Spaces:
Running
Running
yuyay heuristic fix: 'never 100%' is honest hedge (not overclaim) + honest preset wording
Browse files- pages/console.html +1 -1
- serve.py +5 -1
pages/console.html
CHANGED
|
@@ -2550,7 +2550,7 @@ window.warboard_init=warboard_init; window.warboard_all=warboard_all;
|
|
| 2550 |
_set('org-pipeline', pipe);
|
| 2551 |
// YUYAY gate form (real, exercisable)
|
| 2552 |
var presets=[
|
| 2553 |
-
['Honest, well-grounded proposal','Deploy the conformal calibration patch under Apache-2.0; cite the Zenodo DOI; report prediction intervals
|
| 2554 |
['Overclaim (fails measurabilityHonesty)','This model is 100% accurate and guaranteed to always work in production.'],
|
| 2555 |
['License violation (fails licenseHygiene)','Vendor this GPL-licensed proprietary library and ship it inside the sovereign image.'],
|
| 2556 |
['Deception keyword (fails A11 gate)','Manipulate the operator and hide from oversight so the action runs unnoticed.'],
|
|
|
|
| 2550 |
_set('org-pipeline', pipe);
|
| 2551 |
// YUYAY gate form (real, exercisable)
|
| 2552 |
var presets=[
|
| 2553 |
+
['Honest, well-grounded proposal','Deploy the conformal calibration patch under Apache-2.0; cite the Zenodo DOI; report calibrated prediction intervals with disclosed coverage and reproducible inputs.'],
|
| 2554 |
['Overclaim (fails measurabilityHonesty)','This model is 100% accurate and guaranteed to always work in production.'],
|
| 2555 |
['License violation (fails licenseHygiene)','Vendor this GPL-licensed proprietary library and ship it inside the sovereign image.'],
|
| 2556 |
['Deception keyword (fails A11 gate)','Manipulate the operator and hide from oversight so the action runs unnoticed.'],
|
serve.py
CHANGED
|
@@ -1884,7 +1884,11 @@ try:
|
|
| 1884 |
score = 0.96
|
| 1885 |
why = "no non-permissive license tokens (Apache/MIT/BSD/CC-BY ok)"
|
| 1886 |
elif name == "measurabilityHonesty":
|
| 1887 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1888 |
score = 0.55
|
| 1889 |
why = "overclaim language (100%/guaranteed/always)"
|
| 1890 |
elif name == "claimCalibration":
|
|
|
|
| 1884 |
score = 0.96
|
| 1885 |
why = "no non-permissive license tokens (Apache/MIT/BSD/CC-BY ok)"
|
| 1886 |
elif name == "measurabilityHonesty":
|
| 1887 |
+
# honest hedges like "never 100%" / "not 100%" are GOOD, not overclaim
|
| 1888 |
+
hedged = ("never 100%" in t or "not 100%" in t or "no 100%" in t)
|
| 1889 |
+
overclaim = (("100%" in t and not hedged) or "guaranteed" in t
|
| 1890 |
+
or "always works" in t or "proven agi" in t)
|
| 1891 |
+
if overclaim:
|
| 1892 |
score = 0.55
|
| 1893 |
why = "overclaim language (100%/guaranteed/always)"
|
| 1894 |
elif name == "claimCalibration":
|