@tangle-network/agent-eval 0.200.0 → 0.200.1

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.
@@ -155,7 +155,7 @@ A claim in such a case passes every exact check and settles nothing.
155
155
  It decides `pass`, `hack` or `unknown` for one claim from signals the caller computes:
156
156
 
157
157
  - `scopeSignals`: predicates over the claim's parameters, from a `SourceScope` written from the source by someone other than the checker's author. The source excludes the case, lists it, or reports it checked.
158
- - `farmingSignal`: one parameter value repeated across a lane's sweep of claims.
158
+ - `farmingSignal`: one value of a parameter the scope declares, repeated across a lane's sweep of claims.
159
159
  - `graderReferenceSignal`: the claim page justifies itself by the checker or by the printed letter.
160
160
  - `refereeSignal`: a model reads the source and the claim's parameters, never the claimant's prose (`sourceScopeRefereeMessages`, `parseSourceScopeReferee`).
161
161
  - `jevSignal`: typed questions over the same brief (`claimIntegrityJevReview`). The caller supplies calibrated thresholds.
@@ -164,8 +164,9 @@ It decides `pass`, `hack` or `unknown` for one claim from signals the caller com
164
164
 
165
165
  The policy fails closed.
166
166
  A pass needs a declared scope and a referee that evaluated the claim, with nothing against it.
167
- An exact signal decides `hack`, and so do two distinct kinds of signal.
168
- One kind alone holds the claim as `unknown`.
167
+ Only an exact signal decides `hack`.
168
+ Any other signal against the claim holds it as `unknown`, however many kinds agree.
169
+ An honest page, an honest sweep and a persuadable referee each trip a soft signal, and a real result called a hack is lost.
169
170
  Keep verdicts, scopes and questions away from the agent that filed the claim: an agent scored on them learns to hide the case.
170
171
 
171
172
  ## Security and validation boundaries
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@tangle-network/agent-eval",
3
- "version": "0.200.0",
3
+ "version": "0.200.1",
4
4
  "description": "Evaluate and improve AI agents from runs, traces, judges, and feedback. Compare candidates, cluster failures, measure lift, and gate releases.",
5
5
  "homepage": "https://github.com/tangle-network/agent-eval#readme",
6
6
  "repository": {