| # FilingsBench Controls v0.1 candidate report
|
|
|
| Run date: 2026-07-27. This report covers a narrative-controls candidate slice, not the original
|
| 60-filing financial-facts benchmark.
|
|
|
| ## Result
|
|
|
| The deterministic baseline matched all 48 candidate field outcomes across 12 issuer-disjoint SEC
|
| filings. This includes six expected abstentions: the two 10-Qs do not state annual ICFR conclusions
|
| or support material-weakness status, but both do state disclosure-controls effectiveness.
|
|
|
| | Field | Records | Present labels | Abstained labels | Exact outcomes | Evidence on present predictions |
|
| | --- | ---: | ---: | ---: | ---: | ---: |
|
| | ICFR effectiveness | 12 | 10 | 2 | 12/12 | 10/10 |
|
| | Disclosure-controls effectiveness | 12 | 12 | 0 | 12/12 | 12/12 |
|
| | Material weakness disclosed | 12 | 10 | 2 | 12/12 | 10/10 |
|
| | Remediation status | 12 | 10 | 2 | 12/12 | 10/10 |
|
|
|
| Median parser latency was 2,292.04 ms; mean was 2,766.30 ms; maximum was 4,635.59 ms. Median peak
|
| Python memory was 24.32 MB and maximum was 74.30 MB. Inference cost was $0. Network transfer is not
|
| included in per-document parser latency. The observed end-to-end command took 34.6 seconds on the
|
| audit machine.
|
|
|
| ## Dataset and method
|
|
|
| The slice contains ten 10-Ks and two 10-Qs from 12 issuers, split 8/2/2 with no issuer overlap.
|
| Annual records balance five effective and five ineffective ICFR conclusions. The deterministic
|
| system retrieves Item 9A or Part I, Item 4, distinguishes ICFR from disclosure controls, returns
|
| exact normalized-document excerpts and offsets, and abstains when the section does not support a
|
| value.
|
|
|
| The consistency inference from an effective ICFR conclusion to no current material weakness is
|
| explicit in prediction provenance. It follows SEC staff guidance that management may not conclude
|
| ICFR is effective when one or more material weaknesses exist. SEC staff also says an unremediated
|
| material weakness requires an ineffective conclusion and encourages disclosure of its nature,
|
| impact, and remediation plans. Sources: [SEC ICFR FAQ](https://www.sec.gov/oca/controlfaq1004),
|
| [SEC staff statement](https://www.sec.gov/info/accountants/stafficreporting.htm), and
|
| [SEC Financial Reporting Manual](https://www.sec.gov/about/divisions-offices/division-corporation-finance/financial-reporting-manual/frm-topic-4).
|
|
|
| ## Reproducibility
|
|
|
| - dataset SHA-256: `9de0c23ed5e971d9426b7d235b52332fb13c257829d4a09743641eef213717de`
|
| - predictions SHA-256: `caeceffb0e2a2918e6db18dc4ccd4d1c6ecd33bcee1f1dbee811fc5c417a7a7d`
|
| - summary SHA-256: `e6ea9f86da89414fb7a01aada3cee717505829f9f454130030b4c9371db0436f`
|
| - Python 3.11.15; Windows 11 Home 10.0.26200
|
| - Intel Core i5-1135G7, 4 cores / 8 logical processors; 16.95 GB physical RAM
|
|
|
| Commands:
|
|
|
| ```bash
|
| finstruct controls-dataset build
|
| finstruct controls-benchmark run data/filingsbench_controls_v0.1/records.jsonl \
|
| --output results/controls-v0.1.0
|
| ```
|
|
|
| ## Limitations and model gate
|
|
|
| Every record was deliberately selected to exercise a known label state, so this is not a prevalence
|
| sample. The labels were manually read from filed primary HTML but have not yet received independent
|
| double annotation or adjudication; they remain `candidate_pending_independent_review`, not gold.
|
| The slice is small, English-only, and excludes 8-Ks, amendments, foreign-issuer forms, OCR/PDF
|
| filings, management/auditor disagreements, and multi-weakness summary generation.
|
|
|
| Because the deterministic baseline has no residual errors on this candidate slice, an open-model or
|
| A100 experiment is not justified yet. The gate reopens after independent review and a larger blind
|
| adversarial set reveal a measured error surface that a model could improve.
|
|
|
