gemini-academic

name	gemini-academic
description	Delegate ACADEMIC tasks to Gemini CLI. Distinct from gemini-router (engineering/long-context). Three modes — cross-domain mining (Flash), challenger feasibility audit (Pro), arxiv probe (Pro). All challenger-framed; never agreement-seeking. Always paired with codex-academic citation-verify.
version	1.0.0
author	Hermes Agent
license	MIT
metadata	{"hermes":{"tags":["routing","gemini","academic","challenger","cross-domain"],"related_skills":["gemini-router","codex-academic"]}}

gemini-academic

When to use: tasks that need Gemini's strongest academic database + divergent thinking — but with hallucination + sycophancy controls. Per memory feedback_model_routing.md academic-mode roster: Gemini 3.1 Pro has the highest academic ability AND the highest hallucination rate AND tends to please. This wrapper bakes in the controls.

Distinct from gemini-router (engineering): gemini-router is for long-context code / translation review. gemini-academic is for cross-domain mining + challenger feasibility — every prompt is challenger-framed, every emitted citation gets Codex-verified.

Hard control discipline (per memory § Gemini control discipline)

These are not optional — every mode below applies them:

1. Always challenger-framed, never supporter-framed. Prompt must contain "challenge", "find what is wrong/risky/missing", "DO NOT agree-to-please". Never "do you agree", "is this right", "support this".
2. Demand machine-checkable evidence. Every claim cites arxiv-id (or DOI). Soft references like "the literature shows" → reject the output.
3. Always pair with codex-academic citation-verify. Caller MUST run codex-academic Mode 2 on every arxiv-id Gemini emits. Hallucination rate >30% → halt.
4. Never the terminal artifact. Gemini's output is intermediate; an Opus 4.7 review must follow before any decision is made.
5. Stdin discipline: every gemini -p call MUST </dev/null.

Three modes

Mode 1 — cross-domain-mining

Find K breakthrough papers from non-adjacent fields whose CORE METHODOLOGICAL MOVE could plausibly transfer to a target domain. Used by /research-survey Stage 1.

Engine: gemini-3-flash-preview (bulk academic search). Flash is enough — Pro's depth isn't worth the hallucination risk on bulk paper discovery.

Inputs: target_domain, K (default 3) Output contract (strict YAML):

candidates:
  - arxiv: <id>
    title: <verbatim>
    field: <not target_domain>
    core_move: <one sentence, philosophical not technical>
    axiom_attacked: <which default axiom of target_domain>
  - arxiv: ...
    ...

Mode 2 — challenger-feasibility

Probe a proposed reframe for hidden assumptions, infeasibility paths, ablation traps, and contradicting prior work. Used by /research-propose Stage 3.

Engine: gemini-3.1-pro-preview with --approval-mode plan (read-only). Pro's depth is needed to find non-trivial assumption holes; the hallucination is controlled by codex-academic citation-verify on the output.

Inputs: reframe_text, testable_prediction, measurable_contribution Output contract:

hidden_assumptions:
  - <load-bearing assumption inside the reframe — 3-5 items>
infeasibility_paths:
  - <concrete way it fails — engineering / theoretical / empirical — 3 items>
ablation_traps:
  - <ablation that would distinguish reframe's mechanism from naive baseline>
required_citations:
  - arxiv: <id>           # Codex will verify this
    contradiction: <one-line>

Mode 3 — arxiv-probe

Same as challenger-feasibility but applied to an existing kept result (post-experiment). Used by /research-review Stage 6 Pass 2.

Engine: gemini-3.1-pro-preview --approval-mode plan.

Inputs: diff, claim, metric_value, niche, failure_condition, implicit_block Output contract:

R5_transfer:
  status: pass | flag
  missing_transfers:
    - dataset: ...
      arxiv: ...
R9_compression:
  status: pass | flag
  subsumed_or_not: <which prior work this subsumes / fails to subsume>
  reference_arxiv: <id>
R10_suspect_score:
  status: pass | flag
  shortcut_hypotheses:
    - hypothesis: ...
      empirical_signature: ...
      arxiv_support: <id>
verdict: PASS | FLAG | BLOCK

Invocation recipe

MODE="${1:?mode required: cross-domain-mining|challenger-feasibility|arxiv-probe}"
PAYLOAD_FILE="${2:?payload file required}"
TIMEOUT="${TIMEOUT_SECONDS:-600}"

case "$MODE" in
  cross-domain-mining)
    MODEL="gemini-3-flash-preview"
    APPROVAL="default"
    PROMPT_HEADER="You are an academic adversary, not a recommender. Find K=$(awk '/^K:/{print $2; exit}' $PAYLOAD_FILE || echo 3) breakthrough papers from fields ADJACENT BUT DIFFERENT from the target domain."
    ;;
  challenger-feasibility)
    MODEL="gemini-3.1-pro-preview"
    APPROVAL="plan"
    PROMPT_HEADER="You are an academic adversary. CHALLENGE this reframe — DO NOT support it. Find what is wrong / risky / missing."
    ;;
  arxiv-probe)
    MODEL="gemini-3.1-pro-preview"
    APPROVAL="plan"
    PROMPT_HEADER="You are an academic adversary, not a supporter. CHALLENGE this kept result. Probe R5/R9/R10 from constitution.v0.1.0.yaml."
    ;;
  *)
    echo "FAILED: unknown mode '$MODE'"
    exit 2
    ;;
esac

PROMPT="$PROMPT_HEADER

$(cat "$PAYLOAD_FILE")

Hard rules:
- DO NOT agree-to-please. Your job is to find what is wrong / risky / missing.
- DO NOT generalize ('the literature shows' ← rejected). Cite arxiv-ids only — Codex will verify; hallucinated IDs will be dropped.
- Output STRICT YAML matching the contract in gemini-academic § Mode <N>. No prose outside.
"

timeout "$TIMEOUT" gemini --yolo \
  -m "$MODEL" \
  --approval-mode "$APPROVAL" \
  -p "$PROMPT" \
  </dev/null

Output contract

Each mode emits its specific YAML (see Mode 1/2/3 above). The CALLER then MUST:

1. Extract every arxiv: id from the output
2. Pipe each through codex-academic citation-verify
3. Drop any output whose Gemini-cited arxiv-id failed Codex verify
4. If hallucination rate >30%, halt — Gemini is in a bad state

When to choose this over gemini-router

Need	Pick
Code review on 200k-token codebase	gemini-router
Cross-check translation fluency	gemini-router
Find papers from biology that could transfer to ML	gemini-academic (cross-domain-mining)
Probe reframe X for hidden assumptions	gemini-academic (challenger-feasibility)
Adversarial review of kept experiment	gemini-academic (arxiv-probe)

Anti-patterns

• ❌ DO NOT use gemini-academic without paired codex-academic citation-verify (hallucination escapes)
• ❌ DO NOT use agreement-framed prompts ("do you agree", "support this") — defeats the wrapper purpose
• ❌ DO NOT use gemini-academic output as the final decision artifact — Opus 4.7 must review
• ❌ DO NOT remove </dev/null (cron stdin hang)
• ❌ DO NOT use gemini-3.1-pro without -preview suffix (returns 404 per gemini-router note)