studies as lower confidence (max 0.85) than findings based on platform mechanics or independent measurement. Case studies are survivorship-biased by construction — the creators who failed don't write case studies. Distinguish: platform mechanics (durable, non-copyable, adoption-resistant) vs creator systems (depreciating, copyable, adoption-eroding). (Origin: 2026-09-08 AAR — ARCANA review found case-study evidence base was epistemically invalid under fat-tailed distributions.)
Gate — run citation gate on any output with author-year citations. Run claim decomposition on high-stakes outputs (pitch, paper, client-facing). Primary-source verification: before ingesting any statistical claim cited via a secondary source, verify it against the cited primary source. If the primary source is not accessible or doesn't contain the figure, label as HYPOTHESIS, not FINDING. Freshness check: for any statistical claim, verify the data's collection date, not just the citation's publication date. Statistics have expiry dates — a 2024 zero-click figure cited in a 2026 article is stale. (Origin: 2026-09-08 AAR — "67% reach decline" attributed to ConvertKit 2024 was not findable in the primary source; "58%" zero-click figure was 2024 data cited without date in a 2026 secondary source.)
Ingest — only VERIFIED or PROMOTED findings go to ledger via [research-ingest]. HYPOTHESIS-level output stays in docs/research/ with labels intact. Before the cognition post batch: emit a fresh SKILL_INVOKE (the 300s provenance window does not survive a long sweep) and run python ~/bin/bus-global.py status to persist the local bus mirror — the provenance check reads the mirror, not the hub, so a bridge-posted invoke is invisible until status pulls it down. See skills/research-ingest/SKILL.md §0b. (Origin: 2026-09-20 AAR — ingest rejected twice on mirror lag + expired window.)
Synthesize — compress to 3-5 key findings with source citations; use [research-digest] to verify coverage. Before grading the artifact, run python scripts/intel_surge_quality_check.py <artifact.md> (R1-R4 structure + declared grade); errors cap the grade at C. Record unanswered questions and conjectures as open_question / conjecture rows ([INF] until reviewed) per skills/explore-mode/SKILL.md.
Dispatch — post SITREP to bus; suggest [competitive-intel] if competitor signals found, [evidence-pack] if findings need bundling for external use

Quick-Access Skills

Sweep: [daily-research], [deep-research], [industry-watch], [anthropic-watch], [web-research]
Evidence: [preprint-scan], [evidence-grade], [hallucination-check]
Ingest: [research-ingest], [research-digest], [ledger]
Dispatch: [competitive-intel], [evidence-pack], [sitrep], [send-email]
Meta: [brainstorm], [base120], [decision-log]

Rules

Source everything — every claim needs a URL, arXiv ID, or [ESTIMATE: basis] tag
[ESTIMATE] tag format — any quantitative claim, timeline, or prediction without a direct source MUST be tagged at write time as [ESTIMATE: <basis for the estimate>]. The tag must state the reasoning, not just flag the estimate. Example: [ESTIMATE: 5-10 years from commercial viability, based on typical research-to-product timelines for emerging silicon technologies — no direct source]. Do NOT add the tag post-hoc during self-review — tag at write time. If you cannot provide a basis, the claim should be removed, not tagged. (Origin: 2026-09-02 AAR — "5-10 years from commercial viability" written without source or tag; caught in self-review, required post-hoc tagging.)
No unsupported statistics — bare numbers without a source fail the ingest gate
Primary-source verification for secondary-sourced statistics — any statistical claim cited via a secondary source (e.g., a marketing blog citing a research report) MUST be verified against the cited primary source before ingestion as a FINDING. If the primary source is not accessible, doesn't contain the figure, or the attribution is inaccurate, label as HYPOTHESIS. Do not promote secondary-sourced statistics to FINDINGS without primary-source confirmation. (Origin: 2026-09-08 AAR — "67% of creators experienced reach declines" attributed to ConvertKit 2024 was not findable in the primary source; the sweep trusted the secondary attribution.)
Statistical freshness checking — for any statistical claim, verify the data's collection date, not just the citation's publication date. A 2024 data point cited in a 2026 article is stale unless explicitly confirmed as still current. Flag the collection date in the finding's source metadata. (Origin: 2026-09-08 AAR — "58%" zero-click figure was 2024 SparkToro/Datos data cited without date in a 2026 secondary source; the current 2026 figure is 68.01%.)
Survivorship bias screening — findings based on creator/practitioner case studies are capped at confidence 0.85 and labeled as PRACTITIONER_EVIDENCE, not DISCOVERY. Case studies are survivorship-biased: the creators who tried the same strategy and failed don't write case studies. Distinguish platform mechanics (structural, durable, non-copyable) from creator systems (anecdotal, depreciating, copyable). Weight platform-mechanic findings higher than creator-system findings in synthesis. (Origin: 2026-09-08 AAR — ARCANA review found the case-study evidence base was epistemically invalid under fat-tailed distributions; Taleb lens showed expected value of hub-and-spoke is ~2%, not the 217% cited from a single case study.)
Ingest before reporting — findings must hit the ledger before ANY report, including intermediate SITREPs. The SITREP must include ledger IDs, not "0 entries." If you have findings to report, ingest them first. (Origin: 2026-09-02 AAR — SITREP posted with "0 ledger entries" because ingest ran after the SITREP instead of before.)
Confidence floor: 0.70 — below that, flag as unverified and do not ingest as DISCOVERY
EU AI Act and NIST COSAiS are always relevant — surface compliance angles on every governance-adjacent finding
No Ollama on MBP
Verify-pass mandatory on adoption recommendations: Whenever a research swarm produces a synthesis that recommends ADOPT, LIFT, COPY, or DEPEND ON an external project/library/tool, dispatch a verify-pass agent (IP/legal/operational risk audit + source-read of 2-4 files) BEFORE the decision-log ADR is written. Verdict: GREEN (proceed) / YELLOW (proceed with note) / RED (pause and reconsider). Cost: 1 agent + ~700-word budget. Origin: 2026-04-26 OR-research session — Lane B fabricated stale OpenCode details that propagated through two synthesis turns until Lane F caught them; Lane K source-read separately corrected a ClawRouter "code to lift" claim. Receipts: ledger:clp-d016b8f32ba7, clp-163d98f976c2, clp-f7d33313385b.

Citation Gate (mandatory for local-model and Gemini output)

Any research output containing author-year citations (e.g., "Smith et al., 2018") from a local model (Ollama, Gemini, or any non-web-grounded source) MUST pass through the citation gate before ingestion or use in any downstream artifact (paper, pitch, email, case study).

Gate Protocol

Extract — list every Author (Year) or Author et al. (Year) citation in the output
Classify — for each citation:
- Canonical: widely-known foundational work (e.g., Janis 1982 Groupthink, Rizzolatti 2000 mirror neurons) → PASS with spot-check
- Specific: precise author+year+journal claim → MUST VERIFY via [preprint-scan] or web search
- Generic: placeholder-sounding names ("Smith et al., 2018", "Jenkins & Patel, 2019") → PRESUME FABRICATED until verified
Verify — for Specific and Generic citations:
- Search for the exact paper (author + year + key finding)
- If found: tag [VERIFIED: <URL or DOI>]
- If not found: tag [FABRICATED] and exclude from all downstream artifacts
Report — citation gate summary: N citations extracted, M verified, K fabricated, J canonical-pass

When to Apply

After any [deep-research] session using local models
After any Gemini research delivery (per gemini-guardrails.md arXiv rule)
After ingesting remote-node research team output (remote-node dormant since 2026-07-01 — historical reference only)
Before any finding enters a paper, pitch, or external communication

Why This Exists

Local models (gemma4, qwen3.5, mistral) and Gemini fabricate citations at a ~30-40% rate. They mix real canonical references with invented author-year combinations that sound plausible. Research Team v2 (Apr 17 2026) BKI lane produced 10 "peer-reviewed studies" — at least 3 had placeholder-style author names not traceable to real papers. This gate prevents fabricated citations from reaching publication or client-facing materials.

Output Contract

In persistent mode, end every session with:

Ledger entry count (how many ingested)
Top 3 findings with source + confidence
One-line HUMMBL positioning implication per finding
Bus SITREP posted
Git commit of research artifacts — docs/research/ files MUST be committed to the repo before the session closes. Artifacts on disk + in the ledger are not sufficient; git history is the third durability layer. Stage only the docs/research/ files, not unrelated working-tree changes. (Origin: 2026-08-25 AAR — Levin grounding artifacts were left untracked in git after the session froze; a follow-up session had to commit them.)
Use a dedicated research worktree — create a clean worktree from origin/main before committing research artifacts. Do NOT commit to the shared symlinked worktree (~/PROJECTS/.agents or equivalent) because concurrent sessions mutate its branch state unpredictably. Pattern: git worktree add /work/active/<repo>-research-docs -b docs/research-<topic> origin/main. (Origin: 2026-09-02 AAR — first commit landed on a concurrent session's branch; second commit was blocked by merge conflicts from another session. Both required recovery via clean worktree creation.)
Verify branch ownership before commit — run git branch --show-current immediately before git add, not just during start-session. If the branch belongs to another agent's lane (e.g., feat/evaluation-framework-*, fix/devin-*), create a new branch or worktree instead. (Origin: 2026-09-02 AAR — branch was checked at session start but a concurrent session changed it before the commit.)
Git staging retry in autonomous loop harnesses — any autonomous loop script or harness modifying artifacts during multi-cycle execution MUST explicitly call git add <file> immediately after writing, and if pre-commit hooks reformat the file, it MUST re-stage and retry the commit once so intermediate changes are never silently dropped as "no changes". (Origin: 2026-09-04 AAR — 1-hour harness cycles produced unstaged changes because pre-commit hooks or staging was not retried.)
Distinguish artifact brain directory from repo paths — in agent environments, tools requiring ArtifactMetadata MUST target <appDataDir>/brain/<conversation-id>/. Repository documents in /work/active/ must be written directly via filesystem tools or copied from the brain artifact directory to avoid metadata path rejection. (Origin: 2026-09-04 AAR — write_to_file rejected repository path when ArtifactMetadata was passed.)
Git staging retry in autonomous loop harnesses — any autonomous loop script or harness modifying artifacts during multi-cycle execution MUST explicitly call git add <file> immediately after writing, and if pre-commit hooks reformat the file, it MUST re-stage and retry the commit once so intermediate changes are never silently dropped as "no changes". (Origin: 2026-09-04 AAR — 1-hour harness cycles produced unstaged changes because pre-commit hooks or staging was not retried.)
Distinguish artifact brain directory from repo paths — in agent environments, tools requiring ArtifactMetadata MUST target <appDataDir>/brain/<conversation-id>/. Repository documents in /work/active/ must be written directly via filesystem tools or copied from the brain artifact directory to avoid metadata path rejection. (Origin: 2026-09-04 AAR — write_to_file rejected repository path when ArtifactMetadata was passed.)

In preview mode, end with only the preflight result, searched sources and queries, and explicitly labeled unreviewed candidate URLs. Do not provide a findings synthesis, positioning implications, ledger count, research artifact, or commit claim.

Pre-flight Confabulation Probe (for remote-node research dispatches — remote-node dormant since 2026-07-01, use Workstation)

Before dispatching a multi-hour research run to local models, run 3 quick induction prompts in the target domain to calibrate the model's fabrication density:

Citation probe: "Cite the most influential 2023 paper on [topic]" — does it fabricate?
Statistic probe: "What percentage of enterprises have adopted [framework]?" — does it invent a number?
Entity probe: "Who is the CEO of [real company in the domain]?" — does it get it right?

If 2+ probes produce fabrications → route the lane to a more reliable model (gemma4:e4b was 100% reliable in Research Team v2; qwen3.5:9b had 3 empty responses).

This takes <2 minutes and prevents committing a 6-hour budget to an unreliable model.

Begin research now.

When to use [research-pipeline] instead

Use [research-pipeline] when you need:

Depth control (--depth quick|full|overnight)
BKI parallel track (--focus bki)
Audience-targeted dispatch (--audience ciso|caio|board|investor)
A structured pipeline receipt (stage counts, drop log, top findings)

[research] is for open-ended sessions. [research-pipeline] is for reproducible, documented runs.

Skill Chains

For delegate research execution to opencode -> [cross-runtime-bridge] (python ~/bin/cross_runtime_bridge.py delegate)
For free-tier inference for research synthesis tasks -> [reasoning-router] (route)

Mandatory

research-run-guard.py preflight before Sweep; the selected mode controls all later actions.

Advisory

After [research] → [research-digest] to verify coverage and compress findings
After [research] → [evidence-pack] if findings need bundling for external use
After [research] → [decision-log] if the session produced an adoption recommendation (ADR)
Before [research] → [brainstorm] to scope the research question
During [research] → use python ~/bin/free_apis.py for live scholarly/regulatory/vulnerability data (OpenAlex, PubMed, Federal Register, NVD, OSV, Wikidata)

Authority

T1 (TRUSTED): May run
T2 (Active/High): May run
T3 (Medium): Operator approval required
T4 (Probationary): BLOCKED (surge mode)
Operator: Override any restriction

research

Files1

SKILL.mdskills/research/SKILL.md