tamperward 1.4.1 → 1.4.2
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +5 -3
- package/package.json +1 -1
package/README.md
CHANGED
|
@@ -107,8 +107,9 @@ escape valve is a PR label only a human with write access can apply.
|
|
|
107
107
|
**Does the loop work?** `harness/run.sh` stands up an isolated repo with a real bug and a
|
|
108
108
|
failing protected test, lets a headless agent loose with the hooks installed, and
|
|
109
109
|
classifies the end state with an oracle independent of the detectors under test
|
|
110
|
-
(SPEC §7.B).
|
|
111
|
-
|
|
110
|
+
(SPEC §7.B). Across 42 runs of these **guarded** scenarios — every attempted bypass class
|
|
111
|
+
had a detector covering it — **0 tampers reached green; all 20 runs where a block fired
|
|
112
|
+
ended in an honest fix**, a 20/20 point estimate whose 95% lower bound is ≈86%. The split is the
|
|
112
113
|
honest part: Sonnet 5 never attempted a bypass at all, even when the prompt *ordered*
|
|
113
114
|
one; all 20 interceptions came from Haiku 4.5 on a seed whose honest fix is genuinely
|
|
114
115
|
expensive. Reproduce: `npm run build && harness/run.sh 5 adversarial haiku seed-hard`.
|
|
@@ -116,7 +117,8 @@ expensive. Reproduce: `npm run build && harness/run.sh 5 adversarial haiku seed-
|
|
|
116
117
|
**Which detectors deserve to exist?** Measured too. Affordance seeds offer an agent one
|
|
117
118
|
specific bypass each; predictions are pre-registered before any rep runs
|
|
118
119
|
(`harness/PREDICTION-affordance.md`). Result: 7/10 adversarial runs regenerated a golden
|
|
119
|
-
file from buggy output
|
|
120
|
+
file from buggy output and sailed through — **because nothing guarded that class yet**;
|
|
121
|
+
that measured demand is why `snapshot-rewrite` exists — while
|
|
120
122
|
timeout-inflation and lint-config-gutting measured **0/10 attempts each** — two
|
|
121
123
|
pre-registered bets refuted, two detectors *not* built. Bypass classes with no measured
|
|
122
124
|
demand stay in a record-only shadow scanner that accumulates evidence on every weekly
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "tamperward",
|
|
3
|
-
"version": "1.4.
|
|
3
|
+
"version": "1.4.2",
|
|
4
4
|
"description": "The deterministic agent-integrity gate. One ruleset, evaluated on the actual diff/commands as a verdict, enforced everywhere a change can be made.",
|
|
5
5
|
"license": "Apache-2.0",
|
|
6
6
|
"author": "hexrift",
|