codex-academic

name	codex-academic
description	Delegate ACADEMIC tasks to OpenAI Codex CLI. Distinct from codex-router (engineering/debugging). Three modes — novelty audit, citation verify, claim-vs-code-vs-numbers — all conservative + citation-anchored. Used by /research-* commands and Hermes research-cycle cron.
version	1.0.0
author	Hermes Agent
license	MIT
metadata	{"hermes":{"tags":["routing","codex","academic","novelty","citation-verify"],"related_skills":["codex-router","gemini-academic"]}}

codex-academic

When to use: any task in the academic-research pipeline where Codex's conservative, citation-anchored, almost-never-wrong personality is load-bearing. Per memory feedback_model_routing.md academic-mode roster: this is the wrapper for /research-survey, /research-propose, /research-review.

Distinct from codex-router (engineering): codex-router defaults to gpt-5.3-codex for terminal-debug loops. codex-academic uses gpt-5.4 for review-style fact-anchored audit. Both wrappers can coexist.

Three modes

Mode 1 — novelty

Check whether a proposed reframe / cross-domain transfer is already published.

Inputs: reframe_text, target_domain, core_move Output contract (strict YAML):

status: published | near-match | no-match | cannot-verify
existing_work:
  - arxiv: <id>
    claim: <one-line>
    distinguishing_axis: <vs-the-reframe>
near_matches:
  - arxiv: <id>
    distinguishing_axis: <axis>

Mode 2 — citation-verify

Verify whether an arxiv-id resolves to a real paper with the claimed title/claim. Used to catch Gemini hallucination per feedback_model_routing.md § Gemini control discipline.

Inputs: arxiv_id, optional expected_title, optional expected_claim Output contract:

arxiv: <id>
match: yes | no | cannot-verify
title: <actual title if match>
claim: <actual 1-line claim if match>

Mode 3 — claim-vs-code-vs-numbers

Three-column audit: does the diff implement the claim, do the numbers support it.

Inputs: claim_text, diff, metric_value, metric_name, failure_condition, implicit_block Output contract:

checks:
  R1: pass | flag | block       # mechanism present, not just metric
  R2: pass | flag | block       # ablation present
  R6: pass | flag | block       # baseline failure mode named
  R7: pass | flag | block       # failure_condition concrete + testable
  explicit_vs_implicit: consistent | inconsistent
mismatches:
  claim_vs_code: [...]
  claim_vs_numbers: [...]
verdict: PASS | FLAG | BLOCK

Invocation recipe

Stdin discipline mandatory. Model defaults to gpt-5.4 (verified working as of 2026-04-27). Use gpt-5.4-mini for citation-verify only (fact lookup) to save tokens.

MODE="${1:?mode required: novelty|citation-verify|claim-vs-code-vs-numbers}"
PAYLOAD_FILE="${2:?payload file required}"
MODEL="${MODEL:-gpt-5.4}"
TIMEOUT="${TIMEOUT_SECONDS:-300}"

# Compose prompt per mode
case "$MODE" in
  novelty)
    PROMPT=$(cat <<'PROMPT_EOF'
Academic novelty audit (conservative, citation-anchored, no speculation).

Given: $(cat "$PAYLOAD_FILE")

Q1 — Has any peer-reviewed paper published this exact reframe? (published / near-match / no-match)
Q2 — If near-match: list arxiv-id + 1-line distinguishing axis between published work and this reframe.
Q3 — If no-match: list 3 most-related published works with arxiv-id + axis on which they differ.

Output STRICT YAML matching the contract in codex-academic § Mode 1. NO speculation; if unsure on citation existence, say "cannot-verify".
PROMPT_EOF
)
    ;;
  citation-verify)
    PROMPT=$(cat <<'PROMPT_EOF'
Citation verify (academic mode, conservative).

Given: $(cat "$PAYLOAD_FILE")

Q1 — Does the arxiv-id resolve to a real paper?
Q2 — If yes: what is the actual title and 1-line claim? Compare with expected_title / expected_claim if provided — flag mismatches.

Output STRICT YAML matching the contract in codex-academic § Mode 2. DO NOT speculate.
PROMPT_EOF
)
    # Use mini for cheaper citation lookup
    MODEL="${MODEL:-gpt-5.4-mini}"
    ;;
  claim-vs-code-vs-numbers)
    PROMPT=$(cat <<'PROMPT_EOF'
Academic correctness review (claim vs code vs numbers, conservative, citation-anchored).

Given: $(cat "$PAYLOAD_FILE")

Check rules from contracts/constitution.v0.1.0.yaml:
- R1 score_is_not_discovery — claim must include transferable mechanism
- R2 complexity_guilty — added components must have ablation
- R6 explain_why_w2s_fails — writeup must name baseline failure mode + correction
- R7 failure_condition_required — failure_condition must be concrete + testable

Also check Principle 3 cross-consistency: does the explicit reasoning contradict any silent_prior in the implicit block?

Output STRICT YAML matching the contract in codex-academic § Mode 3.
PROMPT_EOF
)
    ;;
  *)
    echo "FAILED: unknown mode '$MODE' — use novelty | citation-verify | claim-vs-code-vs-numbers"
    exit 2
    ;;
esac

timeout "$TIMEOUT" codex exec \
  --model "$MODEL" \
  --skip-git-repo-check \
  </dev/null \
  "$PROMPT"

Hard rules (per memory feedback_model_routing.md academic mode)

1. Stdin discipline: every invocation MUST </dev/null (otherwise codex hangs on "Reading additional input from stdin" — verified 2026-04-26).
2. No speculation: if Codex cannot verify a fact, output cannot-verify — do NOT fabricate. This is the load-bearing personality of this wrapper; if you ask it to speculate, you've defeated the purpose.
3. Citation-only mode: in citation-verify, NEVER ask Codex to "improve" or "rewrite" a citation — only verify it resolves.
4. Pair with gemini-academic: this wrapper is the verifier; gemini-academic is the diverger. Never use codex-academic for divergence (use codex-router or gemini-academic instead). Never use gemini-academic for citation-verify (always use codex-academic).
5. No GPT reviews GPT: per heterogeneous review chain, codex-academic outputs MUST be reviewed by a non-OpenAI model before action (Opus 4.7 or Gemini 3.1 Pro).

Output contract

Codex output ends with one of:

• The final YAML block per mode contract above
• RESULT: cannot-complete — <reason> if the mode cannot be honored

When to choose this over codex-router

Need	Pick
Code review on a PR	codex-router
Run tests + fix iteratively	codex-router
Verify whether arxiv:2505.13447 is real	codex-academic (citation-verify)
Check if reframe X is already published	codex-academic (novelty)
Audit claim-vs-code in research diff	codex-academic (claim-vs-code-vs-numbers)
Debug "why does train.py crash"	codex-router

Anti-patterns

• ❌ Do NOT use codex-academic on engineering tasks (dilutes the academic personality)
• ❌ Do NOT remove </dev/null — same hang issue
• ❌ Do NOT ask Codex to speculate — that's gemini-academic's job
• ❌ Do NOT chain codex-academic → codex-academic for review (no GPT reviews GPT)