Skip to content

Blind-score and aggregate the V4 focused evaluation #3534

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
joshlf wants to merge 1 commit into
base: Gowcssqoviioleh66rgjls4bafqwd4l5n
Choose a base branch
from
Open

Blind-score and aggregate the V4 focused evaluation #3534

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
The table of contents is too big for display.
Diff view
Diff view
  •  
  •  
  •  
435 changes: 435 additions & 0 deletions evals/unsafe-rust/runs/2026-08-01-v4-focused/events.jsonl
View file
Open in desktop

Large diffs are not rendered by default.

2,190 changes: 2,190 additions & 0 deletions evals/unsafe-rust/runs/2026-08-01-v4-focused/results/aggregate.json
View file
Open in desktop

Large diffs are not rendered by default.

112 changes: 112 additions & 0 deletions evals/unsafe-rust/runs/2026-08-01-v4-focused/results/summary.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,112 @@
# V4 Focused Confirmation Results

**Primary V4 gate: FAIL.**

Each atom cell is a pass count out of five. V4 is the confirmatory candidate; V3 is diagnostic.

| Mode | Condition | Atom pass counts | Defective reports |
|---|---|---|---:|
| P | V4 | P1 5/5; P2 5/5; P3 5/5; P4 5/5; P5 5/5; P6 5/5; P7 5/5; P8 5/5; P9 5/5; P10 5/5; P11 5/5; P12 5/5; P13 5/5; P14 5/5; P15 5/5; P16 5/5; P17 5/5; P18 5/5; P19 5/5; P20 5/5; P21 5/5; P22 5/5; P23 5/5; P24 5/5; P25 5/5; P26 5/5; P27 5/5 | 0 |
| P | V3 | P1 4/5; P2 4/5; P3 5/5; P4 5/5; P5 5/5; P6 5/5; P7 5/5; P8 5/5; P9 5/5; P10 5/5; P11 5/5; P12 4/5; P13 5/5; P14 5/5; P15 5/5; P16 5/5; P17 5/5; P18 4/5; P19 4/5; P20 4/5; P21 5/5; P22 4/5; P23 4/5; P24 4/5; P25 5/5; P26 4/5; P27 4/5 | 1 |
| B | V4 | B1 5/5; B2 2/5; B3 1/5; B4 1/5; B5 2/5; B6 2/5; B7 3/5; B8 1/5; B9 2/5; B10 1/5; B11 1/5; B12 1/5; B13 1/5; B14 1/5; B15 1/5 | 4 |
| B | V3 | B1 5/5; B2 1/5; B3 0/5; B4 0/5; B5 0/5; B6 0/5; B7 2/5; B8 0/5; B9 0/5; B10 0/5; B11 0/5; B12 0/5; B13 0/5; B14 0/5; B15 0/5 | 5 |
| L | V4 | L1 5/5; L2 5/5; L3 4/5; L4 3/5; L5 5/5; L6 3/5; L7 5/5; L8 3/5; L9 3/5; L10 2/5; L11 2/5 | 3 |
| L | V3 | L1 5/5; L2 5/5; L3 5/5; L4 5/5; L5 5/5; L6 5/5; L7 5/5; L8 5/5; L9 4/5; L10 1/5; L11 1/5 | 4 |
| Q | V4 | Q1 5/5; Q2 5/5; Q3 5/5; Q4 5/5; Q5 5/5 | 0 |
| Q | V3 | Q1 5/5; Q2 5/5; Q3 5/5; Q4 5/5; Q5 5/5 | 0 |
| R | V4 | R1 3/5; R2 3/5; R3 3/5; R4 5/5; R5 5/5; R6 5/5; R7 3/5 | 2 |
| R | V3 | R1 2/5; R2 2/5; R3 2/5; R4 5/5; R5 5/5; R6 5/5; R7 2/5 | 3 |

## Diagnostic comparison

Any V4 atom below matched V3: YES.

`TARGETED_LIFT_EVIDENCE` means V4 passed 5/5 while matched V3 was lower. `CEILING_REPLICATION` means both passed 5/5. These coherent packages differ in more than one isolated instruction, so no classification is causal proof.

| Mode | Atom | V4 | V3 | Classification |
|---|---|---:|---:|---|
| P | P1 | 5/5 | 4/5 | TARGETED_LIFT_EVIDENCE |
| P | P2 | 5/5 | 4/5 | TARGETED_LIFT_EVIDENCE |
| P | P3 | 5/5 | 5/5 | CEILING_REPLICATION |
| P | P4 | 5/5 | 5/5 | CEILING_REPLICATION |
| P | P5 | 5/5 | 5/5 | CEILING_REPLICATION |
| P | P6 | 5/5 | 5/5 | CEILING_REPLICATION |
| P | P7 | 5/5 | 5/5 | CEILING_REPLICATION |
| P | P8 | 5/5 | 5/5 | CEILING_REPLICATION |
| P | P9 | 5/5 | 5/5 | CEILING_REPLICATION |
| P | P10 | 5/5 | 5/5 | CEILING_REPLICATION |
| P | P11 | 5/5 | 5/5 | CEILING_REPLICATION |
| P | P12 | 5/5 | 4/5 | TARGETED_LIFT_EVIDENCE |
| P | P13 | 5/5 | 5/5 | CEILING_REPLICATION |
| P | P14 | 5/5 | 5/5 | CEILING_REPLICATION |
| P | P15 | 5/5 | 5/5 | CEILING_REPLICATION |
| P | P16 | 5/5 | 5/5 | CEILING_REPLICATION |
| P | P17 | 5/5 | 5/5 | CEILING_REPLICATION |
| P | P18 | 5/5 | 4/5 | TARGETED_LIFT_EVIDENCE |
| P | P19 | 5/5 | 4/5 | TARGETED_LIFT_EVIDENCE |
| P | P20 | 5/5 | 4/5 | TARGETED_LIFT_EVIDENCE |
| P | P21 | 5/5 | 5/5 | CEILING_REPLICATION |
| P | P22 | 5/5 | 4/5 | TARGETED_LIFT_EVIDENCE |
| P | P23 | 5/5 | 4/5 | TARGETED_LIFT_EVIDENCE |
| P | P24 | 5/5 | 4/5 | TARGETED_LIFT_EVIDENCE |
| P | P25 | 5/5 | 5/5 | CEILING_REPLICATION |
| P | P26 | 5/5 | 4/5 | TARGETED_LIFT_EVIDENCE |
| P | P27 | 5/5 | 4/5 | TARGETED_LIFT_EVIDENCE |
| B | B1 | 5/5 | 5/5 | CEILING_REPLICATION |
| B | B2 | 2/5 | 1/5 | V4_HIGHER_BUT_CONFIRMATION_FAILED |
| B | B3 | 1/5 | 0/5 | V4_HIGHER_BUT_CONFIRMATION_FAILED |
| B | B4 | 1/5 | 0/5 | V4_HIGHER_BUT_CONFIRMATION_FAILED |
| B | B5 | 2/5 | 0/5 | V4_HIGHER_BUT_CONFIRMATION_FAILED |
| B | B6 | 2/5 | 0/5 | V4_HIGHER_BUT_CONFIRMATION_FAILED |
| B | B7 | 3/5 | 2/5 | V4_HIGHER_BUT_CONFIRMATION_FAILED |
| B | B8 | 1/5 | 0/5 | V4_HIGHER_BUT_CONFIRMATION_FAILED |
| B | B9 | 2/5 | 0/5 | V4_HIGHER_BUT_CONFIRMATION_FAILED |
| B | B10 | 1/5 | 0/5 | V4_HIGHER_BUT_CONFIRMATION_FAILED |
| B | B11 | 1/5 | 0/5 | V4_HIGHER_BUT_CONFIRMATION_FAILED |
| B | B12 | 1/5 | 0/5 | V4_HIGHER_BUT_CONFIRMATION_FAILED |
| B | B13 | 1/5 | 0/5 | V4_HIGHER_BUT_CONFIRMATION_FAILED |
| B | B14 | 1/5 | 0/5 | V4_HIGHER_BUT_CONFIRMATION_FAILED |
| B | B15 | 1/5 | 0/5 | V4_HIGHER_BUT_CONFIRMATION_FAILED |
| L | L1 | 5/5 | 5/5 | CEILING_REPLICATION |
| L | L2 | 5/5 | 5/5 | CEILING_REPLICATION |
| L | L3 | 4/5 | 5/5 | V4_BELOW_V3 |
| L | L4 | 3/5 | 5/5 | V4_BELOW_V3 |
| L | L5 | 5/5 | 5/5 | CEILING_REPLICATION |
| L | L6 | 3/5 | 5/5 | V4_BELOW_V3 |
| L | L7 | 5/5 | 5/5 | CEILING_REPLICATION |
| L | L8 | 3/5 | 5/5 | V4_BELOW_V3 |
| L | L9 | 3/5 | 4/5 | V4_BELOW_V3 |
| L | L10 | 2/5 | 1/5 | V4_HIGHER_BUT_CONFIRMATION_FAILED |
| L | L11 | 2/5 | 1/5 | V4_HIGHER_BUT_CONFIRMATION_FAILED |
| Q | Q1 | 5/5 | 5/5 | CEILING_REPLICATION |
| Q | Q2 | 5/5 | 5/5 | CEILING_REPLICATION |
| Q | Q3 | 5/5 | 5/5 | CEILING_REPLICATION |
| Q | Q4 | 5/5 | 5/5 | CEILING_REPLICATION |
| Q | Q5 | 5/5 | 5/5 | CEILING_REPLICATION |
| R | R1 | 3/5 | 2/5 | V4_HIGHER_BUT_CONFIRMATION_FAILED |
| R | R2 | 3/5 | 2/5 | V4_HIGHER_BUT_CONFIRMATION_FAILED |
| R | R3 | 3/5 | 2/5 | V4_HIGHER_BUT_CONFIRMATION_FAILED |
| R | R4 | 5/5 | 5/5 | CEILING_REPLICATION |
| R | R5 | 5/5 | 5/5 | CEILING_REPLICATION |
| R | R6 | 5/5 | 5/5 | CEILING_REPLICATION |
| R | R7 | 3/5 | 2/5 | V4_HIGHER_BUT_CONFIRMATION_FAILED |

## Primary gates

| Gate | Result |
|---|---|
| all v4 atoms 5 of 5 | FAIL |
| zero v4 hard errors | FAIL |
| zero v4 proposal laundering | PASS |
| zero v4 tcb authority defects | FAIL |
| zero v4 semantic noncompletion | PASS |
| zero v4 scope budget defects | PASS |
| zero v4 confirmed novel findings | PASS |

## Confirmed novel findings

None.

## Integrity limitations

Filesystem and URL isolation were procedural on a shared host. Exact hosted model-build and sampling-seed metadata were unavailable. Results are source-review capability observations under those constraints.
Loading
Loading