# codex gpt-5.5 on C02 (self-deal identity-collision) — the LLM-blind cell, NOW CAUGHT

2026-06-27. Cell: `S9-xfile-c02-self-deal-commission` (oracle detector, band-3, cross-file). Fix `c2d8b4195`.
Canonical: a BUYER (referee) who owns the deal's vendor self-deals an affiliate commission — `accrueAffiliateCommission`
guards only `isSelfVendorPurchase(referrer,owner)`, NOT `(referee,owner)`. Oracle flag = `guards {owner|referrer} but NOT {owner|referee}`.

## Why this cell matters
Recorded baseline (deployed-gate `VALIDATION.md:137`): **EVERY unseeded Anthropic LLM config MISSED C02** —
Sonnet MED/HIGH/XHIGH, Opus MED/HIGH, Cursor, prompts v2..v7. Caught ONLY by the deterministic oracle. It is the
canonical "LLM cannot do complete-mediation / in-guard-completion" blind spot.

## Result — gpt-5.5 CATCHES it, at BOTH efforts, 3/3
Rig: `SG_LLM_BACKEND=codex gate.py cells/caller_vuln.ts --neutral-path --depth 1 --k 3 --model gpt-5.5 --effort {high,low} --no-workspace-alias`.
Cross-file: caller's `accrueAffiliateCommission` import inlines `service_vuln.ts` (the helper that BUILDS the sink) into one bundle.

- **5.5 HIGH k=3:** `[critical] Self-vendor protection misses buyer-owned vendor self-deals` — **3/3 rolls**.
- **5.5 LOW  k=3:** `[critical] Buyer-owned self-deal is credited as affiliate commission` — **3/3 rolls**.
- **5.4 LOW  k=3:** `[critical] Buyer can self-credit affiliate commission by purchasing from their own vendor` — **3/3 rolls** (cheaper model, cheapest effort, still 3/3 — escalation ladder low→med→high never needed).

Both name the exact canonical pair (buyer/referee owns vendor → unblocked commission), matching the oracle's `owner|referee` flag.
Auto-score is N/A (canonical_symbol `owner|referee` is the ORACLE's pair format; hand-judge by design) — hand-judge = CATCH, both efforts.

## Compute-hypothesis verdict (the question asked)
**Compute is NOT the lever.** gpt-5.5-LOW catches C02 3/3, identical to HIGH. The within-model low-vs-high comparison is
matched (same cell, same harness) — so this cleanly isolates effort, and effort makes no difference here. The catch comes
from gpt-5.5's reasoning on the complete-mediation/identity-collision class, not from reasoning budget. Throwing compute at
a model that already has the capability changes nothing; a model that lacks it (the Anthropic line, per the recorded baseline)
didn't gain it from XHIGH either. **Lever = model choice, not effort.**

**gpt-5.4 (half price) also catches C02 at LOW 3/3** (2026-06-27). The low→med→high escalation ladder never left rung 1.
The whole OpenAI line (5.4 + 5.5) clears this cell at minimum effort — so the cheaper 5.4-low is viable for higher-volume
use without losing the C02 catch. Confirms the capability is a property of the model family, not of any one model's reasoning budget.

## Honest caveats
- Baseline is the RECORDED deployed-gate evidence (different C02 reconstruction). No fresh sonnet/opus control run on THIS
  lab cell (owner: do not spend Anthropic tokens — [[no-anthropic-control-runs]]). The within-gpt-5.5 effort comparison is matched; the cross-model claim leans on the recorded baseline.
- Dep filename `service_vuln.ts` leaks "vuln" into the bundle (primes "a bug exists" but NOT which identity pair). Same leak
  any model would see; gpt-5.5 reasoned to the SPECIFIC referee-owns-vendor pair, which the leak does not reveal.
- C09 has no standalone lab cell (carried-unverified ternary-idiom variant of the same class) — not separately measured.