Model Card: Accounting Conservation Framework

Version: 0.8.0 (Research Preview) Date: February 2026 Maintainer: Nirvan Chitnis (nirvanchitnis@gmail.com)

---

Model Details

• Objective: Quantify accounting conservation by enforcing algebraic consistency across financial statement line items.
• Current Scope: Simplified 15-constraint system spanning 10 core equity and cash-flow tags. The production roadmap extends to the full 54-constraint / 287-tag specification once enriched filings are available.
• Core Components:
  • Constraint matrix generator (src/core/constraint_matrix.py)
  • Linear feasibility gap solver (src/core/feasibility_gap.py)
  • Hierarchical calibration & false discovery correction (src/statistical/*)
  • Temporal drift diagnostics (scripts/temporal_analysis.py)
  • Nullspace and adversarial testing toolchain (src/security/*)
• Key Outputs:
  • Pass/fail indicators for each constraint
  • δ* feasibility gap scores (scaled 0–1)
  • Statistical calibration reports (ICC, FDR-adjusted q-values)
  • Temporal drift plots and Bai–Perron breakpoints
  • Gameability assessments (nullspace dimensionality, adversarial traces)

---

Intended Use

Stakeholder	Primary Decision Support
Audit / Assurance	Flag filings with conservation violations for targeted review
Data Quality Teams	Monitor XBRL tagging regressions across filing cohorts
Academic Researchers	Study structural drivers of accounting misstatements

Non-Goals / Out-of-Scope - Automated fraud adjudication or enforcement actions - Investment recommendations or credit scoring - Real-time ingestion of unaudited / streaming disclosures

---

Training & Evaluation Data

• Data Source: Processed SEC XBRL filings (results/disaggregates/filings_real.csv, 5,533 filings, 2012–2025).
• Gold Labels: Placeholder schema for 300 manually reviewed filings (results/eval/gold_labeled_300.csv). Manual labeling workflow pending completion.
• Evaluation Metrics: Precision / recall / F1 across materiality tiers (0.1%, 0.5%, 1%), Cohen’s κ for inter-rater agreement, temporal pass-rate trends, feasibility gap distribution moments.
• Data Quality Notes: Current dataset contains only 10 of the 54 target tags. Assets/liabilities coverage is incomplete, leading to conservative (high δ*) feasibility scores and 0% pass rates under the simplified model.

---

Ethical Considerations

• Misinterpretation Risk: Constraint violations frequently stem from tagging defects rather than financial misconduct. All flags require human adjudication.
• Reputational Harm: Results are intended for internal audit usage only; public disclosure without context may stigmatize issuers.
• Gaming Risk: The simplified constraint matrix currently has zero-dimensional nullspace (no adversarial freedom). When expanded to 287 tags, nullspace dimensionality is expected to grow; mitigation strategies include dynamic constraint enrichment, anomaly scoring on residual components, and adversarial training.

---

Limitations & Known Issues

1. Constraint Coverage: Only 15 rules implemented; does not capture the full accounting ontology yet.
2. Sparse Tags: Missing balance-sheet tags inflate δ* scores and limit rank diagnostics.
3. Solver Dependencies: CVXPY falls back to SCS when ECOS is unavailable, increasing solve time (~40 ms vs <10 ms target).
4. Statistical Calibration: Severe miscalibration observed (χ²/df ≈ 1.9×10⁵) due to simplified tag universe; expect improvements once additional constraints activate.
5. Temporal Conclusions: Pass-rate improvements cited in legacy documentation (5.1% → 72.4%) are not reproducible with the current data subset.

---

Monitoring & Maintenance

• Operational Metrics: Precision/recall on gold labels (quarterly), pass-rate trend drift, feasibility gap quantiles, constraint rank stability.
• Alert Thresholds: Trigger full review if precision <0.70, average δ*_scaled >0.20, or nullspace dimension increases unexpectedly.
• Release Process: Each material change must:
  1. Pass automated regression tests (pytest suite and REPRODUCE.sh).
  2. Regenerate all artifacts via bash REPRODUCE.sh.
  3. Produce a reproducibility report with tolerance ≤1e-6.
  4. Undergo architect review prior to deployment.
• Version Tracking: Git commit SHA and model version are logged in output metadata (Append to results/REPRODUCIBILITY_REPORT.txt).

---

Change Control & Governance

• Approval Matrix:
  • Minor updates (bug fixes, documentation): Maintainer approval.
  • Major updates (new constraints, optimization logic): Architect review + accounting subject-matter expert.
  • Data changes (new filings ingestion, label refresh): Dual approval (maintainer + auditor lead).
• Release Artifacts: Archive REPRODUCE.sh, MODEL_CARD.md, requirements-freeze.txt, and docs/reproducibility/REPLICATION_GUIDE.md with each tagged release for Zenodo/OSF deposition.
• Gap Remediation Plan: Document mitigation steps for each limitation; open issues tracked in repository roadmap (docs/reproducibility/REPLICATION_GUIDE.md#troubleshooting).

---

References

• Benjamini, Y., & Hochberg, Y. (1995). Controlling the False Discovery Rate.
• Cameron, A. C., Gelbach, J. B., & Miller, D. L. (2008). Bootstrap-Based Improvements for Inference with Clustered Errors.
• Bai, J., & Perron, P. (2003). Computation and Analysis of Multiple Structural Change Models.
• SEC EDGAR US GAAP Taxonomy (2020–2025 editions).

---

Contact & Support

• Primary Contact: Nirvan Chitnis (nirvanchitnis@gmail.com)
• Security Disclosures: security@nchitnis-audit.com
• License Summary: Code (Proprietary), Data (Research Use Only), Documentation (CC BY 4.0). Refer to repository root LICENSE files for full terms.