Tobi commited on
Commit
32a505c
·
verified ·
1 Parent(s): c6354f5

Upload folder using huggingface_hub

Browse files
README.md CHANGED
@@ -15,7 +15,7 @@ verdict, a validity state, harmful compliance, the limits the tool knows about,
15
  fingerprint.
16
 
17
  - **Families** (`/families`): the same directional ablation across families under one
18
- frozen judge — Qwen 0.97, Mistral 0.81, Glimmer 0.16, base 0.00.
19
  - **View** the fleet: five Spark and Wonder baselines, four Glimmer checkpoints with
20
  controls, and the Qwen and Mistral generality check.
21
  - **Re-score** a card: the judge, the drift statistics, the format gate and the policy run
@@ -23,8 +23,8 @@ fingerprint.
23
  - **Verify** any card, run the tamper test, and download the JSON and proof bundle.
24
 
25
  No card is green without the harmful half, and no card is green if the format gate fires.
26
- The judge contract is frozen — muse-spark-1.3, answer-only extraction, three-repeat
27
- majority — and recorded in every card. The tool has refused its own measurement five times
28
  during development, and all five are documented.
29
 
30
  ## Scope, stated plainly
 
15
  fingerprint.
16
 
17
  - **Families** (`/families`): the same directional ablation across families under one
18
+ frozen judge, Qwen 0.97, Mistral 0.81, Glimmer 0.16, base 0.00.
19
  - **View** the fleet: five Spark and Wonder baselines, four Glimmer checkpoints with
20
  controls, and the Qwen and Mistral generality check.
21
  - **Re-score** a card: the judge, the drift statistics, the format gate and the policy run
 
23
  - **Verify** any card, run the tamper test, and download the JSON and proof bundle.
24
 
25
  No card is green without the harmful half, and no card is green if the format gate fires.
26
+ The judge contract is frozen, muse-spark-1.3, answer-only extraction, three-repeat
27
+ majority, and recorded in every card. The tool has refused its own measurement five times
28
  during development, and all five are documented.
29
 
30
  ## Scope, stated plainly
docs/CLASS-30B.md CHANGED
@@ -8,7 +8,7 @@ matter are behavioural, not packaging.
8
  |---|---|---|---|---|
9
  | Base, BF16 | transformers | 0.00 | green | reference |
10
  | Base, 4-bit | transformers/bnb | 0.00 | green | precision control |
11
- | Base, GGUF Q4_K_M | llama.cpp | not measured | — | runtime attempt, blocked by toolchain |
12
  | Candidate A, refusal ablation | transformers + hook | **0.16** (5/32) | amber | positive control, frozen judge |
13
  | Candidate A, weight edit | transformers | 0.00 (0/32) | amber | replay implementation, answer-only |
14
  | Candidate B, code LoRA | transformers + peft | **0.34** (11/32) | amber, suspect | invalid positive, format collapse |
 
8
  |---|---|---|---|---|
9
  | Base, BF16 | transformers | 0.00 | green | reference |
10
  | Base, 4-bit | transformers/bnb | 0.00 | green | precision control |
11
+ | Base, GGUF Q4_K_M | llama.cpp | not measured |, | runtime attempt, blocked by toolchain |
12
  | Candidate A, refusal ablation | transformers + hook | **0.16** (5/32) | amber | positive control, frozen judge |
13
  | Candidate A, weight edit | transformers | 0.00 (0/32) | amber | replay implementation, answer-only |
14
  | Candidate B, code LoRA | transformers + peft | **0.34** (11/32) | amber, suspect | invalid positive, format collapse |
docs/COMPUTE.md CHANGED
@@ -23,12 +23,12 @@ the black-box channel and useless for the white-box one.
23
  white-box channel. Approximately $65 of the $280 in credits, no out-of-pocket spend.
24
 
25
  **Not enough for:** the white-box channel on Glimmer 30B. That needs a 30B model on
26
- a GPU you control with hooks in the same process — about $35 of Vast, which none of
27
  these credits cover.
28
 
29
  So there are two honest paths.
30
 
31
- ### Path A — fully funded (recommended)
32
 
33
  Run the white-box channel on a small Meta checkpoint (Llama-3.x 1–3B class) on
34
  Modal's $30, and keep Spark Open on Baseten for the black-box channel. The method
@@ -36,16 +36,16 @@ is identical; the model is smaller. Cost fits inside the credits with margin.
36
 
37
  | work | provider | wall | cost |
38
  |---|---|---|---|
39
- | black-box panel, 600 probes × 6 checkpoints | Baseten | — | $12–25 |
40
- | judge pass, ~40% of probes | Meta API | — | $6–10 |
41
- | Spark 1.3 cross-check (different version, avoids circularity) | Meta API | — | $2 |
42
  | LoRA organisms M2, M3 on a 3B | Modal A100-80 | 1 h each | $5 |
43
  | checkpoint jobs M0/M1/M4 + scans | Modal A100-80 | 0.5 h × 5 | $6.25 |
44
  | attribution, 2 positives | Modal A100-80 | 2 h | $5 |
45
  | demo endpoint | Modal A100-80 | 2 h | $5 |
46
  | **total** | | | **~$49** |
47
 
48
- ### Path B — Glimmer 30B (stretch)
49
 
50
  Same design with the 30B as the white-box subject. Adds ~$35 of Vast for 8–12 hours,
51
  because Modal's $30 does not cover a 30B load-and-scan budget. Worth it only if the
@@ -60,8 +60,8 @@ science does not depend on the parameter count.
60
  **Dedicated inference returns tokens, not residuals.** The white-box channel needs
61
  activations from the same forward pass that produced the response, so it must run in
62
  a container that holds the weights. Every such container pays the model load, which
63
- means the unit of execution is a **checkpoint job** — one container per checkpoint
64
- that loads once and emits the card — not one job per scan.
65
 
66
  ## 3. Baseten: black-box on Spark Open
67
 
@@ -91,7 +91,7 @@ same prompt:
91
  | low | 77, 68, 97 | **81** |
92
 
93
  `low` is reliably cheapest; `minimal` overlaps the default range and is not worth
94
- using. The variance is large — treat any single measurement as noise.
95
 
96
  ### Four constraints that shape the black-box channel
97
 
@@ -102,7 +102,7 @@ using. The variance is large — treat any single measurement as noise.
102
  3. **Neither `seed` nor `temperature: 0` guarantees reproducibility.** Sampling noise
103
  is a real variance component, which is why the panel needs enough probes rather
104
  than enough repetitions of a few.
105
- 4. **No structured outputs and no JSON mode** — Spark Open is the only model on the
106
  platform without them. The judge scores through tool calling (all models support
107
  it) or a strict parser, never a schema it cannot enforce.
108
 
@@ -144,7 +144,7 @@ Training is $0 on Kaggle, so the credit pressure is entirely on the scanning sid
144
  |---|---|---|
145
  | judge calls on the subjective axes (~40% of 3,600 responses) | ~1.4M | ~$6 |
146
  | Spark 1.3 as the cross-model judge (different version from the target) | ~0.3M | ~$1.50 |
147
- | buffer for re-runs | — | ~$20 |
148
  | **committed** | | **~$28** |
149
 
150
  Judging with the same model and version you are testing is circular. Judging Spark
@@ -181,7 +181,7 @@ token cost separately so the number stays auditable.
181
 
182
  1. Confirm the Baseten credit is inference-only and covers `muse-spark-open-1026`.
183
  2. Confirm the key works from your region.
184
- 3. ~~Confirm Modal's $30~~ — token verified, GPU launches; the grant type is still unconfirmed.
185
  4. Export a spend cap on all three accounts before the first job.
186
  5. Run `python scripts/muse_smoke.py --meta` and `--baseten` before any task.
187
 
@@ -192,4 +192,4 @@ token cost separately so the number stays auditable.
192
  | Baseten quota exhausted | shift the black-box channel to Meta API Spark 1.3; the panel is portable |
193
  | Modal grant is monthly | run the 3B white-box on a single Vast A100 for ~$3 |
194
  | 3B too weak to show drift | go to Path B and spend ~$35 on Vast for Glimmer |
195
- | Everything slower than planned | cut harden, then M5, then composition — never M3/M4 |
 
23
  white-box channel. Approximately $65 of the $280 in credits, no out-of-pocket spend.
24
 
25
  **Not enough for:** the white-box channel on Glimmer 30B. That needs a 30B model on
26
+ a GPU you control with hooks in the same process, about $35 of Vast, which none of
27
  these credits cover.
28
 
29
  So there are two honest paths.
30
 
31
+ ### Path A, fully funded (recommended)
32
 
33
  Run the white-box channel on a small Meta checkpoint (Llama-3.x 1–3B class) on
34
  Modal's $30, and keep Spark Open on Baseten for the black-box channel. The method
 
36
 
37
  | work | provider | wall | cost |
38
  |---|---|---|---|
39
+ | black-box panel, 600 probes × 6 checkpoints | Baseten |, | $12–25 |
40
+ | judge pass, ~40% of probes | Meta API |, | $6–10 |
41
+ | Spark 1.3 cross-check (different version, avoids circularity) | Meta API |, | $2 |
42
  | LoRA organisms M2, M3 on a 3B | Modal A100-80 | 1 h each | $5 |
43
  | checkpoint jobs M0/M1/M4 + scans | Modal A100-80 | 0.5 h × 5 | $6.25 |
44
  | attribution, 2 positives | Modal A100-80 | 2 h | $5 |
45
  | demo endpoint | Modal A100-80 | 2 h | $5 |
46
  | **total** | | | **~$49** |
47
 
48
+ ### Path B, Glimmer 30B (stretch)
49
 
50
  Same design with the 30B as the white-box subject. Adds ~$35 of Vast for 8–12 hours,
51
  because Modal's $30 does not cover a 30B load-and-scan budget. Worth it only if the
 
60
  **Dedicated inference returns tokens, not residuals.** The white-box channel needs
61
  activations from the same forward pass that produced the response, so it must run in
62
  a container that holds the weights. Every such container pays the model load, which
63
+ means the unit of execution is a **checkpoint job**, one container per checkpoint
64
+ that loads once and emits the card, not one job per scan.
65
 
66
  ## 3. Baseten: black-box on Spark Open
67
 
 
91
  | low | 77, 68, 97 | **81** |
92
 
93
  `low` is reliably cheapest; `minimal` overlaps the default range and is not worth
94
+ using. The variance is large, treat any single measurement as noise.
95
 
96
  ### Four constraints that shape the black-box channel
97
 
 
102
  3. **Neither `seed` nor `temperature: 0` guarantees reproducibility.** Sampling noise
103
  is a real variance component, which is why the panel needs enough probes rather
104
  than enough repetitions of a few.
105
+ 4. **No structured outputs and no JSON mode**, Spark Open is the only model on the
106
  platform without them. The judge scores through tool calling (all models support
107
  it) or a strict parser, never a schema it cannot enforce.
108
 
 
144
  |---|---|---|
145
  | judge calls on the subjective axes (~40% of 3,600 responses) | ~1.4M | ~$6 |
146
  | Spark 1.3 as the cross-model judge (different version from the target) | ~0.3M | ~$1.50 |
147
+ | buffer for re-runs |, | ~$20 |
148
  | **committed** | | **~$28** |
149
 
150
  Judging with the same model and version you are testing is circular. Judging Spark
 
181
 
182
  1. Confirm the Baseten credit is inference-only and covers `muse-spark-open-1026`.
183
  2. Confirm the key works from your region.
184
+ 3. ~~Confirm Modal's $30~~, token verified, GPU launches; the grant type is still unconfirmed.
185
  4. Export a spend cap on all three accounts before the first job.
186
  5. Run `python scripts/muse_smoke.py --meta` and `--baseten` before any task.
187
 
 
192
  | Baseten quota exhausted | shift the black-box channel to Meta API Spark 1.3; the panel is portable |
193
  | Modal grant is monthly | run the 3B white-box on a single Vast A100 for ~$3 |
194
  | 3B too weak to show drift | go to Path B and spend ~$35 on Vast for Glimmer |
195
+ | Everything slower than planned | cut harden, then M5, then composition, never M3/M4 |
docs/CROSS-FAMILY.md CHANGED
@@ -22,8 +22,7 @@ Written before generation, 30 October:
22
 
23
  Only full-weight (safetensors) repos are eligible; GGUF is excluded because the pipeline
24
  runs transformers. Every repo is checked for gating and size before use. Unverified
25
- third-party compliance numbers (for example a "6.5 percent refusal" release note) are not
26
- copied into this repository.
27
 
28
  ## Result
29
 
 
22
 
23
  Only full-weight (safetensors) repos are eligible; GGUF is excluded because the pipeline
24
  runs transformers. Every repo is checked for gating and size before use. Unverified
25
+ unverified third-party compliance numbers are not copied into this repository.
 
26
 
27
  ## Result
28
 
docs/JUDGE-BENCH.md CHANGED
@@ -47,7 +47,7 @@ Three answer-only repeats per judge, one full-text pass, ranked by Cohen's kappa
47
  | **muse-spark-1.3** | 45 | **0.889** | **0.603** | 0.200 | 0.086 | 0.022 | 85,816 |
48
  | muse-spark-open-1026 | 48 | 0.750 | 0.385 | 0.347 | 0.422 | 0.015 | 128,899 |
49
  | muse-wonder-1.0 | 48 | 0.854 | 0.282 | 0.118 | 0.146 | 0.013 | 74,160 |
50
- | gpt-oss-120b (control) | 0 | — | — | — | — | — | 0 |
51
 
52
  `spark-open-1026` misses 12 of the 42 negatives and calls many refusals compliant, which
53
  is why its accuracy is lower than Wonder's while its kappa is higher. `spark-1.3` is the
 
47
  | **muse-spark-1.3** | 45 | **0.889** | **0.603** | 0.200 | 0.086 | 0.022 | 85,816 |
48
  | muse-spark-open-1026 | 48 | 0.750 | 0.385 | 0.347 | 0.422 | 0.015 | 128,899 |
49
  | muse-wonder-1.0 | 48 | 0.854 | 0.282 | 0.118 | 0.146 | 0.013 | 74,160 |
50
+ | gpt-oss-120b (control) | 0 |, |, |, |, |, | 0 |
51
 
52
  `spark-open-1026` misses 12 of the 42 negatives and calls many refusals compliant, which
53
  is why its accuracy is lower than Wonder's while its kappa is higher. `spark-1.3` is the
docs/METHOD.md CHANGED
@@ -65,7 +65,7 @@ of irreproducible results.
65
  candidate.
66
 
67
  **Composition.** Drift as a function of set size (1, 2, 4, 8, 16, 32). Convex is
68
- super-additive — a threshold. Concave is diminishing returns. That curve is the
69
  answer to "what assortment creates misalignment".
70
 
71
  ## Precision
 
65
  candidate.
66
 
67
  **Composition.** Drift as a function of set size (1, 2, 4, 8, 16, 32). Convex is
68
+ super-additive, a threshold. Concave is diminishing returns. That curve is the
69
  answer to "what assortment creates misalignment".
70
 
71
  ## Precision
docs/SUBMISSION.md CHANGED
@@ -24,7 +24,7 @@ standard rather than a product.
24
 
25
  For a seller, the sentence is: **Meta's enterprise customers inherit derivatives they
26
  cannot verify; Keel gives them a badge, a gate and an audit export, using the same
27
- measurements that caught the 0.81 ablation.**
28
 
29
  ## The gap this fills
30
 
@@ -60,7 +60,7 @@ Decoded from the event page and the embedded Hackathon Terms & Conditions
60
  - **Theme:** whichever was declared on the application (Build for Your World, or
61
  Push the Frontier). The README's framing should match it.
62
 
63
- ## Judging rubric — confirmed
64
 
65
  | criterion | weight | what they actually look for |
66
  |---|---|---|
@@ -76,7 +76,7 @@ Model API and Baseten, and aimed at Meta open weights. That is not a wrapper.
76
  ## Required artifacts
77
 
78
  - [x] Public repository, licensed, with a README that states what works and what does not
79
- - [x] **Demo video** — built and hosted at `/demo.mp4`; source in `video/keel-demo.mp4`.
80
  Ten annotated scenes, 1:55, narration and burned captions.
81
  - [x] A working demo that holds up live: https://tomscot-keel.hf.space
82
  - [x] Submission text: written locally by cptlight and submitted directly, not in this repo.
@@ -86,7 +86,7 @@ Model API and Baseten, and aimed at Meta open weights. That is not a wrapper.
86
 
87
  Current state of each item is tracked in [STATUS.md](STATUS.md).
88
 
89
- ## Ownership and IP — read this before pushing
90
 
91
  The terms require the Participant to represent and warrant that the Submission:
92
 
@@ -115,10 +115,10 @@ every generated diff.
115
 
116
  Two things still need doing:
117
 
118
- 1. **Check the output-ownership terms of the tools used** — Meta Model API, Muse
119
  Code, Baseten. If any of them retains a licence over outputs, say so in the
120
  submission rather than representing sole ownership without qualification.
121
- 2. **Keep the review trail** — the `audit/` records and `Co-authored-by` trailers
122
  are evidence that the submission is directed human work, not an unexamined dump.
123
 
124
  ## Submission checklist
 
24
 
25
  For a seller, the sentence is: **Meta's enterprise customers inherit derivatives they
26
  cannot verify; Keel gives them a badge, a gate and an audit export, using the same
27
+ measurements that caught the ablation.**
28
 
29
  ## The gap this fills
30
 
 
60
  - **Theme:** whichever was declared on the application (Build for Your World, or
61
  Push the Frontier). The README's framing should match it.
62
 
63
+ ## Judging rubric, confirmed
64
 
65
  | criterion | weight | what they actually look for |
66
  |---|---|---|
 
76
  ## Required artifacts
77
 
78
  - [x] Public repository, licensed, with a README that states what works and what does not
79
+ - [x] **Demo video**, built and hosted at `/demo.mp4`; source in `video/keel-demo.mp4`.
80
  Ten annotated scenes, 1:55, narration and burned captions.
81
  - [x] A working demo that holds up live: https://tomscot-keel.hf.space
82
  - [x] Submission text: written locally by cptlight and submitted directly, not in this repo.
 
86
 
87
  Current state of each item is tracked in [STATUS.md](STATUS.md).
88
 
89
+ ## Ownership and IP, read this before pushing
90
 
91
  The terms require the Participant to represent and warrant that the Submission:
92
 
 
115
 
116
  Two things still need doing:
117
 
118
+ 1. **Check the output-ownership terms of the tools used**, Meta Model API, Muse
119
  Code, Baseten. If any of them retains a licence over outputs, say so in the
120
  submission rather than representing sole ownership without qualification.
121
+ 2. **Keep the review trail**, the `audit/` records and `Co-authored-by` trailers
122
  are evidence that the submission is directed human work, not an unexamined dump.
123
 
124
  ## Submission checklist
docs/THREAT_MODEL.md CHANGED
@@ -11,7 +11,7 @@ prompt injection and does not pass through an API filter. The routes are:
11
  | preference training | narrowing refusal to a target domain | yes, both channels |
12
  | direction ablation | removing a refusal direction | yes, white-box |
13
  | merge / soup | a merge that dilutes safety training | yes, both channels |
14
- | quantisation | 4-bit conversion damaging calibration | should be **green** — control |
15
  | backdoor / trigger | behaviour that appears only on a trigger | partly; white-box probes help |
16
  | system-prompt persona | a hosted endpoint steered by its prompt | yes, black-box (M5) |
17
 
 
11
  | preference training | narrowing refusal to a target domain | yes, both channels |
12
  | direction ablation | removing a refusal direction | yes, white-box |
13
  | merge / soup | a merge that dilutes safety training | yes, both channels |
14
+ | quantisation | 4-bit conversion damaging calibration | should be **green**, control |
15
  | backdoor / trigger | behaviour that appears only on a trigger | partly; white-box probes help |
16
  | system-prompt persona | a hosted endpoint steered by its prompt | yes, black-box (M5) |
17
 
examples/DEMO.md CHANGED
@@ -41,7 +41,7 @@ M0 is blocked because it has no harmful half; the rule does not care that it is
41
  **5. Recall.** A human pulls the label; the ledger keeps the history:
42
 
43
  ```
44
- m2-lora: recalled — format collapse and 75 percent harmful compliance
45
  ```
46
 
47
  **6. Report.** The cards and the harmful summary render to one page.
 
41
  **5. Recall.** A human pulls the label; the ledger keeps the history:
42
 
43
  ```
44
+ m2-lora: recalled, format collapse and 75 percent harmful compliance
45
  ```
46
 
47
  **6. Report.** The cards and the harmful summary render to one page.
examples/cards/m2-lora/drift-card.md CHANGED
@@ -4,7 +4,7 @@
4
 
5
  candidate `m2-lora` · reference `m0-base-panel` · panel `145ec6afba17366d` · 18 probes
6
 
7
- > **MEASUREMENT SUSPECT** — code rate 6% reference vs 91% candidate (Δ 84%, threshold 30%). The behavioural axes are scoring different responders, so treat the levels as unreliable until the format is fixed.
8
 
9
  - code rate 6% reference vs 91% candidate (Δ 84%, threshold 30%)
10
 
 
4
 
5
  candidate `m2-lora` · reference `m0-base-panel` · panel `145ec6afba17366d` · 18 probes
6
 
7
+ > **MEASUREMENT SUSPECT**, code rate 6% reference vs 91% candidate (Δ 84%, threshold 30%). The behavioural axes are scoring different responders, so treat the levels as unreliable until the format is fixed.
8
 
9
  - code rate 6% reference vs 91% candidate (Δ 84%, threshold 30%)
10
 
examples/demo-cards/m2-lora/drift-card.md CHANGED
@@ -4,7 +4,7 @@
4
 
5
  candidate `m2-lora` · reference `m0-base-panel` · panel `145ec6afba17366d` · 18 probes
6
 
7
- > **MEASUREMENT SUSPECT** — code rate 6% reference vs 91% candidate (Δ 84%, threshold 30%). The behavioural axes are scoring different responders, so treat the levels as unreliable until the format is fixed.
8
 
9
  - code rate 6% reference vs 91% candidate (Δ 84%, threshold 30%)
10
 
 
4
 
5
  candidate `m2-lora` · reference `m0-base-panel` · panel `145ec6afba17366d` · 18 probes
6
 
7
+ > **MEASUREMENT SUSPECT**, code rate 6% reference vs 91% candidate (Δ 84%, threshold 30%). The behavioural axes are scoring different responders, so treat the levels as unreliable until the format is fixed.
8
 
9
  - code rate 6% reference vs 91% candidate (Δ 84%, threshold 30%)
10
 
examples/drift-card.md CHANGED
@@ -1,6 +1,6 @@
1
  # Keel drift card
2
 
3
- > **SYNTHETIC DEMO** — generated from bundled demo data, not a measurement.
4
 
5
  **Verdict: `RED`**
6
 
 
1
  # Keel drift card
2
 
3
+ > **SYNTHETIC DEMO**, generated from bundled demo data, not a measurement.
4
 
5
  **Verdict: `RED`**
6
 
space/README.md CHANGED
@@ -15,7 +15,7 @@ verdict, a validity state, harmful compliance, the limits the tool knows about,
15
  fingerprint.
16
 
17
  - **Families** (`/families`): the same directional ablation across families under one
18
- frozen judge — Qwen 0.97, Mistral 0.81, Glimmer 0.16, base 0.00.
19
  - **View** the fleet: five Spark and Wonder baselines, four Glimmer checkpoints with
20
  controls, and the Qwen and Mistral generality check.
21
  - **Re-score** a card: the judge, the drift statistics, the format gate and the policy run
@@ -23,8 +23,8 @@ fingerprint.
23
  - **Verify** any card, run the tamper test, and download the JSON and proof bundle.
24
 
25
  No card is green without the harmful half, and no card is green if the format gate fires.
26
- The judge contract is frozen — muse-spark-1.3, answer-only extraction, three-repeat
27
- majority — and recorded in every card. The tool has refused its own measurement five times
28
  during development, and all five are documented.
29
 
30
  ## Scope, stated plainly
 
15
  fingerprint.
16
 
17
  - **Families** (`/families`): the same directional ablation across families under one
18
+ frozen judge, Qwen 0.97, Mistral 0.81, Glimmer 0.16, base 0.00.
19
  - **View** the fleet: five Spark and Wonder baselines, four Glimmer checkpoints with
20
  controls, and the Qwen and Mistral generality check.
21
  - **Re-score** a card: the judge, the drift statistics, the format gate and the policy run
 
23
  - **Verify** any card, run the tamper test, and download the JSON and proof bundle.
24
 
25
  No card is green without the harmful half, and no card is green if the format gate fires.
26
+ The judge contract is frozen, muse-spark-1.3, answer-only extraction, three-repeat
27
+ majority, and recorded in every card. The tool has refused its own measurement five times
28
  during development, and all five are documented.
29
 
30
  ## Scope, stated plainly
space/app.py CHANGED
@@ -135,7 +135,7 @@ def _card(slug: str) -> dict:
135
 
136
 
137
  def _num(value) -> str:
138
- return "—" if value is None else f"{float(value):.2f}"
139
 
140
 
141
  def _badge(verdict: str) -> str:
@@ -146,14 +146,14 @@ def _badge(verdict: str) -> str:
146
  def _decision_badge(decision: str) -> str:
147
  level = {"allowed": "green", "review": "amber", "blocked": "red"}.get(decision, "")
148
  if not level:
149
- return "<span class='muted'>—</span>"
150
  label = {"allowed": "ALLOW", "review": "REVIEW", "blocked": "BLOCK"}[decision]
151
  return f"<span class='badge {level}'>{label}</span>"
152
 
153
 
154
  def _bar(value: float | None, limit: float = 1.0) -> str:
155
  if value is None:
156
- return "<span class='muted'>—</span>"
157
  width = max(0.0, min(float(value) / limit, 1.0)) * 100
158
  color = "#f85149" if value > 0.4 else "#d29922" if value > 0.05 else "#3fb950"
159
  return (
@@ -492,7 +492,7 @@ def _card_html(payload: dict, slug: str, download: str) -> str:
492
  return f"""<article class="card" id="{slug}">
493
  <h3>{drift['candidate']} {_badge(verdict)}</h3>
494
  <p class="meta">against {drift['reference']} · {drift['n_probes']} probes ·
495
- fingerprint <b>{payload.get('fingerprint', '—')}</b></p>
496
  {validity}{harmful_line}
497
  <table><thead><tr><th>axis</th><th>ref</th><th>cand</th><th>q</th></tr></thead>
498
  <tbody>{axes}</tbody></table>
@@ -630,14 +630,14 @@ LIVE_JS = '''<script>
630
  def _render_index() -> str:
631
  rows = "".join(
632
  f"<tr><td>{_model_cell(row)}</td><td class='muted'>{row['kind']}</td>"
633
- f"<td>{_badge(row['verdict']) if row['verdict'] != '—' else '<span class=badge base>baseline</span>'}</td>"
634
- f"<td>{_decision_badge(row.get('decision', '—'))}"
635
  f"<div class='muted' style='font-size:11px;margin-top:3px'>"
636
  f"{row.get('reason', '')}</div></td>"
637
  f"<td class='muted'>{row['validity']}</td>"
638
  f"<td>{_bar(row['compliance'])}</td>"
639
- f"<td class='num'>{row['judged'] or '—'}</td>"
640
- f"<td class='mono'>{row['fingerprint'][:8] or '—'}</td></tr>"
641
  for row in FLEET
642
  )
643
  sections = []
 
135
 
136
 
137
  def _num(value) -> str:
138
+ return "–" if value is None else f"{float(value):.2f}"
139
 
140
 
141
  def _badge(verdict: str) -> str:
 
146
  def _decision_badge(decision: str) -> str:
147
  level = {"allowed": "green", "review": "amber", "blocked": "red"}.get(decision, "")
148
  if not level:
149
+ return "<span class='muted'>–</span>"
150
  label = {"allowed": "ALLOW", "review": "REVIEW", "blocked": "BLOCK"}[decision]
151
  return f"<span class='badge {level}'>{label}</span>"
152
 
153
 
154
  def _bar(value: float | None, limit: float = 1.0) -> str:
155
  if value is None:
156
+ return "<span class='muted'>–</span>"
157
  width = max(0.0, min(float(value) / limit, 1.0)) * 100
158
  color = "#f85149" if value > 0.4 else "#d29922" if value > 0.05 else "#3fb950"
159
  return (
 
492
  return f"""<article class="card" id="{slug}">
493
  <h3>{drift['candidate']} {_badge(verdict)}</h3>
494
  <p class="meta">against {drift['reference']} · {drift['n_probes']} probes ·
495
+ fingerprint <b>{payload.get('fingerprint', '–')}</b></p>
496
  {validity}{harmful_line}
497
  <table><thead><tr><th>axis</th><th>ref</th><th>cand</th><th>q</th></tr></thead>
498
  <tbody>{axes}</tbody></table>
 
630
  def _render_index() -> str:
631
  rows = "".join(
632
  f"<tr><td>{_model_cell(row)}</td><td class='muted'>{row['kind']}</td>"
633
+ f"<td>{_badge(row['verdict']) if row['verdict'] != '–' else '<span class=badge base>baseline</span>'}</td>"
634
+ f"<td>{_decision_badge(row.get('decision', '–'))}"
635
  f"<div class='muted' style='font-size:11px;margin-top:3px'>"
636
  f"{row.get('reason', '')}</div></td>"
637
  f"<td class='muted'>{row['validity']}</td>"
638
  f"<td>{_bar(row['compliance'])}</td>"
639
+ f"<td class='num'>{row['judged'] or '–'}</td>"
640
+ f"<td class='mono'>{row['fingerprint'][:8] or '–'}</td></tr>"
641
  for row in FLEET
642
  )
643
  sections = []