Final Integrity Stage (4.5) Explained: How It Differs from Pre-Review Integrity (Stage 2.5)

Stage 4.5 is a mandatory post-revision integrity gate that executes a full RAISE compliance audit before final publication, whereas Stage 2.5 is a preliminary checkpoint that verifies basic integrity immediately after the initial draft is produced.

The Imbad0202/academic-research-skills repository defines a 10-stage academic pipeline that inserts two mandatory integrity gates to enforce responsible AI usage. While both Stage 2.5 and Stage 4.5 verify compliance with RAISE principles, they serve distinct purposes at different points in the manuscript lifecycle.

What Is the Final Integrity Stage (Stage 4.5)?

Stage 4.5 is the final integrity gate that runs after the author has incorporated reviewer feedback (Stage 4) and immediately before the manuscript is formatted and released (Stage 5). According to the orchestrator specification in academic-pipeline/agents/pipeline_orchestrator_agent.md, this stage is explicitly listed as non-skippable: "Non-Skippable: … Stage 4.5 (final integrity)"【pipeline_orchestrator_agent.md#L33-L38】.

Timing and Scope

Unlike earlier checkpoints, Stage 4.5 performs a full re-run of the RAISE checklist on the revised draft. This includes validating human-oversight, transparency, reproducibility, and fit-for-purpose requirements. Additionally, it audits any new references introduced during the revision period to ensure they meet citation verification standards.

The shared/handoff_schemas.md file reinforces that this gate cannot be bypassed: "Passport does not grant Stage 4.5 skip… Stage 4.5 always requires full Mode 2 verification"【handoff_schemas.md#L770-L777】.

What Is the Pre-Review Integrity Stage (Stage 2.5)?

Stage 2.5 serves as the pre-review integrity gate, executing immediately after the paper draft is produced (Stage 2) and before the first peer-review round (Stage 3). Its primary function is to catch glaring AI-generated fabrications, missing citations, and methodological transparency gaps before the manuscript reaches reviewers.

This stage focuses on basic compliance with RAISE principles and low-level citation verification. Like Stage 4.5, it is strictly non-skippable—the orchestrator marks it as a mandatory checkpoint that must pass before the pipeline can advance to Stage 3.

Key Differences Between Stage 4.5 and Stage 2.5

While both gates enforce integrity, they differ in timing, scope, and failure handling:

• Timing: Stage 2.5 runs once after the initial draft is complete, while Stage 4.5 runs after revisions and may repeat if the check fails.
• Scope of Checks: Stage 2.5 examines the initial manuscript using a subset of RAISE items (e.g., reviewer-count, basic tool listing). Stage 4.5 executes the full RAISE checklist plus an additional audit-artifact gate for new references.
• Failure Routing: A Stage 2.5 failure triggers a mandatory return to Stage 2 (writing) with integrity violations. A Stage 4.5 failure sends the author back to Stage 5 (revision) with a "final integrity issues" report; if the second check also fails, the pipeline aborts.
• Passport Interaction: Neither stage can be skipped via a Material Passport, but Stage 4.5 is explicitly forbidden from bypass regardless of passport status, whereas Stage 2.5 at least consults the passport for freshness.

Technical Implementation and Source Locations

The integrity gates are hardcoded into the pipeline specification. The test suite in scripts/check_spec_consistency.py validates that the pipeline contains the string "final integrity check (Stage 4.5)"【check_spec_consistency.py#L75-L78】.

Design documentation in docs/design/2026-04-30-ars-v3.6.7-step-6-orchestrator-hooks-spec.md explicitly calls Stage 4.5 a "mandatory integrity gate" and clarifies that it cannot be skipped【2026-04-30-ars-v3.6.7-step-6-orchestrator-hooks-spec.md#L35-L42】.

User-facing examples in examples/showcase/README.md demonstrate a successful run: "Post-Revision (Stage 4.5) – PASS"【examples/showcase/README.md#L18-L20】.

How to Invoke the Integrity Stages

When running the full pipeline, the orchestrator automatically executes both gates in sequence:

# Run the full pipeline – automatically triggers Stage 2.5, then Stage 4.5 later

ars-full

For custom scripts or manual verification, import the integrity checker directly:

from scripts.check_v3_6_8_frontmatter_trust_schema import run_integrity_check

# Pre-review check (Stage 2.5) – run after initial draft

run_integrity_check(stage=2.5, artifact_path="draft/paper.md")

# Final check (Stage 4.5) – run after incorporating reviewer feedback

run_integrity_check(stage=4.5, artifact_path="draft/revised_paper.md")

Both calls raise a FAIL exception that forces the orchestrator to route the user back to the appropriate stage for correction.

Why the Distinction Matters

The dual-gate architecture serves three critical functions:

• Quality Assurance: Stage 2.5 protects the reviewer pool from evaluating papers containing fabricated or undisclosed AI-generated content, ensuring peer review resources are spent on valid submissions.
• Regulatory Compliance: Stage 4.5 guarantees that any changes made during revision—such as added AI-generated analysis—are disclosed in the final AI-use disclosure, satisfying journal requirements that mandate final-stage transparency.
• Audit Trail: Both stages generate an integrity_report artifact incorporated into the AI Self-Reflection Report at Stage 6, with the Stage 4.5 report providing definitive compliance evidence for publication.

Summary

• Stage 4.5 (Final Integrity) runs after revisions and before publication, performing a full RAISE audit that cannot be skipped even with a Material Passport.
• Stage 2.5 (Pre-Review Integrity) runs after the initial draft, catching basic violations before peer review begins.
• Key differences include timing (post-draft vs. post-revision), scope (basic checks vs. full audit), and failure handling (return to writing vs. return to revision).
• Implementation is enforced in pipeline_orchestrator_agent.md and handoff_schemas.md, with validation in check_spec_consistency.py.

Frequently Asked Questions

Can Stage 4.5 be skipped if I have a verified Material Passport?

No. According to shared/handoff_schemas.md, the passport explicitly does not grant Stage 4.5 skip privileges. The specification states that "Stage 4.5 always requires full Mode 2 verification" regardless of passport status【handoff_schemas.md#L770-L777】. This ensures that no manuscript reaches publication without a final integrity validation.

What happens if the Final Integrity check fails after revisions?

If Stage 4.5 fails, the orchestrator returns the manuscript to Stage 5 (revision) with a detailed "final integrity issues" report. The author must address these issues and resubmit. If the second final check also fails, the pipeline aborts entirely, preventing the manuscript from proceeding to formatting and release.

How does Stage 2.5 protect reviewers from AI-generated content?

Stage 2.5 acts as a filter that catches glaring AI-generated fabrications, missing citations, and methodological transparency gaps before the manuscript enters the peer-review pool (Stage 3). By enforcing basic RAISE compliance early, it ensures reviewers evaluate only papers that meet minimum integrity standards, protecting their time from being wasted on irreparable submissions.

Where are the integrity reports stored after Stage 4.5?

Both Stage 2.5 and Stage 4.5 generate an integrity_report artifact that is incorporated into the AI Self-Reflection Report generated at Stage 6. The Stage 4.5 report specifically serves as the definitive compliance evidence for publication, documenting that the final manuscript met all RAISE principles and citation verification requirements before release.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}