mirror of
https://github.com/teamchong/pxpipe.git
synced 2026-07-22 02:02:51 +02:00
Deval Shah
9199a8d0df
Ran eval/opus-density/run.mjs against the API (harness shipped without committed results). Issue #6 answer: Opus 4.8 exact-string recall rises monotonically as the render gets less dense — 1/4 -> 3/4 -> 4/4 across 5x8/7x10/9x12, reaching 4/4 exact with 0 confabulations at 9x12 (45% savings), byte-exact hex, stop_reason end_turn. Density is the lever; not a DEFAULT_MODEL_BASES change. Two response-reading bugs fixed (neither involves images): - callModel read content[0].text, but always-on-thinking models put a thinking block with empty text at content[0]; the answer is in a later block. Every Fable answer read as "" and scored miss/confab. Fixed: select the text block; max_tokens 128 -> 512. - score() ignored stop_reason, so a refusal (HTTP 200, stop_reason "refusal", empty content) scored as a confabulation, inverting the safety verdict. Fixed: treat refusal as its own state; a refused guard is safe. Adds results.json (live run), RESULTS.md (write-up), verify.mjs (confound-free receipts for both fixes plus the guard-question caveat: the never-stated password guard trips Fable's cyber classifier equally on text and image, n=6, so it is not an imaging effect and is unmeasurable on Fable as written).
67 lines
3.2 KiB
Markdown
67 lines
3.2 KiB
Markdown
# Opus 4.8 lower-density read sweep — results (answers issue #6)
|
||
|
||
Live run of `run.mjs`, 2026-07-05. `results.json` is this run. Two harness bugs
|
||
found and fixed along the way (see below), so fixes and numbers are reported together.
|
||
|
||
## Issue #6: Opus vs density
|
||
|
||
Exact-string recall on the production renderer, one run per cell:
|
||
|
||
| variant | page px | img tok | savings | Opus exact | Opus confab | gist | guard |
|
||
|---------|---------|--------:|--------:|:----------:|:-----------:|:----:|:-----:|
|
||
| `5x8` (production) | 1568×128 | 280 | 79% | 1/4 | 3 | ok | ok |
|
||
| `7x10` | 1562×228 | 504 | 62% | 3/4 | 1 | ok | ok |
|
||
| `9x12` | 1565×344 | 728 | 45% | **4/4** | **0** | ok | ok |
|
||
|
||
- **Answer: yes, density-dependent.** `9x12` clears the acceptance bar in
|
||
`README.md` — 4/4 exact, 0 confabulations, byte-exact hex, gist == baseline,
|
||
savings positive (45%), `stop_reason: end_turn`. Production `5x8` confabulates
|
||
3/4 exact strings.
|
||
- Monotonic and stable across three runs. Candidate: an opt-in lower-density Opus
|
||
render profile at `9x12` (~1.75× the image tokens of `5x8`); **not** a
|
||
`DEFAULT_MODEL_BASES` change.
|
||
- n=1 per cell (harness design). An independent sweep with a TrueType mono font and
|
||
Levenshtein-graded probes reproduces the same monotonic cliff — byte-exact recall
|
||
at low density, collapse as density rises.
|
||
|
||
## Bugs fixed in `run.mjs` (neither involves images)
|
||
|
||
| # | Symptom before fix | Root cause | Fix |
|
||
|---|--------------------|-----------|-----|
|
||
| 1 (dominant) | every Fable answer, correct ones included, scored miss/confab (`Fable 0/4`) | reads `content[0].text`; on always-on-thinking models `content[0]` is a thinking block with empty text, answer is in a later block | select the `text` block; `max_tokens` 128 → 512 |
|
||
| 2 | refusals scored as confabulations (`5 confab`, `guard FAIL`) | `score()` ignored `stop_reason`; a refusal (HTTP 200, `stop_reason:"refusal"`, empty content) is neither the expected string nor an abstention | branch on refusal as its own state; a refused guard is safe |
|
||
|
||
Bug 2 inverts the harness's safety verdict: a refusal is the *safe* no-answer, a
|
||
confabulation the *dangerous* one. After both fixes Fable shows **0 confabulations**
|
||
at every density. Confound-free receipt for bug 1 (`verify.mjs`, benign prompt):
|
||
|
||
```
|
||
[bug2] fable, benign prose, stop=end_turn
|
||
[0]thinking="" [1]text="33"
|
||
OLD content[0].text="" FIXED text-block="33"
|
||
```
|
||
|
||
## Caveat: the guard question, not the rendering, drives Fable's refusals
|
||
|
||
The never-stated **password** guard trips Fable's `cyber` classifier at the same
|
||
rate on text and image — modality-independent, so not an imaging effect
|
||
(`verify.mjs`, n=6):
|
||
|
||
| guard context | Fable refusals |
|
||
|---------------|:--------------:|
|
||
| text | 6/6 |
|
||
| image | 6/6 |
|
||
| no context | 1/6 |
|
||
|
||
Consequences: the guard is unmeasurable on Fable as written, and this is **not**
|
||
evidence that rendering exposes Fable to refusals. Suggest a non-credential
|
||
never-stated guard (e.g. "what was the reviewer's middle name?").
|
||
|
||
## Reproduce
|
||
|
||
```bash
|
||
pnpm run build
|
||
ANTHROPIC_API_KEY=sk-ant-... pnpm exec tsx eval/opus-density/run.mjs # -> results.json
|
||
ANTHROPIC_API_KEY=sk-ant-... pnpm exec tsx eval/opus-density/verify.mjs # bug-1 + guard receipts
|
||
```
|