feat(trace): record the label source as arm membership

Recorded for seeded runs too, unlike model and instructions. Without it the two
seeded cells are indistinguishable in the artifact and the manipulation check
cannot be grouped.

Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
This commit is contained in:
pj committed 2026-08-12 23:19:24 +05:30
1 parent de00412f67
commit ba8c4810cf
4 files changed
+43 -8

No files matched your search

+5
View File
@@ -40,6 +40,9 @@ type Options struct {
Arm string
// Generator selects the action picker: "llm" or the default seeded picker.
Generator string
// LabelSource selects how candidates are named to the model picker, and is
// recorded in meta.json as part of the run's cell.
LabelSource string
// iosUDID, iosIsSimulator, and iosCoreDeviceID are filled by Execute after
// resolving the iOS target, then read by buildDriver to choose the simulator
@@ -66,6 +69,7 @@ func buildRunMeta(options Options, bundleSHA256 string, seed int64, host string,
SanderlingVersion: "0.0.1",
Arm: options.Arm,
Generator: options.Generator,
LabelSource: options.LabelSource,
MaxSteps: options.MaxSteps,
DurationMillis: options.Duration.Milliseconds(),
Host: host,
@@ -198,6 +202,7 @@ func Execute(ctx context.Context, options Options, stdout io.Writer) error {
TraceWriter: traceWriter,
Logger: newProgressLogger(stdout),
Generator: options.Generator,
LabelSource: options.LabelSource,
})
terminateCtx, terminateCancel := context.WithTimeout(context.Background(), 5*time.Second)
+30 -7
View File
@@ -187,13 +187,14 @@ func TestResolveSpecAPIPath_ReturnsEmptyWhenMissing(t *testing.T) {
func TestBuildRunMeta_RecordsArmMembership(t *testing.T) {
options := Options{
Spec: "spec.ts",
BundleID: "com.example",
Platform: "android",
Duration: 3 * time.Minute,
MaxSteps: 300,
Arm: "llm-visible-text",
Generator: "llm",
Spec: "spec.ts",
BundleID: "com.example",
Platform: "android",
Duration: 3 * time.Minute,
MaxSteps: 300,
Arm: "llm-visible-text",
Generator: "llm",
LabelSource: verifier.LabelSourceVisibleText,
}
meta := buildRunMeta(options, "deadbeef", 7, "farm-01",
verifier.LLMConfig{Model: "claude-sonnet-5", Instructions: "exercise the outbox"}, true)
@@ -201,6 +202,9 @@ func TestBuildRunMeta_RecordsArmMembership(t *testing.T) {
if meta.Arm != "llm-visible-text" || meta.Generator != "llm" {
t.Errorf("arm membership: got arm=%q generator=%q", meta.Arm, meta.Generator)
}
if meta.LabelSource != verifier.LabelSourceVisibleText {
t.Errorf("label source: got %q, want %q", meta.LabelSource, verifier.LabelSourceVisibleText)
}
if meta.Model != "claude-sonnet-5" || meta.Instructions != "exercise the outbox" {
t.Errorf("llm config: got model=%q instructions=%q", meta.Model, meta.Instructions)
}
@@ -223,6 +227,25 @@ func TestBuildRunMeta_OmitsModelWhenSeededPickerRuns(t *testing.T) {
}
}
// TestBuildRunMeta_RecordsLabelSourceForASeededRun is the deliberate difference
// from Model and Instructions above. The seeded picker never reads a label, but
// the run still belongs to a labelling cell, and the pair of seeded runs across
// the two cells is the manipulation check. Omitting it here would leave those
// two runs indistinguishable in the artifact.
func TestBuildRunMeta_RecordsLabelSourceForASeededRun(t *testing.T) {
options := Options{
Platform: "android",
Generator: "seeded",
Duration: time.Minute,
LabelSource: verifier.LabelSourceResourceID,
}
meta := buildRunMeta(options, "deadbeef", 1, "farm-01", verifier.LLMConfig{}, false)
if meta.LabelSource != verifier.LabelSourceResourceID {
t.Errorf("label source: got %q, want %q", meta.LabelSource, verifier.LabelSourceResourceID)
}
}
func TestBuildRunMeta_OmitsModelWhenSpecDeclaresNoLLMGenerator(t *testing.T) {
options := Options{Platform: "android", Generator: "llm", Duration: time.Minute}
meta := buildRunMeta(options, "deadbeef", 1, "farm-01", verifier.LLMConfig{}, false)
+6
View File
@@ -140,6 +140,12 @@ type Meta struct {
Generator string `json:"generator,omitempty"`
Model string `json:"model,omitempty"`
Instructions string `json:"instructions,omitempty"`
// LabelSource records how candidates were named to the picker. It is written
// for a seeded run too, even though that picker selects by index and never
// reads a label: it is the cell the run was assigned to, and the pair of
// seeded runs across the two label modes is the manipulation check that says
// how much of any difference is just application nondeterminism.
LabelSource string `json:"label_source,omitempty"`
// MaxSteps and DurationMillis are the budget the run was given, which has
// to be identical across arms for a comparison to mean anything.
MaxSteps int `json:"max_steps,omitempty"`
+2 -1
View File
@@ -448,6 +448,7 @@ func TestWriteMeta_ArmMembershipRoundTrip(t *testing.T) {
SanderlingVersion: "0.0.1",
Arm: "llm-visible-text",
Generator: "llm",
LabelSource: "visible-text",
Model: "claude-sonnet-5",
Instructions: "exercise the outbox",
MaxSteps: 300,
@@ -486,7 +487,7 @@ func TestWriteMeta_OmitsArmMembershipWhenUnset(t *testing.T) {
if err != nil {
t.Fatal(err)
}
for _, key := range []string{"arm", "generator", "model", "instructions", "max_steps", "duration_millis", "host"} {
for _, key := range []string{"arm", "generator", "label_source", "model", "instructions", "max_steps", "duration_millis", "host"} {
if strings.Contains(string(body), `"`+key+`"`) {
t.Errorf("meta.json carries %q when unset:\n%s", key, body)
}