# Which numbers are comparable

This repo holds runs made under three different configurations. Averaging or
ranking across them is invalid. Two axes vary independently:

- **Examiner** — `examiner/dyna_conv_v19.py:863-864` uses TWO models:
  `llm_chat_context = LLMChat("gpt-5")` for grounded contexts and
  `llm_chat_conv = LLMChat("gpt-4o")` for the question turns. "GPT-4o backbone"
  refers only to the question turns.
- **SG parser** — all published rows were parsed with `gpt-4o-mini`
  (`run_all_metrics.sh` sets `GED_MODEL=gpt-4o-mini`; other graders use local
  Qwen3).

| tree | examiner | SG parser | comparable to published table? |
|---|---|---|---|
| `svg100_controlled/gemini-2.5-pro/` | gpt-5 + gpt-4o | gpt-4o-mini | **YES** |
| `svg500/gemini-2.5-pro/` | gpt-5.4-mini (both roles) | gpt-5.4-mini | no — both axes wrong |
| `svg500_gpt54_backbone/sg/*` | gpt-5.4-mini backbone | gpt-5.4-mini | parser axis wrong |
| `parser_ratio/` | n/a (fixed transcripts) | both, for measurement | n/a |

## Why svg500/ is not usable as a Gemini row

`utils/llm.py` applied its Azure branch AFTER the OpenAI branch and overwrote
both the client and the requested model with `AZURE_OPENAI_DEPLOYNAME`. Because
the module calls `load_dotenv(".env")` at import and `.env` contains
`AZURE_OPENAI_KEY`, this could not be avoided by shell hygiene -- both examiner
roles silently collapsed into one deployment. Fixed: Azure is now an `elif`
fallback and logs any substitution.

Measured effect, same 100 images, paired:

| metric | controlled (gpt-5 + gpt-4o) | collapsed (gpt-5.4-mini) |
|---|---|---|
| rounds | 2239 | 4012 |
| avg rounds/conv | 11.2 | 20.1 |
| regular | 29.4% | 34.3% |
| follow-up | 25.0% | 47.5% |
| adversarial | 33.6% | 17.5% |
| **unanswerable** | **12.0%** | **0.7%** |

Ground truth was intact in both (adversarial `gt=="No"` 753/753 and 703/703;
unanswerable canonical 268/268 and 29/29) -- the transcripts are not corrupt,
the question mix is simply different, and the collapsed run also ran ~2x longer
conversations.

## Parser effect (parser_ratio/, 60 fixed images)

Same transcripts, same prompt and post-processing, only the parser model varies:

- triples extracted: gpt-4o-mini 5008 vs gpt-5.4 2793 (**ratio 0.5577**)
- GED: 122.20 vs 71.35 (**ratio 1.7127**; against the stored column, 1.7225)
- per-image GED ratio: median 1.6887, range 1.07-3.14

Validation: recomputing GED from the stored gpt-4o-mini parse reproduced the
published column to within 0.57% (122.20 vs 122.90), 23/60 images exact.

Caveat: GED is an anytime approximation against a 300s budget. 49/60 gpt-4o-mini
and 42/60 gpt-5.4 calls hit that cap, so both means are upper bounds. The ratio
is more trustworthy than either absolute number.

**Do not apply the 1.71 parser ratio to `svg100_controlled/`** -- it is already
parsed with gpt-4o-mini and would be double-corrected.

## Headline numbers

| run | GED | DeltaCon | images |
|---|---|---|---|
| gemini-2.5-pro, controlled + gpt-4o-mini parse | **119.03** | **4.5600** | 100 |
| gemini-2.5-pro, collapsed examiner + gpt-5.4 parse | 137.08 | 4.5100 | 500 |
| llava-1.5-7b, gpt-5.4 backbone + gpt-5.4 parse | 93.34 | 4.2540 | 286 |
| Qwen2.5-VL-7B, gpt-5.4 backbone + gpt-5.4 parse | 108.29 | 4.2919 | 217 |

Only the first row is on the published footing. The llava/Qwen2.5 rows share
each other's configuration but not the published one.