diff --git a/cmd/internal-tools/oracle-reduction/load.go b/cmd/internal-tools/oracle-reduction/load.go
new file mode 100644
index 0000000..69358a8
--- /dev/null
+++ b/cmd/internal-tools/oracle-reduction/load.go
@@ -0,0 +1,223 @@
+package main
+
+import (
+	"bufio"
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"sort"
+
+	"github.com/priyanshujain/sanderling/internal/trace"
+	"github.com/priyanshujain/sanderling/internal/verifier"
+)
+
+const maxStepBytes = 64 * 1024 * 1024
+
+// loadedRun is one run directory: its meta, the steps in file order, and the
+// synthetic end-of-run record the runner writes when Finalize convicted a
+// liveness obligation.
+type loadedRun struct {
+	Directory string
+	Meta      trace.Meta
+	Steps     []trace.Step
+	Finalize  *trace.Step
+}
+
+// loadRun reads a run directory and refuses anything E3 cannot replay. A step
+// written before the format change carries no element depths, so its hierarchy
+// decodes with a nil root and every selector resolves to nothing: the refusal
+// has to name the version rather than let the replay report an empty screen.
+func loadRun(directory string) (loadedRun, error) {
+	metaBody, err := os.ReadFile(filepath.Join(directory, "meta.json"))
+	if err != nil {
+		return loadedRun{}, fmt.Errorf("read meta: %w", err)
+	}
+	var meta trace.Meta
+	if err := json.Unmarshal(metaBody, &meta); err != nil {
+		return loadedRun{}, fmt.Errorf("decode meta: %w", err)
+	}
+
+	file, err := os.Open(filepath.Join(directory, "trace.jsonl"))
+	if err != nil {
+		return loadedRun{}, fmt.Errorf("open trace: %w", err)
+	}
+	defer file.Close()
+
+	run := loadedRun{Directory: directory, Meta: meta}
+	scanner := bufio.NewScanner(file)
+	scanner.Buffer(make([]byte, 0, 1024*1024), maxStepBytes)
+	line := 0
+	for scanner.Scan() {
+		line++
+		if len(scanner.Bytes()) == 0 {
+			continue
+		}
+		var step trace.Step
+		if err := json.Unmarshal(scanner.Bytes(), &step); err != nil {
+			return loadedRun{}, fmt.Errorf(
+				"decode step on line %d: %w",
+				line,
+				err,
+			)
+		}
+		if step.TraceVersion != trace.TraceVersion {
+			return loadedRun{}, fmt.Errorf(
+				"step %d is trace_version %d, and E3 replays version %d only: "+
+					"an older step stores no element depths, so its hierarchy decodes "+
+					"with a nil root and every selector resolves to nothing on replay",
+				step.Index, step.TraceVersion, trace.TraceVersion)
+		}
+		run.Steps = append(run.Steps, step)
+	}
+	if err := scanner.Err(); err != nil {
+		return loadedRun{}, fmt.Errorf("read trace: %w", err)
+	}
+	if len(run.Steps) == 0 {
+		return loadedRun{}, fmt.Errorf("trace has no steps")
+	}
+	if last := run.Steps[len(run.Steps)-1]; len(run.Steps) > 1 &&
+		last.Hierarchy == nil &&
+		len(last.Violations) > 0 {
+		run.Finalize = &last
+		run.Steps = run.Steps[:len(run.Steps)-1]
+	}
+	if meta.Seed == 0 {
+		return loadedRun{}, fmt.Errorf(
+			"meta records no seed, so the run's bundle cannot be reproduced",
+		)
+	}
+	return run, nil
+}
+
+// finalizeIndex is the step index the runner gives its end-of-run record: one
+// past the last step it wrote.
+func (r loadedRun) finalizeIndex() int {
+	return r.Steps[len(r.Steps)-1].Index + 1
+}
+
+// extractorFold reconstructs every extractor's value at each step from the
+// recorded per-step diffs. A run's trace stores what changed, so the value an
+// extractor held at a step is the last change at or before it; an extractor
+// that never changed held JSON null throughout, which is what an unwritten
+// diff means.
+func extractorFold(
+	steps []trace.Step,
+	names []string,
+) ([]map[int]json.RawMessage, error) {
+	index := make(map[string]int, len(names))
+	for position, name := range names {
+		index[name] = position
+	}
+	current := make(map[int]json.RawMessage, len(names))
+	for position := range names {
+		current[position] = json.RawMessage("null")
+	}
+	folded := make([]map[int]json.RawMessage, len(steps))
+	for step := range steps {
+		for name, change := range steps[step].ExtractorChanges {
+			position, ok := index[name]
+			if !ok {
+				return nil, fmt.Errorf(
+					"step %d records extractor %q, which the spec does not register; "+
+						"the trace and the spec are not the same bundle",
+					steps[step].Index,
+					name,
+				)
+			}
+			current[position] = change.Curr
+		}
+		snapshot := make(map[int]json.RawMessage, len(current))
+		for position, value := range current {
+			snapshot[position] = value
+		}
+		folded[step] = snapshot
+	}
+	return folded, nil
+}
+
+// lastActionFor rebuilds the action the runner had applied before the next
+// step observed. An action the runner chose but never dispatched left
+// state.lastAction null, and the recorded skip reason is what says so.
+func lastActionFor(step trace.Step) *verifier.Action {
+	if step.NextAction == nil || step.ActionSkipped != "" {
+		return nil
+	}
+	recorded := *step.NextAction
+	action := verifier.Action{
+		Kind:           verifier.ActionKind(recorded.Kind),
+		On:             recorded.Selector,
+		Text:           recorded.Text,
+		X:              recorded.X,
+		Y:              recorded.Y,
+		FromX:          recorded.FromX,
+		FromY:          recorded.FromY,
+		ToX:            recorded.ToX,
+		ToY:            recorded.ToY,
+		Key:            recorded.Key,
+		DurationMillis: recorded.DurationMillis,
+	}
+	return &action
+}
+
+func traceLogs(entries []trace.LogEntry) []verifier.LogEntry {
+	if len(entries) == 0 {
+		return nil
+	}
+	logs := make([]verifier.LogEntry, 0, len(entries))
+	for _, entry := range entries {
+		logs = append(logs, verifier.LogEntry{
+			UnixMillis: entry.UnixMillis,
+			Level:      entry.Level,
+			Tag:        entry.Tag,
+			Message:    entry.Message,
+		})
+	}
+	return logs
+}
+
+func traceExceptions(entries []trace.Exception) []verifier.Exception {
+	if len(entries) == 0 {
+		return nil
+	}
+	exceptions := make([]verifier.Exception, 0, len(entries))
+	for _, entry := range entries {
+		exceptions = append(exceptions, verifier.Exception{
+			Class:      entry.Class,
+			Message:    entry.Message,
+			StackTrace: entry.StackTrace,
+			UnixMillis: entry.UnixMillis,
+		})
+	}
+	return exceptions
+}
+
+// discoverRuns finds every run directory at or below root, a run directory
+// being one holding both meta.json and trace.jsonl.
+func discoverRuns(root string) ([]string, error) {
+	var directories []string
+	err := filepath.Walk(
+		root,
+		func(path string, info os.FileInfo, err error) error {
+			if err != nil {
+				return err
+			}
+			if !info.IsDir() {
+				return nil
+			}
+			if _, statErr := os.Stat(filepath.Join(path, "trace.jsonl")); statErr != nil {
+				return nil
+			}
+			if _, statErr := os.Stat(filepath.Join(path, "meta.json")); statErr != nil {
+				return nil
+			}
+			directories = append(directories, path)
+			return nil
+		},
+	)
+	if err != nil {
+		return nil, err
+	}
+	sort.Strings(directories)
+	return directories, nil
+}
diff --git a/cmd/internal-tools/oracle-reduction/load_test.go b/cmd/internal-tools/oracle-reduction/load_test.go
new file mode 100644
index 0000000..7ddea11
--- /dev/null
+++ b/cmd/internal-tools/oracle-reduction/load_test.go
@@ -0,0 +1,162 @@
+package main
+
+import (
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/priyanshujain/sanderling/internal/trace"
+	"github.com/priyanshujain/sanderling/internal/verifier"
+)
+
+func writeRunFiles(t *testing.T, directory, meta string, steps ...string) {
+	t.Helper()
+	if err := os.WriteFile(filepath.Join(directory, "meta.json"), []byte(meta), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	body := strings.Join(steps, "\n") + "\n"
+	if err := os.WriteFile(filepath.Join(directory, "trace.jsonl"), []byte(body), 0o644); err != nil {
+		t.Fatal(err)
+	}
+}
+
+func TestLoadRunRefusesAStepFromBeforeTheFormatChange(t *testing.T) {
+	directory := t.TempDir()
+	writeRunFiles(
+		t,
+		directory,
+		`{"seed": 3}`,
+		`{"step":1,"timestamp":"2026-08-15T12:00:00Z"}`,
+	)
+
+	_, err := loadRun(directory)
+
+	if err == nil {
+		t.Fatal("a version-0 step must be refused")
+	}
+	if !strings.Contains(err.Error(), "trace_version 0") ||
+		!strings.Contains(err.Error(), "depths") {
+		t.Errorf(
+			"the refusal must name the version and why it cannot replay: %v",
+			err,
+		)
+	}
+}
+
+func TestLoadRunRefusesARunWithNoSeed(t *testing.T) {
+	directory := t.TempDir()
+	writeRunFiles(
+		t,
+		directory,
+		`{}`,
+		`{"step":1,"trace_version":1,"timestamp":"2026-08-15T12:00:00Z"}`,
+	)
+
+	_, err := loadRun(directory)
+
+	if err == nil || !strings.Contains(err.Error(), "seed") {
+		t.Errorf("a run without a seed cannot be bundled as it was: %v", err)
+	}
+}
+
+func TestLoadRunSplitsOffTheEndOfRunRecord(t *testing.T) {
+	directory := t.TempDir()
+	writeRunFiles(
+		t,
+		directory,
+		`{"seed": 3}`,
+		`{"step":1,"trace_version":1,"timestamp":"2026-08-15T12:00:00Z","hierarchy":{"elements":[],"depths":[]}}`,
+		`{"step":2,"trace_version":1,"timestamp":"2026-08-15T12:00:01Z","violations":["reachable"]}`,
+	)
+
+	run, err := loadRun(directory)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	if len(run.Steps) != 1 {
+		t.Fatalf("observation steps: got %d, want 1", len(run.Steps))
+	}
+	if run.Finalize == nil || run.Finalize.Index != 2 {
+		t.Fatalf("finalize record: %+v", run.Finalize)
+	}
+	if run.finalizeIndex() != 2 {
+		t.Errorf("finalize index: got %d, want 2", run.finalizeIndex())
+	}
+}
+
+func TestExtractorFoldCarriesAValueForwardUntilItChanges(t *testing.T) {
+	steps := []trace.Step{
+		{Index: 1, ExtractorChanges: map[string]trace.ExtractorChange{
+			"count": {
+				Prev: json.RawMessage("null"),
+				Curr: json.RawMessage("1"),
+			},
+		}},
+		{Index: 2},
+		{Index: 3, ExtractorChanges: map[string]trace.ExtractorChange{
+			"count": {Prev: json.RawMessage("1"), Curr: json.RawMessage("4")},
+		}},
+	}
+
+	folded, err := extractorFold(steps, []string{"count", "unseen"})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	want := []string{"1", "1", "4"}
+	for position, expected := range want {
+		if got := string(folded[position][0]); got != expected {
+			t.Errorf(
+				"count at step %d: got %s, want %s",
+				position+1,
+				got,
+				expected,
+			)
+		}
+		if got := string(folded[position][1]); got != "null" {
+			t.Errorf(
+				"an extractor that never changed must fold to null, got %s",
+				got,
+			)
+		}
+	}
+}
+
+func TestExtractorFoldRefusesAValueTheSpecCannotPlace(t *testing.T) {
+	steps := []trace.Step{
+		{Index: 1, ExtractorChanges: map[string]trace.ExtractorChange{
+			"gone": {Curr: json.RawMessage("1")},
+		}},
+	}
+
+	_, err := extractorFold(steps, []string{"count"})
+
+	if err == nil || !strings.Contains(err.Error(), "gone") {
+		t.Errorf("an unplaceable extractor value must be refused: %v", err)
+	}
+}
+
+func TestLastActionIsEmptyWhenTheRunnerNeverDispatchedIt(t *testing.T) {
+	dispatched := trace.Step{
+		NextAction: &trace.Action{Kind: "Tap", Selector: "id:save", X: 4, Y: 5},
+	}
+	skipped := trace.Step{
+		NextAction:    &trace.Action{Kind: "Tap", Selector: "id:save"},
+		ActionSkipped: foregroundLossReason,
+	}
+
+	action := lastActionFor(dispatched)
+	if action == nil || action.Kind != verifier.ActionKindTap ||
+		action.On != "id:save" ||
+		action.X != 4 {
+		t.Fatalf("dispatched action: %+v", action)
+	}
+	if lastActionFor(skipped) != nil {
+		t.Error(
+			"an action the runner threw away never reached state.lastAction",
+		)
+	}
+}
diff --git a/cmd/internal-tools/oracle-reduction/main.go b/cmd/internal-tools/oracle-reduction/main.go
new file mode 100644
index 0000000..8e5b00f
--- /dev/null
+++ b/cmd/internal-tools/oracle-reduction/main.go
@@ -0,0 +1,312 @@
+// Command oracle-reduction re-evaluates stored traces offline under four
+// oracles and reports what each one refutes: the full engine, a crash-only
+// detector, a single-state check, and a single-step property triple. The
+// oracles vary while the traces stay fixed, which is what separates a defect an
+// oracle cannot express from one an explorer never reached.
+//
+// The offline engine has to reproduce the verdicts each run recorded. A
+// disagreement is a bug here or a gap in the trace, so it is reported as a
+// mismatch and exits nonzero rather than being counted as a finding.
+package main
+
+import (
+	"encoding/json"
+	"flag"
+	"fmt"
+	"io"
+	"os"
+	"sort"
+	"strings"
+
+	"github.com/priyanshujain/sanderling/internal/testrun"
+)
+
+const usage = `oracle-reduction replays stored traces under reduced oracles.
+
+Usage:
+  oracle-reduction --runs <dir> [--output <path>] [--spec <path>]
+
+--runs is scanned recursively; every directory holding meta.json and
+trace.jsonl is one trace. Each trace's spec is bundled from the path its
+meta.json records unless --spec overrides it.
+
+Exit status is 2 when the offline engine disagreed with any run's recorded
+verdicts, which blocks the experiment rather than reporting the difference as
+noise.
+
+A reduced oracle whose rewrite no longer states a property reports
+cannot_express for it rather than a verdict, and the property is temporal-only
+when the other reductions also fail to state or refute it. A property whose
+window is longer than the two observations a triple spans is reported with
+single_step_truncates_window, which is the form that decides it.
+`
+
+type config struct {
+	runsRoot   string
+	outputPath string
+	specPath   string
+	// hostExtractors re-runs the spec's extractor getters over each stored
+	// hierarchy instead of replaying the values a web run's page computed. It
+	// asks a different question of the same trace: whether the stored tree
+	// alone carries what the properties read.
+	hostExtractors bool
+}
+
+func parseArguments(arguments []string, stderr io.Writer) (config, error) {
+	flagSet := flag.NewFlagSet("oracle-reduction", flag.ContinueOnError)
+	flagSet.SetOutput(stderr)
+	flagSet.Usage = func() {
+		fmt.Fprint(stderr, usage)
+		flagSet.PrintDefaults()
+	}
+	var configuration config
+	flagSet.StringVar(
+		&configuration.runsRoot,
+		"runs",
+		"",
+		"directory tree holding the run directories to replay (required)",
+	)
+	flagSet.StringVar(
+		&configuration.outputPath,
+		"output",
+		"",
+		"file to write the per-trace JSONL to (default stdout)",
+	)
+	flagSet.StringVar(
+		&configuration.specPath,
+		"spec",
+		"",
+		"spec to bundle instead of the one each meta.json records",
+	)
+	flagSet.BoolVar(
+		&configuration.hostExtractors,
+		"host-extractors",
+		false,
+		"re-run the spec's extractors over each stored hierarchy instead of replaying the values a web run's page computed",
+	)
+	if err := flagSet.Parse(arguments); err != nil {
+		return config{}, err
+	}
+	if configuration.runsRoot == "" {
+		return config{}, fmt.Errorf("--runs is required")
+	}
+	return configuration, nil
+}
+
+func main() {
+	configuration, err := parseArguments(os.Args[1:], os.Stderr)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "oracle-reduction: %v\n", err)
+		os.Exit(1)
+	}
+	code, err := run(configuration, os.Stdout, os.Stderr)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "oracle-reduction: %v\n", err)
+		os.Exit(1)
+	}
+	os.Exit(code)
+}
+
+func run(configuration config, stdout, stderr io.Writer) (int, error) {
+	directories, err := discoverRuns(configuration.runsRoot)
+	if err != nil {
+		return 1, err
+	}
+	if len(directories) == 0 {
+		return 1, fmt.Errorf(
+			"no run directories under %s",
+			configuration.runsRoot,
+		)
+	}
+
+	output := stdout
+	if configuration.outputPath != "" {
+		file, createErr := os.Create(configuration.outputPath)
+		if createErr != nil {
+			return 1, createErr
+		}
+		defer file.Close()
+		output = file
+	}
+	encoder := json.NewEncoder(output)
+
+	var reports []runReport
+	rejected := 0
+	for _, directory := range directories {
+		loaded, loadErr := loadRun(directory)
+		if loadErr != nil {
+			rejected++
+			fmt.Fprintf(stderr, "skipped %s: %v\n", directory, loadErr)
+			continue
+		}
+		specPath := loaded.Meta.SpecPath
+		if configuration.specPath != "" {
+			specPath = configuration.specPath
+		}
+		bundle, bundleErr := testrun.BundleSpec(specPath, loaded.Meta.Seed)
+		if bundleErr != nil {
+			return 1, fmt.Errorf(
+				"%s: bundle %s: %w",
+				directory,
+				specPath,
+				bundleErr,
+			)
+		}
+		if loaded.Meta.BundleSHA256 != "" &&
+			bundle.SHA256 != loaded.Meta.BundleSHA256 {
+			// The bundler writes each module's path into the output relative to
+			// the working directory, so this differs whenever the replay is
+			// invoked from somewhere else than the run was. What a changed spec
+			// would actually break is caught by the property set, the extractor
+			// names and the residual comparison below.
+			fmt.Fprintf(stderr,
+				"note: %s bundles to %s here and the run recorded %s\n",
+				directory, bundle.SHA256[:12], loaded.Meta.BundleSHA256[:12])
+		}
+		report, replayErr := replay(
+			loaded,
+			string(bundle.JavaScript),
+			configuration.hostExtractors,
+		)
+		if replayErr != nil {
+			return 1, fmt.Errorf("%s: %w", directory, replayErr)
+		}
+		if err := encoder.Encode(report); err != nil {
+			return 1, err
+		}
+		reports = append(reports, report)
+	}
+
+	summarize(reports, rejected, stderr)
+	for _, report := range reports {
+		if !report.Valid {
+			return 2, nil
+		}
+	}
+	if len(reports) == 0 {
+		return 1, fmt.Errorf(
+			"every run directory was rejected; nothing was replayed",
+		)
+	}
+	return 0, nil
+}
+
+func summarize(reports []runReport, rejected int, out io.Writer) {
+	invalid := 0
+	crashed := 0
+	weakest := map[string]int{}
+	byClass := map[string]map[string]int{}
+	inexpressible := map[string]map[string]bool{}
+	unmatched := 0
+	engineRefutations := 0
+	for _, report := range reports {
+		if !report.Valid {
+			invalid++
+		}
+		if report.CrashOnly.Fired {
+			crashed++
+		}
+		for _, property := range report.Properties {
+			recordInexpressible(
+				inexpressible,
+				"single-state",
+				property.SingleState,
+				property.Property,
+			)
+			recordInexpressible(
+				inexpressible,
+				"single-step",
+				property.SingleStep,
+				property.Property,
+			)
+			if !property.Engine.Refuted {
+				if property.SingleState.Refuted || property.SingleStep.Refuted {
+					unmatched++
+				}
+				continue
+			}
+			engineRefutations++
+			weakest[property.Weakest]++
+			if byClass[property.Class] == nil {
+				byClass[property.Class] = map[string]int{}
+			}
+			byClass[property.Class][property.Weakest]++
+		}
+	}
+
+	fmt.Fprintf(
+		out,
+		"\ntraces replayed: %d (rejected: %d)\n",
+		len(reports),
+		rejected,
+	)
+	fmt.Fprintf(
+		out,
+		"validity: %d of %d reproduced the recorded verdicts exactly\n",
+		len(reports)-invalid,
+		len(reports),
+	)
+	fmt.Fprintf(out, "traces where crash-only fired: %d\n", crashed)
+	fmt.Fprintf(out, "engine refutations: %d\n", engineRefutations)
+	if engineRefutations > 0 {
+		fmt.Fprintf(out, "weakest refuting oracle: %s\n", counts(weakest))
+		for _, class := range sortedKeys(byClass) {
+			fmt.Fprintf(out, "  %s: %s\n", class, counts(byClass[class]))
+		}
+		fmt.Fprintf(out, "temporal-only fraction: %.3f\n",
+			float64(weakest["temporal-only"])/float64(engineRefutations))
+	}
+	fmt.Fprintf(
+		out,
+		"properties a reduced oracle cannot express: %s\n",
+		counts(distinct(inexpressible)),
+	)
+	fmt.Fprintf(
+		out,
+		"reduced-oracle refutations the engine did not make: %d\n",
+		unmatched,
+	)
+}
+
+// recordInexpressible counts a property once per oracle however many traces it
+// appears on, because whether an oracle can state a property is a fact about
+// the property and not about the run.
+func recordInexpressible(
+	seen map[string]map[string]bool,
+	oracle string,
+	finding refutation,
+	property string,
+) {
+	if !finding.CannotExpress {
+		return
+	}
+	if seen[oracle] == nil {
+		seen[oracle] = map[string]bool{}
+	}
+	seen[oracle][property] = true
+}
+
+func distinct(seen map[string]map[string]bool) map[string]int {
+	sizes := map[string]int{}
+	for oracle, properties := range seen {
+		sizes[oracle] = len(properties)
+	}
+	return sizes
+}
+
+func counts(values map[string]int) string {
+	parts := make([]string, 0, len(values))
+	for _, key := range sortedKeys(values) {
+		parts = append(parts, fmt.Sprintf("%s=%d", key, values[key]))
+	}
+	return strings.Join(parts, " ")
+}
+
+func sortedKeys[V any](values map[string]V) []string {
+	keys := make([]string, 0, len(values))
+	for key := range values {
+		keys = append(keys, key)
+	}
+	sort.Strings(keys)
+	return keys
+}
diff --git a/cmd/internal-tools/oracle-reduction/oracle.go b/cmd/internal-tools/oracle-reduction/oracle.go
new file mode 100644
index 0000000..e9e53f6
--- /dev/null
+++ b/cmd/internal-tools/oracle-reduction/oracle.go
@@ -0,0 +1,549 @@
+package main
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"sort"
+
+	"github.com/priyanshujain/sanderling/internal/ltl"
+	"github.com/priyanshujain/sanderling/internal/trace"
+	"github.com/priyanshujain/sanderling/internal/verifier"
+)
+
+// foregroundLossReason is the skip reason internal/runner records when the app
+// was no longer the foreground process at action time. It is the only
+// foreground signal a stored trace carries, and it is what the crash-only
+// oracle reads for "the application is no longer the foreground process".
+const foregroundLossReason = "app_left_foreground"
+
+// refutation is one oracle's finding for one property on one trace. A reduced
+// oracle has three of them and not two: CannotExpress says its rewrite no
+// longer states the property, which is neither a refutation nor a clean bill.
+type refutation struct {
+	Refuted       bool `json:"refuted"`
+	CannotExpress bool `json:"cannot_express,omitempty"`
+	// Step is the observation whose evaluation produced the violation;
+	// OriginStep is the observation whose obligation failed, which for a
+	// deferred one is earlier.
+	Step       int    `json:"step,omitempty"`
+	OriginStep int    `json:"origin_step,omitempty"`
+	Reason     string `json:"reason,omitempty"`
+	IsError    bool   `json:"is_error,omitempty"`
+}
+
+type propertyReport struct {
+	Property    string     `json:"property"`
+	Class       string     `json:"class"`
+	TopLevel    string     `json:"top_level"`
+	Engine      refutation `json:"engine"`
+	SingleState refutation `json:"single_state"`
+	SingleStep  refutation `json:"single_step"`
+	// SingleStepTruncatesWindow marks a property whose window the triple had to
+	// shorten to two observations, so what its single-step column refutes is a
+	// stronger property than the one the author wrote.
+	SingleStepTruncatesWindow bool `json:"single_step_truncates_window,omitempty"`
+	// Weakest names the weakest oracle that refutes this property on this
+	// trace, and is empty unless the engine refuted it. "temporal-only" means
+	// no reduced oracle did.
+	Weakest string `json:"weakest_refuting_oracle,omitempty"`
+}
+
+type crashReport struct {
+	Fired               bool  `json:"fired"`
+	FirstStep           int   `json:"first_step,omitempty"`
+	ExceptionSteps      []int `json:"exception_steps,omitempty"`
+	ForegroundLossSteps []int `json:"foreground_loss_steps,omitempty"`
+	ErrorLogSteps       []int `json:"error_log_steps,omitempty"`
+}
+
+// mismatch is one disagreement between the offline engine and the verdicts the
+// run recorded. Any of these blocks the trace's result.
+type mismatch struct {
+	Property string `json:"property"`
+	Field    string `json:"field"`
+	Online   string `json:"online"`
+	Offline  string `json:"offline"`
+}
+
+type runReport struct {
+	Run       string `json:"run"`
+	Seed      int64  `json:"seed"`
+	Platform  string `json:"platform"`
+	Arm       string `json:"arm,omitempty"`
+	Generator string `json:"generator,omitempty"`
+	// ReplayMode says where the extractor values under replay came from:
+	// "page-extractor-values" reuses what the page computed in V8 and the run
+	// evaluated against, "host-extractors" re-runs the spec's getters over the
+	// stored hierarchy.
+	ReplayMode    string     `json:"replay_mode"`
+	StepsObserved int        `json:"steps_observed"`
+	StepsSkipped  int        `json:"steps_skipped"`
+	Valid         bool       `json:"valid"`
+	Mismatches    []mismatch `json:"mismatches,omitempty"`
+	// ResidualMismatches counts every step whose replayed residual formula
+	// differed from the recorded one, of which Mismatches carries the first
+	// few. The recorded residual is the engine's whole pending state, so
+	// agreement on it is a stronger claim than agreement on the verdicts.
+	ResidualMismatches int              `json:"residual_mismatches"`
+	CrashOnly          crashReport      `json:"crash_only"`
+	Properties         []propertyReport `json:"properties"`
+	ActionsUnbounded   int              `json:"actions_without_resolved_bounds"`
+	ScrollActions      int              `json:"scroll_last_actions"`
+}
+
+// witnessRecord is one violation as either side reports it: the step it was
+// recorded at, plus the witness the engine attached.
+type witnessRecord struct {
+	RecordedStep int
+	OriginStep   int
+	DetectedStep int
+	Reason       string
+	IsError      bool
+}
+
+// replay re-evaluates one run offline under all four oracles. The engine's
+// offline verdicts are compared against the ones the run recorded, and the
+// report is marked invalid on any disagreement: a mismatch is a bug here or a
+// gap in the trace, not a finding.
+func replay(
+	run loadedRun,
+	bundleJavaScript string,
+	hostExtractors bool,
+) (runReport, error) {
+	engine, err := verifier.New(
+		verifier.WithSeed(uint64(run.Meta.Seed)),
+		verifier.WithPlatform(run.Meta.Platform),
+		verifier.WithAppPackage(run.Meta.BundleID),
+	)
+	if err != nil {
+		return runReport{}, fmt.Errorf("verifier: %w", err)
+	}
+	if err := engine.Load(bundleJavaScript); err != nil {
+		return runReport{}, fmt.Errorf("load spec: %w", err)
+	}
+	formulas, err := engine.PropertyFormulas()
+	if err != nil {
+		return runReport{}, err
+	}
+
+	singleState := map[string]*ltl.Evaluator{}
+	singleStep := map[string]*ltl.Evaluator{}
+	for name, formula := range formulas {
+		singleState[name] = ltl.NewEvaluator(singleStateFormula(formula))
+		singleStep[name] = ltl.NewEvaluator(singleStepFormula(formula))
+	}
+
+	usePageValues := run.Meta.Platform == "web" && !hostExtractors
+	report := runReport{
+		Run:        run.Directory,
+		Seed:       run.Meta.Seed,
+		Platform:   run.Meta.Platform,
+		Arm:        run.Meta.Arm,
+		Generator:  run.Meta.Generator,
+		ReplayMode: "host-extractors",
+	}
+	var folded []map[int]json.RawMessage
+	if usePageValues {
+		report.ReplayMode = "page-extractor-values"
+		folded, err = extractorFold(run.Steps, engine.ExtractorNames())
+		if err != nil {
+			return runReport{}, err
+		}
+	}
+
+	offline := map[string]witnessRecord{}
+	stateFired := map[string]int{}
+	stepFired := map[string]int{}
+	var residualMismatches []mismatch
+	var lastAction *verifier.Action
+	for position, step := range run.Steps {
+		if step.NextAction != nil && step.NextAction.Selector != "" &&
+			step.NextAction.ResolvedBounds == nil {
+			report.ActionsUnbounded++
+		}
+		if step.SkippedVerification {
+			report.StepsSkipped++
+			residualMismatches = append(
+				residualMismatches,
+				compareResiduals(step, engine.Residuals())...)
+			lastAction = lastActionFor(step)
+			continue
+		}
+		if lastAction != nil && lastAction.Kind == verifier.ActionKindScroll {
+			report.ScrollActions++
+		}
+		if err := engine.PushSnapshot(verifier.SnapshotInput{
+			Tree:       step.Hierarchy,
+			LastAction: lastAction,
+			StepTime:   step.Timestamp,
+			StepIndex:  step.Index,
+			RunStart:   run.Meta.StartedAt,
+			Logs:       traceLogs(step.Logs),
+			Exceptions: traceExceptions(step.Exceptions),
+		}); err != nil {
+			return runReport{}, fmt.Errorf("step %d push: %w", step.Index, err)
+		}
+		if usePageValues {
+			skipped, overrideErr := engine.OverrideExtractorValues(
+				folded[position],
+			)
+			if overrideErr != nil {
+				return runReport{}, fmt.Errorf(
+					"step %d override: %w",
+					step.Index,
+					overrideErr,
+				)
+			}
+			if skipped > 0 {
+				return runReport{}, fmt.Errorf(
+					"step %d: %d recorded extractor values fell outside the spec's extractor list",
+					step.Index,
+					skipped,
+				)
+			}
+		}
+		engine.EvaluateProperties()
+		for _, name := range engine.NewlyViolatedProperties() {
+			offline[name] = witnessFrom(engine.Witness(name), step.Index)
+		}
+		residualMismatches = append(
+			residualMismatches,
+			compareResiduals(step, engine.Residuals())...)
+		for name := range formulas {
+			recordFiring(
+				stateFired,
+				name,
+				singleState[name].ObserveAtStep(step.Timestamp, step.Index),
+				step.Index,
+			)
+			recordFiring(
+				stepFired,
+				name,
+				singleStep[name].ObserveAtStep(step.Timestamp, step.Index),
+				step.Index,
+			)
+		}
+		report.StepsObserved++
+		lastAction = lastActionFor(step)
+	}
+
+	finalizeIndex := run.finalizeIndex()
+	for _, name := range engine.Finalize() {
+		offline[name] = witnessFrom(engine.Witness(name), finalizeIndex)
+	}
+	for name := range formulas {
+		recordFiring(
+			stateFired,
+			name,
+			singleState[name].Finalize(),
+			finalizeIndex,
+		)
+		recordFiring(
+			stepFired,
+			name,
+			singleStep[name].Finalize(),
+			finalizeIndex,
+		)
+	}
+
+	report.CrashOnly = crashOnly(run)
+	report.Mismatches = compareVerdicts(onlineVerdicts(run), offline)
+	report.ResidualMismatches = len(residualMismatches)
+	if len(residualMismatches) > reportedResiduals {
+		residualMismatches = residualMismatches[:reportedResiduals]
+	}
+	report.Mismatches = append(report.Mismatches, residualMismatches...)
+	report.Valid = len(report.Mismatches) == 0
+
+	names := make([]string, 0, len(formulas))
+	for name := range formulas {
+		names = append(names, name)
+	}
+	sort.Strings(names)
+	for _, name := range names {
+		property := propertyReport{
+			Property: name,
+			Class:    propertyClass(formulas[name]),
+			TopLevel: topLevelForm(formulas[name]),
+			Engine:   engineRefutation(offline, name),
+			SingleState: reducedRefutation(
+				singleStateExpresses(formulas[name]),
+				singleState[name],
+				stateFired[name],
+			),
+			SingleStep: reducedRefutation(
+				singleStepExpresses(formulas[name]),
+				singleStep[name],
+				stepFired[name],
+			),
+
+			SingleStepTruncatesWindow: truncatesWindow(formulas[name]),
+		}
+		property.Weakest = weakestOracle(property, report.CrashOnly)
+		report.Properties = append(report.Properties, property)
+	}
+	return report, nil
+}
+
+// reportedResiduals caps how many residual differences one trace lists. The
+// count of all of them is reported alongside; the first few are what says
+// where the two engines parted.
+const reportedResiduals = 5
+
+// compareResiduals checks the replayed pending state against the one the step
+// recorded. A step the verifier skipped still recorded the residual it was
+// holding, so those steps assert that a skipped step advanced nothing.
+func compareResiduals(
+	step trace.Step,
+	replayed map[string]ltl.Formula,
+) []mismatch {
+	if len(step.Residuals) == 0 {
+		return nil
+	}
+	var mismatches []mismatch
+	names := make([]string, 0, len(step.Residuals))
+	for name := range step.Residuals {
+		names = append(names, name)
+	}
+	sort.Strings(names)
+	for _, name := range names {
+		formula, ok := replayed[name]
+		if !ok {
+			mismatches = append(mismatches, mismatch{
+				Property: name,
+				Field:    fmt.Sprintf("residual at step %d", step.Index),
+				Online:   string(step.Residuals[name]),
+				Offline:  "property not registered",
+			})
+			continue
+		}
+		encoded, err := json.Marshal(formula)
+		if err != nil {
+			encoded = []byte(fmt.Sprintf("%q", err.Error()))
+		}
+		if !bytes.Equal(encoded, step.Residuals[name]) {
+			mismatches = append(mismatches, mismatch{
+				Property: name,
+				Field:    fmt.Sprintf("residual at step %d", step.Index),
+				Online:   string(step.Residuals[name]),
+				Offline:  string(encoded),
+			})
+		}
+	}
+	return mismatches
+}
+
+func witnessFrom(witness *verifier.Witness, recordedStep int) witnessRecord {
+	record := witnessRecord{RecordedStep: recordedStep}
+	if witness == nil {
+		return record
+	}
+	record.OriginStep = witness.Step
+	record.DetectedStep = witness.DetectedStep
+	record.Reason = witness.Reason
+	record.IsError = witness.IsError
+	return record
+}
+
+// onlineVerdicts reads the violations the run recorded, including the
+// end-of-run record a finalized liveness obligation is written to.
+func onlineVerdicts(run loadedRun) map[string]witnessRecord {
+	recorded := map[string]witnessRecord{}
+	steps := run.Steps
+	if run.Finalize != nil {
+		steps = append(append([]trace.Step(nil), steps...), *run.Finalize)
+	}
+	for _, step := range steps {
+		for _, name := range step.Violations {
+			record := witnessRecord{RecordedStep: step.Index}
+			if witness, ok := step.Witnesses[name]; ok {
+				record.OriginStep = witness.Step
+				record.DetectedStep = witness.DetectedStep
+				record.Reason = witness.Reason
+				record.IsError = witness.IsError
+			}
+			recorded[name] = record
+		}
+	}
+	return recorded
+}
+
+func compareVerdicts(online, offline map[string]witnessRecord) []mismatch {
+	var mismatches []mismatch
+	names := map[string]bool{}
+	for name := range online {
+		names[name] = true
+	}
+	for name := range offline {
+		names[name] = true
+	}
+	ordered := make([]string, 0, len(names))
+	for name := range names {
+		ordered = append(ordered, name)
+	}
+	sort.Strings(ordered)
+
+	for _, name := range ordered {
+		recorded, wasRecorded := online[name]
+		replayed, wasReplayed := offline[name]
+		switch {
+		case wasRecorded && !wasReplayed:
+			mismatches = append(mismatches, mismatch{
+				Property: name, Field: "violated",
+				Online: fmt.Sprintf(
+					"violated at step %d",
+					recorded.RecordedStep,
+				),
+				Offline: "not violated",
+			})
+			continue
+		case !wasRecorded && wasReplayed:
+			mismatches = append(mismatches, mismatch{
+				Property: name, Field: "violated",
+				Online: "not violated",
+				Offline: fmt.Sprintf(
+					"violated at step %d",
+					replayed.RecordedStep,
+				),
+			})
+			continue
+		case !wasRecorded:
+			continue
+		}
+		for _, field := range []struct {
+			name    string
+			online  string
+			offline string
+		}{
+			{"step", fmt.Sprint(recorded.RecordedStep), fmt.Sprint(replayed.RecordedStep)},
+			{"origin_step", fmt.Sprint(recorded.OriginStep), fmt.Sprint(replayed.OriginStep)},
+			{"detected_step", fmt.Sprint(recorded.DetectedStep), fmt.Sprint(replayed.DetectedStep)},
+			{"reason", recorded.Reason, replayed.Reason},
+		} {
+			if field.online != field.offline {
+				mismatches = append(mismatches, mismatch{
+					Property: name, Field: field.name,
+					Online: field.online, Offline: field.offline,
+				})
+			}
+		}
+	}
+	return mismatches
+}
+
+// crashOnly fires where the application left the foreground or an error
+// surface was recorded. Error-level log lines are reported alongside rather
+// than folded in: a console error is not a crash, and an analysis that wants
+// the looser detector can read the steps from here.
+func crashOnly(run loadedRun) crashReport {
+	report := crashReport{}
+	for _, step := range run.Steps {
+		if len(step.Exceptions) > 0 {
+			report.ExceptionSteps = append(report.ExceptionSteps, step.Index)
+		}
+		if step.ActionSkipped == foregroundLossReason {
+			report.ForegroundLossSteps = append(
+				report.ForegroundLossSteps,
+				step.Index,
+			)
+		}
+		for _, entry := range step.Logs {
+			if entry.Level == "E" || entry.Level == "F" {
+				report.ErrorLogSteps = append(report.ErrorLogSteps, step.Index)
+				break
+			}
+		}
+	}
+	first := 0
+	for _, step := range append(append([]int(nil), report.ExceptionSteps...), report.ForegroundLossSteps...) {
+		if first == 0 || step < first {
+			first = step
+		}
+	}
+	report.Fired = first != 0
+	report.FirstStep = first
+	return report
+}
+
+func engineRefutation(
+	offline map[string]witnessRecord,
+	name string,
+) refutation {
+	record, ok := offline[name]
+	if !ok {
+		return refutation{}
+	}
+	return refutation{
+		Refuted:    true,
+		Step:       record.RecordedStep,
+		OriginStep: record.OriginStep,
+		Reason:     record.Reason,
+		IsError:    record.IsError,
+	}
+}
+
+// reducedRefutation reports a reduced oracle's finding, or its inability to
+// state the property at all. The evaluator is driven either way so that the
+// verdict a reduction would have reached is never what decides whether it was
+// entitled to reach one.
+func reducedRefutation(
+	expresses bool,
+	evaluator *ltl.Evaluator,
+	firedAt int,
+) refutation {
+	if !expresses {
+		return refutation{CannotExpress: true}
+	}
+	return evaluatorRefutation(evaluator, firedAt)
+}
+
+func evaluatorRefutation(evaluator *ltl.Evaluator, firedAt int) refutation {
+	violation := evaluator.Violation()
+	if violation == nil {
+		return refutation{}
+	}
+	return refutation{
+		Refuted:    true,
+		Step:       firedAt,
+		OriginStep: violation.Step,
+		Reason:     violation.Reason,
+		IsError:    violation.IsError,
+	}
+}
+
+// recordFiring keeps the first observation at which a reduced oracle latched,
+// which the evaluator itself does not carry.
+func recordFiring(
+	fired map[string]int,
+	name string,
+	verdict ltl.Verdict,
+	step int,
+) {
+	if verdict != ltl.VerdictViolated {
+		return
+	}
+	if _, ok := fired[name]; ok {
+		return
+	}
+	fired[name] = step
+}
+
+// weakestOracle names the weakest oracle that refutes a defect the engine
+// refuted, in the order a crash detector, a single-state check and a
+// single-step triple are weaker than the engine.
+func weakestOracle(property propertyReport, crash crashReport) string {
+	if !property.Engine.Refuted {
+		return ""
+	}
+	switch {
+	case crash.Fired:
+		return "crash-only"
+	case property.SingleState.Refuted:
+		return "single-state"
+	case property.SingleStep.Refuted:
+		return "single-step"
+	default:
+		return "temporal-only"
+	}
+}
diff --git a/cmd/internal-tools/oracle-reduction/oracle_test.go b/cmd/internal-tools/oracle-reduction/oracle_test.go
new file mode 100644
index 0000000..1af0b92
--- /dev/null
+++ b/cmd/internal-tools/oracle-reduction/oracle_test.go
@@ -0,0 +1,433 @@
+package main
+
+import (
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/priyanshujain/sanderling/internal/hierarchy"
+	"github.com/priyanshujain/sanderling/internal/trace"
+	"github.com/priyanshujain/sanderling/internal/verifier"
+)
+
+// fixtureSpec counts rows on screen and asserts three things about them: a
+// bound a single observation can check, a reachability goal, and a growth step
+// that only two consecutive observations can see.
+const fixtureSpec = `
+const rows = __sanderling__.extract((s) => s.ax.findAll({ text: "row" }).length).named("rows");
+globalThis.properties = {
+  fewRows: __sanderling__.always(() => rows.current < 3),
+  rowsAppear: __sanderling__.eventually(() => rows.current > 0).within(500, "steps"),
+  rowsKeepGrowing: __sanderling__.always(
+    __sanderling__.now(() => rows.current === 1).implies(
+      __sanderling__.next(() => rows.current === 2))),
+};
+`
+
+func rowTree(t *testing.T, rows int) *hierarchy.Tree {
+	t.Helper()
+	children := make([]string, 0, rows)
+	for range rows {
+		children = append(
+			children,
+			`{"attributes": {"text": "row", "bounds": "[0,0,10,10]"}, "children": []}`,
+		)
+	}
+	source := fmt.Sprintf(
+		`{"attributes": {"resource-id": "root", "bounds": "[0,0,100,100]"}, "children": [%s]}`,
+		strings.Join(children, ","),
+	)
+	tree, err := hierarchy.Parse(source)
+	if err != nil {
+		t.Fatalf("parse tree: %v", err)
+	}
+	return tree
+}
+
+// writeFixtureRun records a run the way internal/runner does: every step's
+// hierarchy, the violations that fired at it, their witnesses, and the residual
+// each property was left holding.
+func writeFixtureRun(t *testing.T, directory string, rowCounts []int) {
+	t.Helper()
+	engine, err := verifier.New(verifier.WithPlatform("android"))
+	if err != nil {
+		t.Fatalf("verifier: %v", err)
+	}
+	if err := engine.Load(fixtureSpec); err != nil {
+		t.Fatalf("load spec: %v", err)
+	}
+	writer, err := trace.NewWriter(directory)
+	if err != nil {
+		t.Fatalf("trace writer: %v", err)
+	}
+	defer writer.Close()
+
+	runStart := time.Date(2026, 8, 15, 12, 0, 0, 0, time.UTC)
+	if err := writer.WriteMeta(trace.Meta{
+		Seed:      11,
+		SpecPath:  "fixture.ts",
+		Platform:  "android",
+		BundleID:  "com.example.fixture",
+		StartedAt: runStart,
+	}); err != nil {
+		t.Fatalf("meta: %v", err)
+	}
+
+	lastIndex := 0
+	for position, rows := range rowCounts {
+		index := position + 1
+		lastIndex = index
+		stepTime := runStart.Add(time.Duration(index) * time.Second)
+		tree := rowTree(t, rows)
+		if err := engine.PushSnapshot(verifier.SnapshotInput{
+			Tree:      tree,
+			StepTime:  stepTime,
+			StepIndex: index,
+			RunStart:  runStart,
+		}); err != nil {
+			t.Fatalf("push step %d: %v", index, err)
+		}
+		engine.EvaluateProperties()
+		violations := engine.NewlyViolatedProperties()
+		step := trace.Step{
+			Index:      index,
+			Timestamp:  stepTime,
+			Hierarchy:  tree,
+			Violations: violations,
+			Witnesses:  fixtureWitnesses(engine, violations, index),
+			Residuals:  fixtureResiduals(t, engine),
+		}
+		if err := writer.WriteStep(step); err != nil {
+			t.Fatalf("write step %d: %v", index, err)
+		}
+	}
+	if ended := engine.Finalize(); len(ended) > 0 {
+		if err := writer.WriteStep(trace.Step{
+			Index:      lastIndex + 1,
+			Timestamp:  runStart.Add(time.Duration(lastIndex+1) * time.Second),
+			Violations: ended,
+			Witnesses:  fixtureWitnesses(engine, ended, lastIndex+1),
+		}); err != nil {
+			t.Fatalf("write finalize step: %v", err)
+		}
+	}
+}
+
+func fixtureWitnesses(
+	engine *verifier.Verifier,
+	properties []string,
+	index int,
+) map[string]trace.Witness {
+	if len(properties) == 0 {
+		return nil
+	}
+	witnesses := map[string]trace.Witness{}
+	for _, name := range properties {
+		witness := engine.Witness(name)
+		if witness == nil {
+			continue
+		}
+		detected := witness.DetectedStep
+		if detected == 0 {
+			detected = index
+		}
+		witnesses[name] = trace.Witness{
+			Reason:       witness.Reason,
+			IsError:      witness.IsError,
+			Step:         witness.Step,
+			DetectedStep: detected,
+			Extractors:   witness.Extractors,
+		}
+	}
+	return witnesses
+}
+
+func fixtureResiduals(
+	t *testing.T,
+	engine *verifier.Verifier,
+) map[string]json.RawMessage {
+	t.Helper()
+	residuals := map[string]json.RawMessage{}
+	for name, formula := range engine.Residuals() {
+		body, err := json.Marshal(formula)
+		if err != nil {
+			t.Fatalf("marshal residual %q: %v", name, err)
+		}
+		residuals[name] = body
+	}
+	return residuals
+}
+
+func replayFixture(t *testing.T, directory string) runReport {
+	t.Helper()
+	loaded, err := loadRun(directory)
+	if err != nil {
+		t.Fatalf("load run: %v", err)
+	}
+	report, err := replay(loaded, fixtureSpec, false)
+	if err != nil {
+		t.Fatalf("replay: %v", err)
+	}
+	return report
+}
+
+func propertyByName(
+	t *testing.T,
+	report runReport,
+	name string,
+) propertyReport {
+	t.Helper()
+	for _, property := range report.Properties {
+		if property.Property == name {
+			return property
+		}
+	}
+	t.Fatalf("property %q missing from the report", name)
+	return propertyReport{}
+}
+
+func TestReplayReproducesTheRecordedVerdicts(t *testing.T) {
+	directory := t.TempDir()
+	writeFixtureRun(t, directory, []int{0, 1, 1, 3})
+
+	report := replayFixture(t, directory)
+
+	if !report.Valid {
+		t.Fatalf("replay disagreed with the run: %+v", report.Mismatches)
+	}
+	if report.ResidualMismatches != 0 {
+		t.Errorf("residual mismatches: %d", report.ResidualMismatches)
+	}
+	if report.StepsObserved != 4 {
+		t.Errorf("steps observed: got %d, want 4", report.StepsObserved)
+	}
+
+	growth := propertyByName(t, report, "rowsKeepGrowing")
+	if !growth.Engine.Refuted || growth.Engine.Step != 3 {
+		t.Errorf("engine on rowsKeepGrowing: %+v", growth.Engine)
+	}
+	if growth.SingleState.Refuted {
+		t.Errorf(
+			"single-state refuted a growth step it cannot see: %+v",
+			growth.SingleState,
+		)
+	}
+	if !growth.SingleStep.Refuted {
+		t.Error("single-step did not refute a one-step obligation")
+	}
+	if growth.Weakest != "single-step" {
+		t.Errorf("weakest oracle for rowsKeepGrowing: got %q", growth.Weakest)
+	}
+
+	bound := propertyByName(t, report, "fewRows")
+	if !bound.SingleState.Refuted || bound.Weakest != "single-state" {
+		t.Errorf(
+			"fewRows: single-state %+v weakest %q",
+			bound.SingleState,
+			bound.Weakest,
+		)
+	}
+}
+
+func TestReplayReportsAViolationTheRunNeverRecorded(t *testing.T) {
+	directory := t.TempDir()
+	writeFixtureRun(t, directory, []int{0, 1, 1, 3})
+	dropRecordedViolation(t, directory, "rowsKeepGrowing")
+
+	report := replayFixture(t, directory)
+
+	if report.Valid {
+		t.Fatal("a trace missing a recorded violation must not replay as valid")
+	}
+	var found bool
+	for _, entry := range report.Mismatches {
+		if entry.Property == "rowsKeepGrowing" && entry.Field == "violated" {
+			found = true
+		}
+	}
+	if !found {
+		t.Errorf(
+			"no mismatch named the dropped violation: %+v",
+			report.Mismatches,
+		)
+	}
+}
+
+func TestReplayReportsAResidualTheRunNeverHeld(t *testing.T) {
+	directory := t.TempDir()
+	writeFixtureRun(t, directory, []int{0, 1, 1, 3})
+	rewriteResidual(t, directory, 1, "fewRows", `{"op":"false"}`)
+
+	report := replayFixture(t, directory)
+
+	if report.Valid {
+		t.Fatal(
+			"a trace whose recorded residual differs must not replay as valid",
+		)
+	}
+	if report.ResidualMismatches != 1 {
+		t.Errorf(
+			"residual mismatches: got %d, want 1",
+			report.ResidualMismatches,
+		)
+	}
+}
+
+// dropRecordedViolation removes one property from every step's violations,
+// which is what a trace that failed to record a verdict looks like.
+func dropRecordedViolation(t *testing.T, directory, property string) {
+	t.Helper()
+	rewriteSteps(t, directory, func(step *trace.Step) {
+		kept := step.Violations[:0]
+		for _, name := range step.Violations {
+			if name != property {
+				kept = append(kept, name)
+			}
+		}
+		step.Violations = kept
+		delete(step.Witnesses, property)
+	})
+}
+
+func rewriteResidual(
+	t *testing.T,
+	directory string,
+	index int,
+	property, residual string,
+) {
+	t.Helper()
+	rewriteSteps(t, directory, func(step *trace.Step) {
+		if step.Index == index {
+			step.Residuals[property] = json.RawMessage(residual)
+		}
+	})
+}
+
+func rewriteSteps(t *testing.T, directory string, edit func(*trace.Step)) {
+	t.Helper()
+	path := filepath.Join(directory, "trace.jsonl")
+	body, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatalf("read trace: %v", err)
+	}
+	var rewritten strings.Builder
+	for _, line := range strings.Split(strings.TrimSpace(string(body)), "\n") {
+		var step trace.Step
+		if err := json.Unmarshal([]byte(line), &step); err != nil {
+			t.Fatalf("decode step: %v", err)
+		}
+		edit(&step)
+		encoded, err := json.Marshal(step)
+		if err != nil {
+			t.Fatalf("encode step: %v", err)
+		}
+		rewritten.Write(encoded)
+		rewritten.WriteByte('\n')
+	}
+	if err := os.WriteFile(path, []byte(rewritten.String()), 0o644); err != nil {
+		t.Fatalf("write trace: %v", err)
+	}
+}
+
+func TestCrashOnlyFiresOnAnErrorSurfaceAndOnLeavingTheForeground(t *testing.T) {
+	run := loadedRun{Steps: []trace.Step{
+		{Index: 1, Logs: []trace.LogEntry{{Level: "E", Message: "noisy"}}},
+		{Index: 2, ActionSkipped: foregroundLossReason},
+		{Index: 3, Exceptions: []trace.Exception{{Class: "TypeError"}}},
+	}}
+
+	report := crashOnly(run)
+
+	if !report.Fired || report.FirstStep != 2 {
+		t.Errorf("crash-only: %+v", report)
+	}
+	if len(report.ExceptionSteps) != 1 || report.ExceptionSteps[0] != 3 {
+		t.Errorf("exception steps: %v", report.ExceptionSteps)
+	}
+	if len(report.ErrorLogSteps) != 1 || report.ErrorLogSteps[0] != 1 {
+		t.Errorf("error log steps: %v", report.ErrorLogSteps)
+	}
+}
+
+func TestCrashOnlyIgnoresAnErrorLogOnItsOwn(t *testing.T) {
+	run := loadedRun{
+		Steps: []trace.Step{{Index: 1, Logs: []trace.LogEntry{{Level: "E"}}}},
+	}
+
+	if crashOnly(run).Fired {
+		t.Error("an error-level log line is not a crash")
+	}
+}
+
+// A reachability goal the run reaches at the fourth observation is clean under
+// the window its author wrote and refuted by the same window shortened to two
+// observations. Reporting that shortened refutation would convict every clean
+// trace, so the single-step column has to admit it cannot state the property.
+func TestSingleStepDoesNotConvictACleanTraceOfATruncatedWindow(t *testing.T) {
+	directory := t.TempDir()
+	writeFixtureRun(t, directory, []int{0, 0, 0, 1})
+
+	report := replayFixture(t, directory)
+
+	if !report.Valid {
+		t.Fatalf("replay disagreed with the run: %+v", report.Mismatches)
+	}
+	appear := propertyByName(t, report, "rowsAppear")
+	if appear.Engine.Refuted {
+		t.Fatalf(
+			"the trace reaches the goal inside the 500-step window: %+v",
+			appear.Engine,
+		)
+	}
+	if appear.SingleStep.Refuted {
+		t.Errorf(
+			"single-step refuted a clean trace by shortening the window: %+v",
+			appear.SingleStep,
+		)
+	}
+	if !appear.SingleStep.CannotExpress {
+		t.Error(
+			"single-step must record a window it cannot state as inexpressible",
+		)
+	}
+	if !appear.SingleStepTruncatesWindow {
+		t.Error("the truncated-window marker must stay visible on the property")
+	}
+}
+
+func TestSingleStateAdmitsWhatOneObservationCannotState(t *testing.T) {
+	directory := t.TempDir()
+	writeFixtureRun(t, directory, []int{0, 1, 1, 3})
+
+	report := replayFixture(t, directory)
+
+	growth := propertyByName(t, report, "rowsKeepGrowing")
+	if !growth.SingleState.CannotExpress {
+		t.Errorf(
+			"single-state kept a next obligation it erases to nothing: %+v",
+			growth.SingleState,
+		)
+	}
+	appear := propertyByName(t, report, "rowsAppear")
+	if !appear.SingleState.CannotExpress {
+		t.Errorf(
+			"single-state kept a reachability goal it erases to nothing: %+v",
+			appear.SingleState,
+		)
+	}
+	bound := propertyByName(t, report, "fewRows")
+	if bound.SingleState.CannotExpress {
+		t.Error("single-state states a bound read at one observation")
+	}
+	if !bound.SingleState.Refuted || bound.Weakest != "single-state" {
+		t.Errorf(
+			"fewRows: single-state %+v weakest %q",
+			bound.SingleState,
+			bound.Weakest,
+		)
+	}
+}
diff --git a/cmd/internal-tools/oracle-reduction/reduce.go b/cmd/internal-tools/oracle-reduction/reduce.go
new file mode 100644
index 0000000..6c61b4b
--- /dev/null
+++ b/cmd/internal-tools/oracle-reduction/reduce.go
@@ -0,0 +1,281 @@
+package main
+
+import (
+	"fmt"
+	"time"
+
+	"github.com/priyanshujain/sanderling/internal/ltl"
+)
+
+// tripleWindow is the horizon a property triple has: an obligation armed at one
+// observation must discharge at the next, and nothing outlives that.
+const tripleWindow = 2
+
+// singleStateFormula is what a checker holding one observation can refute.
+// Every operator that defers or repeats an obligation is replaced by the
+// constant that makes the formula around it trivially satisfied, so the only
+// refutation left is a predicate read at a single observation. The outermost
+// always survives because it is what says "check this at each observation",
+// which costs no history.
+func singleStateFormula(formula ltl.Formula) ltl.Formula {
+	if always, ok := formula.(ltl.AlwaysFormula); ok {
+		always.Inner = stateless(always.Inner, true)
+		return always
+	}
+	return stateless(formula, true)
+}
+
+// singleStepFormula is the property triple: one step of history, and no
+// obligation surviving past the next observation. A next keeps its one-step
+// deferral, and an eventually of any window shrinks to the two observations a
+// triple spans.
+func singleStepFormula(formula ltl.Formula) ltl.Formula {
+	if always, ok := formula.(ltl.AlwaysFormula); ok {
+		always.Inner = oneStep(always.Inner, true)
+		return always
+	}
+	return oneStep(formula, true)
+}
+
+// singleStateExpresses and singleStepExpresses decide, from the property's form
+// alone, whether the reduced oracle can still state the property after its
+// rewrite. An oracle that cannot state a property does not get to refute it,
+// and is reported as silent on it rather than as either verdict.
+//
+// Two shapes defeat a reduction. The rewrite shortens an obligation window the
+// oracle's horizon cannot hold, leaving it to refute a property strictly
+// stronger than the one the author wrote. Or the rewrite leaves a formula whose
+// verdict no longer depends on the trace, leaving it to report the same answer
+// everywhere. Either way the column would carry a verdict about a property
+// nobody wrote, which is worth less than an admission that the oracle is out of
+// its depth.
+func singleStateExpresses(formula ltl.Formula) bool {
+	return dependsOnTrace(singleStateFormula(formula))
+}
+
+func singleStepExpresses(formula ltl.Formula) bool {
+	return !truncatesWindow(formula) &&
+		dependsOnTrace(singleStepFormula(formula))
+}
+
+// truncatesWindow reports whether the single-step rewrite had to shorten a
+// window to fit a triple. Where it did, the reduced oracle is checking a
+// stronger property than the author wrote, so a refutation of it is not the
+// same event as a refutation of the property.
+func truncatesWindow(formula ltl.Formula) bool {
+	if always, ok := formula.(ltl.AlwaysFormula); ok {
+		return shortensWindow(always.Inner, true)
+	}
+	return shortensWindow(formula, true)
+}
+
+// shortensWindow walks the same shape oneStep rewrites, and reports only the
+// shortening that strengthens the formula. Polarity is what separates the two:
+// a shorter window under an even number of negations demands the same thing
+// sooner, so a refutation of it need not be a refutation of the property, while
+// under an odd number it asks for less and its refutations stay sound. The
+// operators oneStep erases rather than shortens are erased to the constant its
+// position is satisfied by, which also only asks for less.
+func shortensWindow(formula ltl.Formula, positive bool) bool {
+	switch concrete := formula.(type) {
+	case ltl.EventuallyFormula:
+		return positive && windowOutlastsTriple(concrete)
+	case ltl.NowFormula:
+		return shortensWindow(concrete.Inner, positive)
+	case ltl.NotFormula:
+		return shortensWindow(concrete.Inner, !positive)
+	case ltl.AndFormula:
+		return shortensWindow(concrete.Left, positive) || shortensWindow(concrete.Right, positive)
+	case ltl.OrFormula:
+		return shortensWindow(concrete.Left, positive) || shortensWindow(concrete.Right, positive)
+	case ltl.ImpliesFormula:
+		return shortensWindow(concrete.Antecedent, !positive) || shortensWindow(concrete.Consequent, positive)
+	default:
+		return false
+	}
+}
+
+// windowOutlastsTriple asks whether an obligation can still be open after the
+// two observations a triple spans. A window counted in time is outside what a
+// triple can state whatever its length, because a triple has no clock: it can
+// say "at the next observation" and nothing about when that arrives.
+func windowOutlastsTriple(formula ltl.EventuallyFormula) bool {
+	if formula.HasStepBound {
+		return formula.StepBound > tripleWindow
+	}
+	return true
+}
+
+// dependsOnTrace reports whether a rewritten formula can still read the trace.
+// Constants fold through the connectives, and one that folds away to a constant
+// answers the same on every trace: silence that reads as "did not refute", or a
+// refutation of everything. Neither is a verdict about the run.
+func dependsOnTrace(formula ltl.Formula) bool {
+	_, constant := fold(formula).(ltl.PureFormula)
+	return !constant
+}
+
+// fold propagates the constants the rewrites substituted for erased operators.
+// An obligation whose inner formula folded to false keeps its shape, because a
+// deferred false is a refutation still owed; only the side that can no longer
+// fail folds away.
+func fold(formula ltl.Formula) ltl.Formula {
+	switch concrete := formula.(type) {
+	case ltl.AlwaysFormula:
+		concrete.Inner = fold(concrete.Inner)
+		if pure, ok := concrete.Inner.(ltl.PureFormula); ok {
+			return pure
+		}
+		return concrete
+	case ltl.EventuallyFormula:
+		concrete.Inner = fold(concrete.Inner)
+		if isPure(concrete.Inner, true) {
+			return ltl.Pure(true)
+		}
+		return concrete
+	case ltl.NextFormula:
+		inner := fold(concrete.Inner)
+		if isPure(inner, true) {
+			return ltl.Pure(true)
+		}
+		return ltl.Next(inner)
+	case ltl.NowFormula:
+		inner := fold(concrete.Inner)
+		if _, ok := inner.(ltl.PureFormula); ok {
+			return inner
+		}
+		return ltl.Now(inner)
+	case ltl.NotFormula:
+		inner := fold(concrete.Inner)
+		if pure, ok := inner.(ltl.PureFormula); ok {
+			return ltl.Pure(!pure.Value)
+		}
+		return ltl.Not(inner)
+	case ltl.AndFormula:
+		left, right := fold(concrete.Left), fold(concrete.Right)
+		switch {
+		case isPure(left, false) || isPure(right, false):
+			return ltl.Pure(false)
+		case isPure(left, true):
+			return right
+		case isPure(right, true):
+			return left
+		}
+		return ltl.And(left, right)
+	case ltl.OrFormula:
+		left, right := fold(concrete.Left), fold(concrete.Right)
+		switch {
+		case isPure(left, true) || isPure(right, true):
+			return ltl.Pure(true)
+		case isPure(left, false):
+			return right
+		case isPure(right, false):
+			return left
+		}
+		return ltl.Or(left, right)
+	case ltl.ImpliesFormula:
+		antecedent, consequent := fold(concrete.Antecedent), fold(concrete.Consequent)
+		switch {
+		case isPure(antecedent, false) || isPure(consequent, true):
+			return ltl.Pure(true)
+		case isPure(antecedent, true):
+			return consequent
+		}
+		return ltl.Implies(antecedent, consequent)
+	default:
+		return formula
+	}
+}
+
+func isPure(formula ltl.Formula, value bool) bool {
+	pure, ok := formula.(ltl.PureFormula)
+	return ok && pure.Value == value
+}
+
+// stateless erases every temporal operator. The replacement constant follows
+// the position's polarity: under an even number of negations a temporal
+// sub-formula is dropped as satisfied, and under an odd number as failed, so
+// that in both cases its negation cannot refute anything either.
+func stateless(formula ltl.Formula, positive bool) ltl.Formula {
+	switch concrete := formula.(type) {
+	case ltl.AlwaysFormula, ltl.NextFormula, ltl.EventuallyFormula:
+		return ltl.Pure(positive)
+	case ltl.NowFormula:
+		return ltl.Now(stateless(concrete.Inner, positive))
+	case ltl.NotFormula:
+		return ltl.Not(stateless(concrete.Inner, !positive))
+	case ltl.AndFormula:
+		return ltl.And(stateless(concrete.Left, positive), stateless(concrete.Right, positive))
+	case ltl.OrFormula:
+		return ltl.Or(stateless(concrete.Left, positive), stateless(concrete.Right, positive))
+	case ltl.ImpliesFormula:
+		return ltl.Implies(
+			stateless(concrete.Antecedent, !positive),
+			stateless(concrete.Consequent, positive),
+		)
+	default:
+		return formula
+	}
+}
+
+func oneStep(formula ltl.Formula, positive bool) ltl.Formula {
+	switch concrete := formula.(type) {
+	case ltl.AlwaysFormula:
+		return ltl.Pure(positive)
+	case ltl.NextFormula:
+		return ltl.Next(stateless(concrete.Inner, positive))
+	case ltl.EventuallyFormula:
+		return ltl.EventuallyWithinSteps(stateless(concrete.Inner, positive), tripleWindow)
+	case ltl.NowFormula:
+		return ltl.Now(oneStep(concrete.Inner, positive))
+	case ltl.NotFormula:
+		return ltl.Not(oneStep(concrete.Inner, !positive))
+	case ltl.AndFormula:
+		return ltl.And(oneStep(concrete.Left, positive), oneStep(concrete.Right, positive))
+	case ltl.OrFormula:
+		return ltl.Or(oneStep(concrete.Left, positive), oneStep(concrete.Right, positive))
+	case ltl.ImpliesFormula:
+		return ltl.Implies(
+			oneStep(concrete.Antecedent, !positive),
+			oneStep(concrete.Consequent, positive),
+		)
+	default:
+		return formula
+	}
+}
+
+// propertyClass splits safety from liveness by the property's top-level form:
+// a reachability goal is liveness, everything else is a safety obligation
+// re-asserted at each observation.
+func propertyClass(formula ltl.Formula) string {
+	if _, ok := formula.(ltl.EventuallyFormula); ok {
+		return "liveness"
+	}
+	return "safety"
+}
+
+func topLevelForm(formula ltl.Formula) string {
+	switch concrete := formula.(type) {
+	case ltl.AlwaysFormula:
+		return "always" + boundSuffix(concrete.HasStepBound, concrete.StepBound, concrete.Duration)
+	case ltl.EventuallyFormula:
+		return "eventually" + boundSuffix(concrete.HasStepBound, concrete.StepBound, concrete.Duration)
+	default:
+		return "predicate"
+	}
+}
+
+func boundSuffix(
+	hasStepBound bool,
+	stepBound int,
+	duration time.Duration,
+) string {
+	switch {
+	case hasStepBound:
+		return fmt.Sprintf(" within %d steps", stepBound)
+	case duration > 0:
+		return fmt.Sprintf(" within %s", duration)
+	default:
+		return ""
+	}
+}
diff --git a/cmd/internal-tools/oracle-reduction/reduce_test.go b/cmd/internal-tools/oracle-reduction/reduce_test.go
new file mode 100644
index 0000000..6216e30
--- /dev/null
+++ b/cmd/internal-tools/oracle-reduction/reduce_test.go
@@ -0,0 +1,222 @@
+package main
+
+import (
+	"testing"
+	"time"
+
+	"github.com/priyanshujain/sanderling/internal/ltl"
+)
+
+// observe drives an evaluator over a fixed sequence of predicate readings and
+// returns the observation at which it latched, or 0 if it never did.
+func observe(formula ltl.Formula, steps int) int {
+	evaluator := ltl.NewEvaluator(formula)
+	base := time.Unix(0, 0)
+	for step := 1; step <= steps; step++ {
+		if evaluator.ObserveAtStep(
+			base.Add(time.Duration(step)*time.Second),
+			step,
+		) == ltl.VerdictViolated {
+			return step
+		}
+	}
+	if evaluator.Finalize() == ltl.VerdictViolated {
+		return steps + 1
+	}
+	return 0
+}
+
+func constant(value bool) ltl.Formula {
+	return ltl.ThunkNamed("p", func() (bool, error) { return value, nil })
+}
+
+func TestSingleStateHoldsWhereRefutationNeedsTheNextObservation(t *testing.T) {
+	property := ltl.Always(
+		ltl.Implies(ltl.Now(constant(true)), ltl.Next(constant(false))),
+	)
+
+	if engine := observe(property, 4); engine == 0 {
+		t.Fatal(
+			"the engine was expected to refute a next obligation that never held",
+		)
+	}
+	if reduced := observe(singleStateFormula(property), 4); reduced != 0 {
+		t.Errorf(
+			"single-state refuted at observation %d, and one observation cannot see a next",
+			reduced,
+		)
+	}
+	if reduced := observe(singleStepFormula(property), 4); reduced == 0 {
+		t.Error("single-step was expected to refute a one-step obligation")
+	}
+}
+
+func TestSingleStateRefutesAPredicateReadAtOneObservation(t *testing.T) {
+	property := ltl.Always(constant(false))
+
+	if reduced := observe(singleStateFormula(property), 3); reduced != 1 {
+		t.Errorf(
+			"single-state latched at %d, want the first observation",
+			reduced,
+		)
+	}
+}
+
+func TestSingleStateKeepsANegatedTemporalHarmless(t *testing.T) {
+	property := ltl.Always(ltl.Not(ltl.Eventually(constant(false))))
+
+	if reduced := observe(singleStateFormula(property), 3); reduced != 0 {
+		t.Errorf(
+			"single-state refuted at observation %d; erasing a negated eventually must not manufacture a violation",
+			reduced,
+		)
+	}
+}
+
+func TestSingleStepShrinksAReachabilityGoalToTwoObservations(t *testing.T) {
+	property := ltl.EventuallyWithinSteps(constant(false), 500)
+
+	if engine := observe(property, 4); engine != 5 {
+		t.Fatalf(
+			"a 500-step window closes only at run end, latched at %d",
+			engine,
+		)
+	}
+	if reduced := observe(singleStepFormula(property), 4); reduced != 2 {
+		t.Errorf(
+			"single-step latched at %d, want the second observation",
+			reduced,
+		)
+	}
+	if !truncatesWindow(property) {
+		t.Error(
+			"a 500-step window shortened to two observations must be reported as truncated",
+		)
+	}
+}
+
+// sequence returns a predicate that reads the given values, one per call, so
+// each evaluator gets its own reading counter.
+func sequence(readings []bool) ltl.Formula {
+	position := 0
+	return ltl.ThunkNamed("p", func() (bool, error) {
+		value := readings[position]
+		position++
+		return value, nil
+	})
+}
+
+func TestSingleStepConvictsAGoalTheEngineSeesReached(t *testing.T) {
+	readings := []bool{false, false, true, true}
+
+	if engine := observe(ltl.EventuallyWithinSteps(sequence(readings), 4), 4); engine != 0 {
+		t.Fatalf(
+			"the engine latched at %d, and the goal is reached inside its window",
+			engine,
+		)
+	}
+	if reduced := observe(singleStepFormula(ltl.EventuallyWithinSteps(sequence(readings), 4)), 4); reduced != 2 {
+		t.Errorf(
+			"single-step latched at %d; an obligation armed at the first observation must not reach the third",
+			reduced,
+		)
+	}
+}
+
+func TestTruncatesWindowIgnoresAnObligationThatAlreadyFits(t *testing.T) {
+	property := ltl.Always(
+		ltl.Implies(ltl.Now(constant(true)), ltl.Next(constant(true))),
+	)
+
+	if truncatesWindow(property) {
+		t.Error("a next spans two observations already and is not truncated")
+	}
+}
+
+func TestPropertyClassAndFormComeFromTheTopLevel(t *testing.T) {
+	safety := ltl.Always(constant(true))
+	liveness := ltl.EventuallyWithinSteps(constant(true), 575)
+
+	if got := propertyClass(safety); got != "safety" {
+		t.Errorf("class of an always: got %q", got)
+	}
+	if got := propertyClass(liveness); got != "liveness" {
+		t.Errorf("class of an eventually: got %q", got)
+	}
+	if got := topLevelForm(liveness); got != "eventually within 575 steps" {
+		t.Errorf("form: got %q", got)
+	}
+	if got := topLevelForm(ltl.EventuallyWithin(constant(true), 3*time.Second)); got != "eventually within 3s" {
+		t.Errorf("form: got %q", got)
+	}
+}
+
+func TestSingleStepCannotExpressAWindowLongerThanATriple(t *testing.T) {
+	for _, testCase := range []struct {
+		name      string
+		property  ltl.Formula
+		expresses bool
+	}{
+		{"reachability goal in steps", ltl.EventuallyWithinSteps(constant(true), 575), false},
+		{"reachability goal in time", ltl.EventuallyWithin(constant(true), 3*time.Second), false},
+		{"unbounded reachability goal", ltl.Eventually(constant(true)), false},
+		{"window a triple spans exactly", ltl.EventuallyWithinSteps(constant(true), tripleWindow), true},
+		{"deadline nested under an always", ltl.Always(ltl.Implies(
+			ltl.Now(constant(true)), ltl.EventuallyWithin(constant(true), 3*time.Second))), false},
+		{"one-step obligation under an always", ltl.Always(ltl.Implies(
+			ltl.Now(constant(true)), ltl.Next(constant(true)))), true},
+		{"predicate under an always", ltl.Always(constant(true)), true},
+	} {
+		t.Run(testCase.name, func(t *testing.T) {
+			if singleStepExpresses(testCase.property) != testCase.expresses {
+				t.Errorf("single-step expresses %s: got %t, want %t",
+					testCase.name, !testCase.expresses, testCase.expresses)
+			}
+		})
+	}
+}
+
+// A shorter window under a negation asks for less rather than more, so the
+// triple's refutations of it stay sound and the property stays expressible.
+// Reporting it as inexpressible would push a defect the single-step oracle can
+// genuinely catch into the temporal-only column.
+func TestSingleStepStillExpressesANegatedWindow(t *testing.T) {
+	property := ltl.Always(
+		ltl.Not(ltl.EventuallyWithinSteps(constant(false), 575)),
+	)
+
+	if !singleStepExpresses(property) {
+		t.Error(
+			"a window shortened under a negation is weakened, not strengthened",
+		)
+	}
+	if truncatesWindow(property) {
+		t.Error(
+			"the truncation marker must not fire where shortening only weakens",
+		)
+	}
+}
+
+func TestSingleStateCannotExpressWhatItsRewriteEmpties(t *testing.T) {
+	for _, testCase := range []struct {
+		name      string
+		property  ltl.Formula
+		expresses bool
+	}{
+		{"predicate at one observation", ltl.Always(constant(true)), true},
+		{"reachability goal", ltl.EventuallyWithinSteps(constant(true), 575), false},
+		{"next obligation under an always", ltl.Always(ltl.Implies(
+			ltl.Now(constant(true)), ltl.Next(constant(true)))), false},
+		{"deadline under an always", ltl.Always(ltl.Implies(
+			ltl.Now(constant(true)), ltl.EventuallyWithin(constant(true), 3*time.Second))), false},
+		{"negated eventually under an always", ltl.Always(ltl.Not(ltl.Eventually(constant(false)))), false},
+		{"predicate conjoined with a next", ltl.Always(ltl.And(constant(true), ltl.Next(constant(true)))), true},
+	} {
+		t.Run(testCase.name, func(t *testing.T) {
+			if singleStateExpresses(testCase.property) != testCase.expresses {
+				t.Errorf("single-state expresses %s: got %t, want %t",
+					testCase.name, !testCase.expresses, testCase.expresses)
+			}
+		})
+	}
+}