Verasic Deep Research

Security: see references/scanner-notes.md and upstream SECURITY.md for expected scanner signals and trust model.

Verasic Deep Research — Ledger-Backed Research

Source of truth

Never duplicate the protocol in chat — follow it.

Workflow

1. Help path — if the user invokes deep research with no question or asks for help, relay references/helper.md (adjust path prefix for install root).
2. Pre-flight gate — require all pre-flight fields before any fetch or spawn: question, depth, output, source-boundary, languages. No defaults on output or source-boundary (recommend public-standard only after asking). Collect output path when output includes files. Apply honesty notices per protocol (see references/helper.md ## honesty).
3. Read protocol — read references/research-protocol.md, active workflows/<depth>.md, and supporting references before any fetch or spawn.
4. Dispatch T2/T3 per tier — T1 main orchestrates; spawn T2 workers per depth preset via Task. T3 leaf jobs (fetch-url, extract-excerpt, single-query-search, verify-one-claim) are spawned by T2 (grandchild, no Task). T1 batches WebFetch when T2 unavailable. standard-research: Hunter + Practitioner parallel → Skeptic sequential (mandatory). T2 direct fallback when T3 fails — not a blocker.
5. Verify ledger — every citation must pass verify-before-cite; no cite without a ledger row (see references/citation-protocol.md).
6. Score — apply 5-axis confidence per references/confidence-rubric.md; sensitive domains enforce floor 60.
7. Drill — per references/drill-protocol.md when thresholds hit; max 2 rounds (auto-at-threshold auto-executes round 1; offers round 2).
8. Deliver — chat per output format; write files to ./docs/research/<slug>/ when chosen. Use templates/deep-research-brief.md sections. Ask mode = no file writes.

Orchestration (Cursor)

Read .cursor/skills/verasic-deep-research/references/research-protocol.md first and follow it exactly.

Help

If the user sent no question, only help, or an empty invocation — relay .cursor/skills/verasic-deep-research/references/helper.md verbatim (adjust the path prefix if this skill is installed elsewhere).

Required from the user's message

Parse or ask for:

• depth: one of quick-scan, standard-research, adversarial-deep, custom
• output: one of chat-only, chat+files, files-only, custom — no default
• source-boundary: one of public-standard, public-extended, aggressive-scrape (+ optional free text)
• languages: cite / search / report languages
• The question body

Optional: models:, domain:, drill: (auto-at-threshold | off | always-offer), claims: list, output path, custom-roles: when depth: custom

No default output format. No default source boundary. Recommend public-standard. If required fields missing, ask before any fetch or spawn. Include honesty notices from helper ## honesty at each pre-flight step.

Ask mode

If the user is in Ask (read-only) mode — no file writes. Deliver chat only even if they chose chat+files or files-only. Say so and offer Agent mode or chat-only.

Steps

1. Pre-flight per protocol — all required fields; path ./docs/research/<slug>/ when files requested.
2. Read supporting refs: citation-protocol, confidence-rubric, source-tiers, drill-protocol, workflows/<depth>.md, deliver template.
3. Dispatch T2 workers via Task per depth tier:
   • quick-scan: Hunter only (no Skeptic)
   • standard-research: Hunter + Practitioner parallel → Skeptic sequential (mandatory 7a)
   • adversarial-deep: Hunter + Practitioner + Skeptic + Arbiter (4 parallel)
4. T3 leaf jobs spawned by T2 (grandchild, no Task): fetch-url, extract-excerpt, single-query-search, verify-one-claim. T1 batches WebFetch when T2 unavailable. T2 direct fallback on failure.
5. Verify ledger — verify-before-cite; no [Sn] without ledger row; IEEE citations; snippet-only headline hard cap 40.
6. Score — 5-axis per claim (SQ, EC, CG, CO, VR); sensitive domain floor 60; chat shows headline + full axes.
7. Drill when thresholds hit — auto-at-threshold auto-executes round 1; offer round 2; max 2 rounds (drill-protocol.md).
8. Deliver per output — use templates/deep-research-brief.md sections.
9. ## unverified → suggest manual fusion only if verasic-fusion installed (fusion-handoff.md).
10. If Task unavailable — ask before degraded sequential single-context research.

T2 subagent task prompt (one per role)

Replace <skill-root> with the absolute skill directory (contains references/research-protocol.md).

Readonly deep-research T2 (<role>). Follow <skill-root>/references/research-protocol.md.

Supporting refs (readonly):
- <skill-root>/references/citation-protocol.md
- <skill-root>/references/source-tiers.md

Role: <Hunter|Practitioner|Skeptic|Arbiter>
Source boundary: <boundary>
Languages: <languages>
Domain pack: <inferred or user domain>

## Packaged prompt
<question + pre-flight context>

Return candidate sources for main-agent verify. Do not mutate the repo.

T3 leaf jobs (spawned by T2, no Task)

T3 must not synthesize, merge ledger, spawn subagents, or score confidence.

Deliver

• chat-only — full brief sections in chat; headline + per-claim 5-axis in ## confidence.
• chat+files — chat summary + write ./docs/research/<slug>/deep-research-brief.md and ledger files in the same turn (Agent mode). Do not end with chat-only and write files later unless the user explicitly asks for staged delivery.
• files-only — write files; chat gets short pointer only.

Hard rules: verify-before-cite; no cite without ledger row; refuse insider/illegal sources.

Recommend composer-2.5-fast in optional model rosters when suggesting models.

Without subagents

Read references/research-protocol.md and execute the full pipeline yourself, including degraded confirmation when Task is unavailable.

Hard rules

• Verify-before-cite — no citation without a verified ledger row.
• No default output format — user must pick chat-only, chat+files, files-only, or custom.
• No default source boundary — user must pick; agent recommends public-standard.
• Ask mode = no file writes — deliver in chat only.
• composer-2.5-fast should appear in suggested model rosters when optional roster is offered.
• Fusion handoff — suggest manually on ## unverified only if verasic-fusion is installed; never assume it is present.

Refuse

Insider or non-public proprietary data, illegal collection without an explicit user-stated boundary, and research that requires violating the chosen source boundary.