Rodin
78 lines
5.3 KiB
Markdown
78 lines
5.3 KiB
Markdown
# Finding #61: Regulatory Completeness Analysis
|
|
|
|
**Date:** 2026-05-09
|
|
**Lens:** Regulatory completeness analysis — does the document correctly implement referenced regulatory requirements?
|
|
**Document:** gargoyle wash-sale-tracking.md (159 lines)
|
|
|
|
## Task
|
|
|
|
Analyze wash sale tracking design document for regulatory compliance gaps — does the implementation correctly capture all requirements from IRC §1091, Treasury Regulations §1.1091-1, and IRS Publication 550?
|
|
|
|
## Results
|
|
|
|
| Model | Output tokens | Reasoning tokens | Findings |
|
|
|---|---|---|---|
|
|
| GPT-5 | 10,460 | 8,832 | 10 |
|
|
| Claude Opus 4.5 | 2,227 | (internal) | 11 |
|
|
| Claude Sonnet 4 | 1,276 | (internal) | 10 |
|
|
|
|
## Common Ground (all 3 identified)
|
|
|
|
- Cross-account/IRA wash sale detection missing (CRITICAL) — IRC §1091 applies across all taxpayer accounts; IRA losses are **permanently** disallowed, not deferred
|
|
- Substantially identical definition too narrow — misses ADRs vs ordinary shares, preferred vs common, etc.
|
|
- Trade date vs settlement date ambiguity — IRS uses trade date for 61-day window
|
|
- Multiple replacement lots allocation unclear — no FIFO or ordering rule specified
|
|
- 1099-B reconciliation requirements missing — broker vs platform calculations may differ
|
|
|
|
## GPT-5 Unique Findings
|
|
|
|
- **Pairwise detection over-disallowance:** Share-level ledger needed; current pairwise model can disallow same loss multiple times across multiple replacement lots
|
|
- **Lot-level vs share-level adjustments:** Basis and holding period adjustments described at lot level, not share level; partial overlap would incorrectly adjust entire replacement lot
|
|
- **Corporate action false positives:** Splits/dividends creating "new lots" would trigger false wash sales (these aren't purchases under IRC §1091)
|
|
- **Short sale wash sale rules:** Window measured differently for shorts (30 days before short sale through 30 days after close); document silent on shorts
|
|
- **Pre-sale/post-sale allocation determinism:** No deterministic rule for allocating disallowed loss when both pre-sale and post-sale purchases exist
|
|
|
|
## Opus Unique Findings
|
|
|
|
- **IRA loss permanently lost:** Uniquely emphasized that IRA wash sales don't just defer losses — the loss is **permanently unrecoverable** (can't add to IRA basis). Document's unconditional basis adjustment would mislead users.
|
|
- **Option exercise/assignment as purchases:** Treasury Reg §1.1091-1(a) explicitly includes "contract or option" acquisitions; exercising a call or being assigned on a short put is a purchase for wash sale purposes
|
|
- **Merger continuity:** Company A → Company B reorganization may leave A and B "substantially identical"; stable instrument identifier approach may miss this
|
|
- **Chained wash sales:** Replacement lot sold at a loss triggering another wash sale with third lot — holding period chains through multiple replacements
|
|
|
|
## Sonnet Unique Findings
|
|
|
|
- **Constructive ownership rules:** Treasury Reg §1.1091-1(c) covers related parties (spouse, controlled corps, partnerships) — missing from design
|
|
- **Stock rights and dividends:** Treasury Reg §1.1091-1(g) addresses how these affect "substantially identical" and basis calculations
|
|
- **Dealer exception:** Treasury Reg §1.1091-1(b) exempts securities dealers — out of scope but not documented
|
|
|
|
## Key Insights
|
|
|
|
### GPT-5's Exhaustive Regulatory Enumeration
|
|
GPT-5's 8,832 reasoning tokens enabled systematic cross-referencing between document claims and IRC/Treasury Reg sections. It explicitly enumerated pairwise allocation edge cases that would cause mathematical over-disallowance. This is the same exhaustive enumeration pattern seen in Finding #58 (state machine completeness) and Finding #59 (convention rule gaps).
|
|
|
|
### Opus Traces Regulatory Gaps to Operational Consequences
|
|
Opus uniquely emphasized the **permanent** nature of IRA wash sale loss disallowance and traced option exercise scenarios through to incorrect tax reporting. Consistent with previous findings, Opus finds what the document can't see about itself — the gap between "deferral" language and the IRA case where deferral doesn't apply.
|
|
|
|
### Sonnet Finds Structural Regulatory Categories
|
|
Sonnet uniquely identified constructive ownership rules (Treasury Reg §1.1091-1(c)) as a missing category — a structural gap in the regulatory coverage. However, Sonnet didn't trace this to specific failure modes like the other models.
|
|
|
|
## Task Taxonomy Update
|
|
|
|
**Regulatory completeness analysis** → GPT-5 for exhaustive IRC/CFR cross-referencing, Opus for tracing gaps to operational consequences, Sonnet for structural category identification
|
|
|
|
This lens is distinct from other document analysis types:
|
|
- **State machine completeness** (#58) tests transition coverage
|
|
- **Convention rule gaps** (#59) tests specification consistency
|
|
- **Event ordering** (#60) tests temporal failure modes
|
|
- **Regulatory completeness** tests legal/regulatory implementation correctness
|
|
|
|
## Practical Implication
|
|
|
|
For regulatory compliance analysis of financial systems, use GPT-5 for exhaustive regulation cross-referencing, then Opus to trace gaps to operational/legal consequences. Sonnet provides efficient structural overview but insufficient depth for compliance work.
|
|
|
|
## Efficiency
|
|
|
|
- GPT-5: 1,046 tokens/finding (verbose but exhaustive)
|
|
- Opus: 202 tokens/finding (good detail-to-length ratio)
|
|
- Sonnet: 128 tokens/finding (efficient but surface-level)
|