feat(reconcile): merge two transcripts with a model choosing between disagreements

Base text stands except where the model picks the alternative's reading; trivial
disputes skip the model; the reading order alternates; the base's context goes to
the model. Fails soft: three consecutive model errors return the base text
unreconciled (none of the partial changes) with the reason; a verdict returned with
an error is ignored; a cancelled job is an error, not 'skipped'. Counts are exact,
the stored list of contested spots is capped at 200 and carries the second each
starts at. Mutation-checked on nine behaviours.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

internal/domain/domain.go

@@ -129,6 +129,50 @@ type Transcription struct {
 	Segments     []Segment
 	Endpoint     string
 	AudioSeconds float64
+	// Reconciliation is set when two engines transcribed the same audio and a local
+	// model chose between their disagreements (ADR-034). Nil for a single engine.
+	Reconciliation *Reconciliation
+}
+
+// ReconcileStatus says what became of an attempt to merge two transcripts.
+type ReconcileStatus string
+
+const (
+	// ReconcileDone: both engines produced text and the disagreements were resolved.
+	ReconcileDone ReconcileStatus = "reconciled"
+	// ReconcileSingle: only one engine produced text, so there was nothing to merge.
+	ReconcileSingle ReconcileStatus = "single_engine"
+	// ReconcileSkipped: both engines produced text but the model could not be used;
+	// the base text stands, unreconciled.
+	ReconcileSkipped ReconcileStatus = "skipped"
+)
+
+// ContestedSpot is a place the engines disagreed that a reader may want to check
+// against the audio: one the model changed, or one it was unsure about.
+type ContestedSpot struct {
+	Seconds float64 `json:"seconds"` // roughly where in the recording
+	Base    string  `json:"base"`    // what the base engine heard
+	Alt     string  `json:"alt"`     // what the other engine heard
+	Used    string  `json:"used"`    // "base" or "alt": the reading now in the text
+	Unsure  bool    `json:"unsure"`  // the model could not decide; the base was kept
+}
+
+// Reconciliation records how a merged transcript came about, so the page can say so
+// plainly and the user can check the uncertain spots. The counts are exact; Contested
+// is bounded.
+type Reconciliation struct {
+	Status       ReconcileStatus `json:"status"`
+	Reason       string          `json:"reason,omitempty"` // why single_engine or skipped
+	Base         string          `json:"base"`             // engine whose text is the base
+	Alt          string          `json:"alt"`
+	Reconciler   string          `json:"reconciler"`
+	Disputes     int             `json:"disputes"`      // places the engines disagreed
+	AutoResolved int             `json:"auto_resolved"` // trivial: compounds, filler words
+	Adjudicated  int             `json:"adjudicated"`   // decided by the model
+	Changed      int             `json:"changed"`       // the alternative's reading was used
+	Unsure       int             `json:"unsure"`        // base kept, flagged
+	Errors       int             `json:"errors"`        // model calls that failed
+	Contested    []ContestedSpot `json:"contested,omitempty"`
 }
 
 // UploadStatus is the state of an uploaded-audio transcription job.

internal/reconcile/adjudicate.go

@@ -0,0 +1,119 @@
+package reconcile
+
+import (
+	"context"
+	"fmt"
+	"regexp"
+	"strings"
+)
+
+// Verdict is an adjudicator's decision about one dispute. The zero value is
+// VerdictUnsure: with no decision the base reading stands and the spot is flagged.
+type Verdict int
+
+const (
+	VerdictUnsure Verdict = iota
+	VerdictBase
+	VerdictAlt
+)
+
+// Request is one dispute as the adjudicator sees it: the surrounding words, the two
+// readings, and which of them is shown as "Reading 1". Alternating AltFirst between
+// disputes averages out position bias.
+type Request struct {
+	Before, After string
+	Base, Alt     string
+	AltFirst      bool
+}
+
+// Adjudicator decides which reading of a dispute the speaker most likely said.
+type Adjudicator interface {
+	Adjudicate(ctx context.Context, r Request) (Verdict, error)
+}
+
+// Completer is the chat surface the LLM adjudicator needs. *llm.Client satisfies it.
+type Completer interface {
+	Complete(ctx context.Context, system, user string) (string, error)
+}
+
+// systemPrompt is the wording that was MEASURED (docs/research/
+// local-model-adjudicator-sweep-2026-10-01.md): gemma4-26b 92% and gemma4-31b 97% against
+// the speaker's blind answers. Changing it invalidates those figures.
+const systemPrompt = "You compare two automatic transcriptions of the same Swedish speech. A single person is dictating project notes " +
+	"(finance and payments jargon, with English terms mixed into Swedish). The two transcriptions differ at ONE marked spot. " +
+	"Decide which reading is what the speaker most likely actually said, using the surrounding text and ordinary Swedish. " +
+	"Reply with exactly one line, either 'ANSWER: 1', 'ANSWER: 2' or 'ANSWER: UNSURE', then one short sentence of reasoning."
+
+var answerRe = regexp.MustCompile(`(?i)ANSWER:\s*(1|2|UNSURE)\b`)
+
+// LLMAdjudicator asks a chat model. It never invents an answer: a reply it cannot read
+// is VerdictUnsure, and a transport error is returned as an error so the caller can tell
+// "the model is down" from "the model is unsure".
+type LLMAdjudicator struct{ c Completer }
+
+// NewLLMAdjudicator wraps a Completer.
+func NewLLMAdjudicator(c Completer) *LLMAdjudicator { return &LLMAdjudicator{c: c} }
+
+// Adjudicate implements Adjudicator.
+func (a *LLMAdjudicator) Adjudicate(ctx context.Context, r Request) (Verdict, error) {
+	first, second := r.Base, r.Alt
+	if r.AltFirst {
+		first, second = r.Alt, r.Base
+	}
+	user := fmt.Sprintf("Text before the spot:\n%s\n\n[[ THE SPOT ]]\n\nText after the spot:\n%s\n\nReading 1: %s\nReading 2: %s\n",
+		r.Before, r.After, describe(first), describe(second))
+	reply, err := a.c.Complete(ctx, systemPrompt, user)
+	if err != nil {
+		return VerdictUnsure, fmt.Errorf("reconcile: adjudicate: %w", err)
+	}
+	m := answerRe.FindStringSubmatch(reply)
+	if m == nil {
+		return VerdictUnsure, nil
+	}
+	switch strings.ToUpper(m[1]) {
+	case "1":
+		if r.AltFirst {
+			return VerdictAlt, nil
+		}
+		return VerdictBase, nil
+	case "2":
+		if r.AltFirst {
+			return VerdictBase, nil
+		}
+		return VerdictAlt, nil
+	}
+	return VerdictUnsure, nil
+}
+
+func describe(reading string) string {
+	if strings.TrimSpace(reading) == "" {
+		return "(nothing: the words are simply absent)"
+	}
+	return reading
+}
+
+// OrderChecked asks the inner adjudicator twice with the reading order swapped and
+// keeps a verdict only when both agree. A model that flips with the order is showing
+// position bias, not judgement, so disagreement is VerdictUnsure. It doubles the time.
+type OrderChecked struct{ inner Adjudicator }
+
+// NewOrderChecked wraps an adjudicator.
+func NewOrderChecked(inner Adjudicator) *OrderChecked { return &OrderChecked{inner: inner} }
+
+// Adjudicate implements Adjudicator.
+func (o *OrderChecked) Adjudicate(ctx context.Context, r Request) (Verdict, error) {
+	r.AltFirst = false
+	v1, err := o.inner.Adjudicate(ctx, r)
+	if err != nil {
+		return VerdictUnsure, err
+	}
+	r.AltFirst = true
+	v2, err := o.inner.Adjudicate(ctx, r)
+	if err != nil {
+		return VerdictUnsure, err
+	}
+	if v1 == v2 {
+		return v1, nil
+	}
+	return VerdictUnsure, nil
+}

internal/reconcile/adjudicate_test.go

@@ -0,0 +1,147 @@
+package reconcile_test
+
+import (
+	"context"
+	"errors"
+	"strings"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/reconcile"
+)
+
+type fakeCompleter struct {
+	replies []string
+	err     error
+	systems []string
+	users   []string
+}
+
+func (f *fakeCompleter) Complete(_ context.Context, system, user string) (string, error) {
+	f.systems = append(f.systems, system)
+	f.users = append(f.users, user)
+	if f.err != nil {
+		return "", f.err
+	}
+	i := len(f.users) - 1
+	if i >= len(f.replies) {
+		i = len(f.replies) - 1
+	}
+	return f.replies[i], nil
+}
+
+var req = reconcile.Request{Before: "kunderna ska då", After: "De ska kunna betala", Base: "onboardade", Alt: "omvårdade"}
+
+func TestLLMAdjudicator_SendsTheMeasuredPromptWithTheReadingsInTheRequestedOrder(t *testing.T) {
+	fc := &fakeCompleter{replies: []string{"ANSWER: 1"}}
+	a := reconcile.NewLLMAdjudicator(fc)
+
+	_, err := a.Adjudicate(context.Background(), req)
+
+	require.NoError(t, err)
+	require.Len(t, fc.users, 1)
+	assert.Contains(t, fc.systems[0], "two automatic transcriptions of the same Swedish speech")
+	assert.Contains(t, fc.systems[0], "ANSWER: 1")
+	u := fc.users[0]
+	assert.Contains(t, u, "Text before the spot:\nkunderna ska då")
+	assert.Contains(t, u, "[[ THE SPOT ]]")
+	assert.Contains(t, u, "Text after the spot:\nDe ska kunna betala")
+	assert.Contains(t, u, "Reading 1: onboardade\nReading 2: omvårdade", "base first when AltFirst is false")
+}
+
+func TestLLMAdjudicator_AltFirstSwapsWhichReadingIsNumberOne(t *testing.T) {
+	fc := &fakeCompleter{replies: []string{"ANSWER: 1"}}
+	r := req
+	r.AltFirst = true
+
+	_, err := reconcile.NewLLMAdjudicator(fc).Adjudicate(context.Background(), r)
+
+	require.NoError(t, err)
+	assert.Contains(t, fc.users[0], "Reading 1: omvårdade\nReading 2: onboardade")
+}
+
+func TestLLMAdjudicator_MapsTheAnswerBackToBaseOrAltWhicheverOrderWasShown(t *testing.T) {
+	tests := []struct {
+		name     string
+		altFirst bool
+		reply    string
+		want     reconcile.Verdict
+	}{
+		{"1 with base first is base", false, "ANSWER: 1", reconcile.VerdictBase},
+		{"2 with base first is alt", false, "ANSWER: 2", reconcile.VerdictAlt},
+		{"1 with alt first is alt", true, "ANSWER: 1", reconcile.VerdictAlt},
+		{"2 with alt first is base", true, "ANSWER: 2", reconcile.VerdictBase},
+		{"unsure is unsure", false, "ANSWER: UNSURE", reconcile.VerdictUnsure},
+		{"case and trailing reasoning are tolerated", false, "answer: 2\nThe second reading fits the context.", reconcile.VerdictAlt},
+		{"reasoning before the answer line is tolerated", true, "Hmm.\nANSWER: 2", reconcile.VerdictBase},
+		{"no answer line is unsure, not a guess", false, "I think the first one.", reconcile.VerdictUnsure},
+		{"an empty reply is unsure", false, "", reconcile.VerdictUnsure},
+		{"an out of range number is unsure", false, "ANSWER: 3", reconcile.VerdictUnsure},
+		{"a number that merely starts with 1 is not 1", false, "ANSWER: 12", reconcile.VerdictUnsure},
+		{"a word that merely starts with 2 is not 2", false, "ANSWER: 2nd option", reconcile.VerdictUnsure},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			r := req
+			r.AltFirst = tt.altFirst
+
+			got, err := reconcile.NewLLMAdjudicator(&fakeCompleter{replies: []string{tt.reply}}).Adjudicate(context.Background(), r)
+
+			require.NoError(t, err)
+			assert.Equal(t, tt.want, got)
+		})
+	}
+}
+
+func TestLLMAdjudicator_AnEmptyReadingIsDescribedNotLeftBlank(t *testing.T) {
+	fc := &fakeCompleter{replies: []string{"ANSWER: 1"}}
+	r := req
+	r.Base = ""
+
+	_, err := reconcile.NewLLMAdjudicator(fc).Adjudicate(context.Background(), r)
+
+	require.NoError(t, err)
+	assert.Contains(t, fc.users[0], "Reading 1: (nothing: the words are simply absent)")
+}
+
+func TestLLMAdjudicator_ACompleterErrorIsReturnedNotTurnedIntoAVerdict(t *testing.T) {
+	boom := errors.New("gateway down")
+
+	_, err := reconcile.NewLLMAdjudicator(&fakeCompleter{err: boom}).Adjudicate(context.Background(), req)
+
+	assert.ErrorIs(t, err, boom, "the caller must be able to tell 'the model is down' from 'the model is unsure'")
+}
+
+func TestOrderChecked_AsksTwiceWithTheOrderSwappedAndKeepsOnlyAgreement(t *testing.T) {
+	tests := []struct {
+		name    string
+		replies []string // for [base-first call, alt-first call]
+		want    reconcile.Verdict
+	}{
+		{"both orders pick the base", []string{"ANSWER: 1", "ANSWER: 2"}, reconcile.VerdictBase},
+		{"both orders pick the alternative", []string{"ANSWER: 2", "ANSWER: 1"}, reconcile.VerdictAlt},
+		{"it flips with the order, so it is position bias, not a judgement", []string{"ANSWER: 1", "ANSWER: 1"}, reconcile.VerdictUnsure},
+		{"one order unsure is unsure", []string{"ANSWER: 1", "ANSWER: UNSURE"}, reconcile.VerdictUnsure},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			fc := &fakeCompleter{replies: tt.replies}
+
+			got, err := reconcile.NewOrderChecked(reconcile.NewLLMAdjudicator(fc)).Adjudicate(context.Background(), req)
+
+			require.NoError(t, err)
+			assert.Equal(t, tt.want, got)
+			require.Len(t, fc.users, 2)
+			assert.True(t, strings.Contains(fc.users[0], "Reading 1: onboardade"), "first call shows the base first")
+			assert.True(t, strings.Contains(fc.users[1], "Reading 1: omvårdade"), "second call shows the alternative first")
+		})
+	}
+}
+
+func TestOrderChecked_AnErrorOnEitherCallIsAnError(t *testing.T) {
+	_, err := reconcile.NewOrderChecked(reconcile.NewLLMAdjudicator(&fakeCompleter{err: errors.New("down")})).Adjudicate(context.Background(), req)
+
+	assert.Error(t, err)
+}

internal/reconcile/diff.go

@@ -0,0 +1,187 @@
+package reconcile
+
+// OpKind is the kind of one run in an edit script.
+type OpKind int
+
+const (
+	// Equal: A[AStart:AEnd] and B[BStart:BEnd] are the same words.
+	Equal OpKind = iota
+	// Replace: different words on both sides.
+	Replace
+	// Delete: words only in A.
+	Delete
+	// Insert: words only in B.
+	Insert
+)
+
+// Op is one run of the edit script. The ops of a Diff tile both inputs: each op starts
+// where the previous one ended, in A and in B.
+type Op struct {
+	Kind         OpKind
+	AStart, AEnd int
+	BStart, BEnd int
+}
+
+// maxEdits bounds the Myers search. Memory grows with the square of the edit
+// distance, so two texts that share almost nothing (one engine produced garbage)
+// are reported as one Replace instead of spending gigabytes proving it. A real
+// hour disagrees in roughly 400 places, i.e. about 800 edits, far below this.
+const maxEdits = 3000
+
+// Diff returns a minimal edit script from a to b (Myers, O(ND)). Adjacent deletions
+// and insertions are reported as one Replace.
+func Diff(a, b []string) []Op {
+	pre := 0
+	for pre < len(a) && pre < len(b) && a[pre] == b[pre] {
+		pre++
+	}
+	suf := 0
+	for suf < len(a)-pre && suf < len(b)-pre && a[len(a)-1-suf] == b[len(b)-1-suf] {
+		suf++
+	}
+	midA, midB := a[pre:len(a)-suf], b[pre:len(b)-suf]
+
+	steps, ok := myers(midA, midB)
+	if !ok {
+		steps = steps[:0]
+		for range midA {
+			steps = append(steps, stepDel)
+		}
+		for range midB {
+			steps = append(steps, stepIns)
+		}
+	}
+
+	var ops []Op
+	ai, bi := 0, 0
+	emit := func(k OpKind, da, db int) {
+		if da == 0 && db == 0 {
+			return
+		}
+		if k == Equal && len(ops) > 0 && ops[len(ops)-1].Kind == Equal {
+			ops[len(ops)-1].AEnd += da
+			ops[len(ops)-1].BEnd += db
+		} else {
+			ops = append(ops, Op{Kind: k, AStart: ai, AEnd: ai + da, BStart: bi, BEnd: bi + db})
+		}
+		ai += da
+		bi += db
+	}
+	emit(Equal, pre, pre)
+	for i := 0; i < len(steps); {
+		if steps[i] == stepEq {
+			j := i
+			for j < len(steps) && steps[j] == stepEq {
+				j++
+			}
+			emit(Equal, j-i, j-i)
+			i = j
+			continue
+		}
+		dels, inss := 0, 0
+		for i < len(steps) && steps[i] != stepEq {
+			if steps[i] == stepDel {
+				dels++
+			} else {
+				inss++
+			}
+			i++
+		}
+		switch {
+		case dels > 0 && inss > 0:
+			emit(Replace, dels, inss)
+		case dels > 0:
+			emit(Delete, dels, 0)
+		default:
+			emit(Insert, 0, inss)
+		}
+	}
+	emit(Equal, suf, suf)
+	return ops
+}
+
+type step int8
+
+const (
+	stepEq step = iota
+	stepDel
+	stepIns
+)
+
+// myers returns one step per element of the edit script, in order. ok is false when
+// the edit distance exceeds maxEdits.
+func myers(a, b []string) ([]step, bool) {
+	n, m := len(a), len(b)
+	if n == 0 && m == 0 {
+		return nil, true
+	}
+	limit := min(n+m, maxEdits)
+	off := limit + 1
+	v := make([]int32, 2*limit+3)
+	var trace [][]int32 // trace[d] is v as it stood at the start of round d, for k in [-d, d]
+
+	for d := 0; d <= limit; d++ {
+		snap := make([]int32, 2*d+1)
+		for k := -d; k <= d; k++ {
+			snap[k+d] = v[off+k]
+		}
+		trace = append(trace, snap)
+		for k := -d; k <= d; k += 2 {
+			var x int
+			if k == -d || (k != d && v[off+k-1] < v[off+k+1]) {
+				x = int(v[off+k+1])
+			} else {
+				x = int(v[off+k-1]) + 1
+			}
+			y := x - k
+			for x < n && y < m && a[x] == b[y] {
+				x++
+				y++
+			}
+			v[off+k] = int32(x)
+			if x >= n && y >= m {
+				return backtrack(trace, n, m, d), true
+			}
+		}
+	}
+	return nil, false
+}
+
+func backtrack(trace [][]int32, n, m, d int) []step {
+	var rev []step
+	x, y := n, m
+	for ; d >= 0; d-- {
+		snap := trace[d]
+		at := func(k int) int { return int(snap[k+d]) }
+		k := x - y
+		var prevK int
+		if k == -d || (k != d && at(k-1) < at(k+1)) {
+			prevK = k + 1
+		} else {
+			prevK = k - 1
+		}
+		prevX := 0
+		if d > 0 {
+			prevX = at(prevK)
+		}
+		prevY := prevX - prevK
+		for x > prevX && y > prevY {
+			rev = append(rev, stepEq)
+			x--
+			y--
+		}
+		if d > 0 {
+			if x == prevX {
+				rev = append(rev, stepIns)
+			} else {
+				rev = append(rev, stepDel)
+			}
+		}
+		x, y = prevX, prevY
+	}
+	out := make([]step, len(rev))
+	for i, s := range rev {
+		out[len(rev)-1-i] = s
+	}
+	return out
+}

internal/reconcile/diff_test.go

@@ -0,0 +1,139 @@
+package reconcile_test
+
+import (
+	"math/rand"
+	"strings"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/reconcile"
+)
+
+func words(s string) []string { return strings.Fields(s) }
+
+// apply rebuilds b from a using the ops, copying equal runs from a and the
+// replacement/insertion runs from b. If Diff is right this must equal b.
+func apply(a, b []string, ops []reconcile.Op) []string {
+	var out []string
+	for _, o := range ops {
+		switch o.Kind {
+		case reconcile.Equal:
+			out = append(out, a[o.AStart:o.AEnd]...)
+		default:
+			out = append(out, b[o.BStart:o.BEnd]...)
+		}
+	}
+	return out
+}
+
+func TestDiff_Cases(t *testing.T) {
+	tests := []struct {
+		name string
+		a, b string
+		want []reconcile.Op
+	}{
+		{"identical", "a b c", "a b c", []reconcile.Op{{Kind: reconcile.Equal, AStart: 0, AEnd: 3, BStart: 0, BEnd: 3}}},
+		{"both empty", "", "", nil},
+		{"one word replaced", "a b c", "a x c", []reconcile.Op{
+			{Kind: reconcile.Equal, AStart: 0, AEnd: 1, BStart: 0, BEnd: 1},
+			{Kind: reconcile.Replace, AStart: 1, AEnd: 2, BStart: 1, BEnd: 2},
+			{Kind: reconcile.Equal, AStart: 2, AEnd: 3, BStart: 2, BEnd: 3},
+		}},
+		{"word only in b", "a c", "a b c", []reconcile.Op{
+			{Kind: reconcile.Equal, AStart: 0, AEnd: 1, BStart: 0, BEnd: 1},
+			{Kind: reconcile.Insert, AStart: 1, AEnd: 1, BStart: 1, BEnd: 2},
+			{Kind: reconcile.Equal, AStart: 1, AEnd: 2, BStart: 2, BEnd: 3},
+		}},
+		{"word only in a", "a b c", "a c", []reconcile.Op{
+			{Kind: reconcile.Equal, AStart: 0, AEnd: 1, BStart: 0, BEnd: 1},
+			{Kind: reconcile.Delete, AStart: 1, AEnd: 2, BStart: 1, BEnd: 1},
+			{Kind: reconcile.Equal, AStart: 2, AEnd: 3, BStart: 1, BEnd: 2},
+		}},
+		{"nothing in common", "a b", "x y z", []reconcile.Op{
+			{Kind: reconcile.Replace, AStart: 0, AEnd: 2, BStart: 0, BEnd: 3},
+		}},
+		{"a is empty", "", "x y", []reconcile.Op{{Kind: reconcile.Insert, AStart: 0, AEnd: 0, BStart: 0, BEnd: 2}}},
+		{"b is empty", "x y", "", []reconcile.Op{{Kind: reconcile.Delete, AStart: 0, AEnd: 2, BStart: 0, BEnd: 0}}},
+		{"adjacent delete and insert read as one replace", "a b c d", "a x y d", []reconcile.Op{
+			{Kind: reconcile.Equal, AStart: 0, AEnd: 1, BStart: 0, BEnd: 1},
+			{Kind: reconcile.Replace, AStart: 1, AEnd: 3, BStart: 1, BEnd: 3},
+			{Kind: reconcile.Equal, AStart: 3, AEnd: 4, BStart: 3, BEnd: 4},
+		}},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			assert.Equal(t, tt.want, reconcile.Diff(words(tt.a), words(tt.b)))
+		})
+	}
+}
+
+// lcsLen is a plain dynamic-programming reference. Myers must find an edit script
+// whose equal runs add up to exactly this: a diff that is merely valid but not
+// minimal would turn real agreement into disputes the adjudicator is then asked about.
+func lcsLen(a, b []string) int {
+	prev := make([]int, len(b)+1)
+	for i := 1; i <= len(a); i++ {
+		cur := make([]int, len(b)+1)
+		for j := 1; j <= len(b); j++ {
+			if a[i-1] == b[j-1] {
+				cur[j] = prev[j-1] + 1
+			} else {
+				cur[j] = max(prev[j], cur[j-1])
+			}
+		}
+		prev = cur
+	}
+	return prev[len(b)]
+}
+
+func TestDiff_RebuildsBAndKeepsEveryCommonWordOnRandomInputs(t *testing.T) {
+	rng := rand.New(rand.NewSource(7))
+	alphabet := []string{"a", "b", "c", "d", "e", "f"}
+	gen := func(n int) []string {
+		out := make([]string, n)
+		for i := range out {
+			out[i] = alphabet[rng.Intn(len(alphabet))]
+		}
+		return out
+	}
+	for i := 0; i < 400; i++ {
+		a, b := gen(rng.Intn(25)), gen(rng.Intn(25))
+
+		ops := reconcile.Diff(a, b)
+
+		require.Equal(t, append([]string{}, b...), append([]string{}, apply(a, b, ops)...), "a=%v b=%v ops=%v", a, b, ops)
+		common := 0
+		prevA, prevB := 0, 0
+		for _, o := range ops {
+			require.Equal(t, prevA, o.AStart, "ops must tile a without gaps: %v", ops)
+			require.Equal(t, prevB, o.BStart, "ops must tile b without gaps: %v", ops)
+			prevA, prevB = o.AEnd, o.BEnd
+			if o.Kind == reconcile.Equal {
+				common += o.AEnd - o.AStart
+				require.Equal(t, a[o.AStart:o.AEnd], b[o.BStart:o.BEnd], "an Equal op must really be equal")
+			}
+		}
+		require.Equal(t, len(a), prevA)
+		require.Equal(t, len(b), prevB)
+		require.Equal(t, lcsLen(a, b), common, "the diff must be minimal: a=%v b=%v", a, b)
+	}
+}
+
+func TestDiff_HandlesAnHoursWorthOfWordsFast(t *testing.T) {
+	rng := rand.New(rand.NewSource(3))
+	a := make([]string, 6500)
+	for i := range a {
+		a[i] = string(rune('a'+rng.Intn(26))) + string(rune('a'+rng.Intn(26)))
+	}
+	b := append([]string{}, a...)
+	for i := 0; i < 380; i++ { // ~380 disputes, as measured on a real hour
+		b[rng.Intn(len(b))] = "zz" + string(rune('a'+rng.Intn(26)))
+	}
+
+	ops := reconcile.Diff(a, b)
+
+	assert.NotEmpty(t, ops)
+	assert.Equal(t, append([]string{}, b...), append([]string{}, apply(a, b, ops)...))
+}

internal/reconcile/disputes.go

@@ -0,0 +1,100 @@
+package reconcile
+
+import "strings"
+
+// Dispute is one place where the two transcripts disagree. BaseStart:BaseEnd and
+// AltStart:AltEnd are token ranges into the two Tokenize results; an empty range
+// means that engine heard nothing there. Base and Alt are the two readings as each
+// engine wrote them, so a replaced span takes the alternative's own spelling.
+type Dispute struct {
+	BaseStart, BaseEnd int
+	AltStart, AltEnd   int
+	Base, Alt          string
+}
+
+// FindDisputes aligns base against alt word by word and returns the disagreements.
+// Differences separated by a single shared word are merged into one dispute: for a
+// reader that is one decision, and asking about "b" and "d" separately in "b c d"
+// hides that "c" sits between them. Two shared words keep disputes apart.
+func FindDisputes(base, alt []Token) []Dispute {
+	a, b := keysOf(base), keysOf(alt)
+	ops := Diff(a, b)
+	var out []Dispute
+	for i := 0; i < len(ops); i++ {
+		if ops[i].Kind == Equal {
+			continue
+		}
+		cur := ops[i]
+		for i+2 < len(ops) && ops[i+1].Kind == Equal && ops[i+1].AEnd-ops[i+1].AStart == 1 && ops[i+2].Kind != Equal {
+			cur.AEnd, cur.BEnd = ops[i+2].AEnd, ops[i+2].BEnd
+			i += 2
+		}
+		out = append(out, Dispute{
+			BaseStart: cur.AStart, BaseEnd: cur.AEnd,
+			AltStart: cur.BStart, AltEnd: cur.BEnd,
+			Base: join(base[cur.AStart:cur.AEnd]),
+			Alt:  join(alt[cur.BStart:cur.BEnd]),
+		})
+	}
+	return out
+}
+
+// fillers are Swedish discourse particles. A dispute that is only one or two of them
+// on one side changes no meaning, so it is resolved without asking a model. The list
+// is deliberately short and leaves out words that can be an answer ("nej") or that
+// carry content.
+var fillers = map[string]bool{
+	"ju": true, "då": true, "liksom": true, "alltså": true, "så": true, "ja": true,
+	"eh": true, "ehm": true, "öh": true, "mm": true, "hm": true, "typ": true, "va": true, "asså": true,
+}
+
+// maxFillerRun is how many filler words may be waved through together.
+const maxFillerRun = 2
+
+// AutoResolve reports whether a dispute needs no model, in which case the base
+// reading is kept: either the two sides are the same letters once spaces are removed
+// (a compound written as one word or two), or one side is empty and the other is a
+// short run of filler words.
+func AutoResolve(d Dispute) bool {
+	baseKeys, altKeys := keysOfText(d.Base), keysOfText(d.Alt)
+	if strings.Join(baseKeys, "") == strings.Join(altKeys, "") {
+		return true
+	}
+	return (len(baseKeys) == 0 && onlyFillers(altKeys)) || (len(altKeys) == 0 && onlyFillers(baseKeys))
+}
+
+func onlyFillers(keys []string) bool {
+	if len(keys) == 0 || len(keys) > maxFillerRun {
+		return false
+	}
+	for _, k := range keys {
+		if !fillers[k] {
+			return false
+		}
+	}
+	return true
+}
+
+// ContextAround returns up to window words of the base before and after the dispute,
+// as the base wrote them.
+func ContextAround(base []Token, d Dispute, window int) (before, after string) {
+	return join(base[max(0, d.BaseStart-window):d.BaseStart]), join(base[d.BaseEnd:min(len(base), d.BaseEnd+window)])
+}
+
+func keysOf(ts []Token) []string {
+	out := make([]string, len(ts))
+	for i, t := range ts {
+		out[i] = t.Key
+	}
+	return out
+}
+
+func keysOfText(s string) []string { return keysOf(Tokenize(s)) }
+
+func join(ts []Token) string {
+	parts := make([]string, len(ts))
+	for i, t := range ts {
+		parts[i] = t.Text
+	}
+	return strings.Join(parts, " ")
+}

internal/reconcile/disputes_test.go

@@ -0,0 +1,131 @@
+package reconcile_test
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/reconcile"
+)
+
+func disputes(base, alt string) []reconcile.Dispute {
+	return reconcile.FindDisputes(reconcile.Tokenize(base), reconcile.Tokenize(alt))
+}
+
+func TestFindDisputes_NoneWhenTheTextsAgreeWordForWord(t *testing.T) {
+	assert.Empty(t, disputes("Vi går igenom läget.", "vi går igenom läget"),
+		"casing and punctuation are not disagreements")
+}
+
+func TestFindDisputes_OneReplacedWordIsOneDisputeCarryingBothReadings(t *testing.T) {
+	got := disputes("Vi ska onboardade kunderna.", "Vi ska omvårdade kunderna.")
+
+	require.Len(t, got, 1)
+	d := got[0]
+	assert.Equal(t, "onboardade", d.Base)
+	assert.Equal(t, "omvårdade", d.Alt)
+	assert.Equal(t, 2, d.BaseStart)
+	assert.Equal(t, 3, d.BaseEnd)
+	assert.Equal(t, 2, d.AltStart)
+	assert.Equal(t, 3, d.AltEnd)
+}
+
+func TestFindDisputes_ReadingsKeepTheirOwnSpellingAndPunctuation(t *testing.T) {
+	got := disputes("Det blir T+1, alltså.", "Det blir T 1, alltså.")
+
+	require.Len(t, got, 1)
+	assert.Equal(t, "T+1,", got[0].Base)
+	assert.Equal(t, "T 1,", got[0].Alt, "the alternative's words replace the span as the engine wrote them")
+}
+
+func TestFindDisputes_AWordOnlyOneEngineHeardHasAnEmptyOtherSide(t *testing.T) {
+	got := disputes("Man får tillbaka ett förslag.", "Man får tillbaka ett coolt förslag.")
+
+	require.Len(t, got, 1)
+	assert.Equal(t, "", got[0].Base)
+	assert.Equal(t, "coolt", got[0].Alt)
+	assert.Equal(t, got[0].BaseStart, got[0].BaseEnd, "an insertion has an empty base range at its position")
+	assert.Equal(t, 4, got[0].BaseStart)
+}
+
+func TestFindDisputes_TwoDifferencesSeparatedByOneSharedWordAreOneDispute(t *testing.T) {
+	// "x ... y" around a single agreed word is one decision for the reader, not two.
+	got := disputes("a b c d e", "a x c y e")
+
+	require.Len(t, got, 1)
+	assert.Equal(t, "b c d", got[0].Base)
+	assert.Equal(t, "x c y", got[0].Alt)
+}
+
+func TestFindDisputes_TwoSharedWordsKeepTheDisputesApart(t *testing.T) {
+	got := disputes("a b c d e f", "a x c d y f")
+
+	require.Len(t, got, 2)
+	assert.Equal(t, "b", got[0].Base)
+	assert.Equal(t, "e", got[1].Base)
+}
+
+func TestFindDisputes_ChainsOfCloseDifferencesMergeIntoOne(t *testing.T) {
+	got := disputes("a b c d e f g", "a x c y e z g")
+
+	require.Len(t, got, 1)
+	assert.Equal(t, "b c d e f", got[0].Base)
+}
+
+func TestFindDisputes_LeavesIndicesThatAddressBothTokenSlices(t *testing.T) {
+	base := reconcile.Tokenize("Hej Anders och Maria")
+	alt := reconcile.Tokenize("Hej Sanders och Maria")
+
+	got := reconcile.FindDisputes(base, alt)
+
+	require.Len(t, got, 1)
+	assert.Equal(t, "Anders", base[got[0].BaseStart].Text)
+	assert.Equal(t, "Sanders", alt[got[0].AltStart].Text)
+}
+
+func TestAutoResolve(t *testing.T) {
+	tests := []struct {
+		name      string
+		base, alt string
+		resolved  bool
+	}{
+		{"a compound written as two words or one", "i crossborder delen", "i crossborderdelen", true},
+		{"a compound with a hyphen", "crossborder-delen", "crossborder delen", true},
+		{"a filler only the alternative heard", "vi ska ju betala", "vi ska betala", true},
+		{"a filler only the base heard", "vi ska betala", "vi ska ju betala", true},
+		{"two fillers in a row", "vi ska ja ju betala", "vi ska betala", true},
+		{"a different word is for the model", "vi ska onboardade", "vi ska omvårdade", false},
+		{"a different number is for the model", "totalt 1050 kronor", "totalt 2050 kronor", false},
+		{"a content word only one side heard is for the model", "ett förslag", "ett coolt förslag", false},
+		{"three filler words are too many to wave through", "vi ska ju då liksom betala", "vi ska betala", false},
+		{"a filler swapped for a content word is for the model", "vi ska ju betala", "vi ska nu betala", false},
+		{"nej is an answer, not a filler", "men nej det går inte", "men det går inte", false},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := disputes(tt.base, tt.alt)
+			require.Len(t, got, 1, "the fixture must be exactly one dispute")
+
+			assert.Equal(t, tt.resolved, reconcile.AutoResolve(got[0]))
+		})
+	}
+}
+
+func TestContextAround_TakesWordsFromTheBaseOnBothSidesAndStopsAtTheEdges(t *testing.T) {
+	base := reconcile.Tokenize("ett två tre fyra fem sex sju åtta nio tio")
+	d := reconcile.Dispute{BaseStart: 4, BaseEnd: 6} // "fem sex"
+
+	before, after := reconcile.ContextAround(base, d, 3)
+
+	assert.Equal(t, "två tre fyra", before)
+	assert.Equal(t, "sju åtta nio", after)
+
+	b2, a2 := reconcile.ContextAround(base, reconcile.Dispute{BaseStart: 0, BaseEnd: 1}, 3)
+	assert.Equal(t, "", b2)
+	assert.Equal(t, "två tre fyra", a2)
+
+	b3, a3 := reconcile.ContextAround(base, reconcile.Dispute{BaseStart: 9, BaseEnd: 10}, 3)
+	assert.Equal(t, "sju åtta nio", b3)
+	assert.Equal(t, "", a3)
+}

internal/reconcile/reconcile.go

@@ -0,0 +1,152 @@
+package reconcile
+
+import (
+	"context"
+	"fmt"
+	"strings"
+
+	"git.d-ma.be/mathias/tapir/internal/domain"
+)
+
+// MaxContested bounds the stored list of contested spots. The counts on the
+// Reconciliation are always exact; only this list is capped.
+const MaxContested = 200
+
+// Options configure Reconcile. Zero Window and BreakAfter take their defaults.
+type Options struct {
+	// Window is how many words of the base to show on each side of a dispute. 100 is
+	// what was measured (docs/research/local-model-adjudicator-sweep-2026-10-01.md).
+	Window int
+	// BreakAfter is how many consecutive model errors stop the reconciliation, so a
+	// model that is down costs three timeouts, not one per dispute.
+	BreakAfter int
+	// BaseEngine, AltEngine and Reconciler are recorded on the result.
+	BaseEngine, AltEngine, Reconciler string
+}
+
+// Reconcile merges two transcripts of the same speech. The base text stands except
+// where the model chooses the alternative's reading of a dispute; trivial disputes
+// (a compound split, a filler word) are resolved without the model. It never fails a
+// transcription: if the model cannot be used the base text comes back unreconciled
+// and the result says why. The only error it returns is a cancelled context, which
+// is the job stopping, not a verdict on the model.
+func Reconcile(ctx context.Context, base, alt domain.Transcription, adj Adjudicator, o Options) (domain.Transcription, error) {
+	window, breakAfter := o.Window, o.BreakAfter
+	if window <= 0 {
+		window = 100
+	}
+	if breakAfter <= 0 {
+		breakAfter = 3
+	}
+	bt, at := Tokenize(base.Text), Tokenize(alt.Text)
+	ds := FindDisputes(bt, at)
+	rec := &domain.Reconciliation{
+		Status: domain.ReconcileDone, Base: o.BaseEngine, Alt: o.AltEngine, Reconciler: o.Reconciler, Disputes: len(ds),
+	}
+	seconds := positioner(base, len(bt))
+
+	verdicts := make([]Verdict, len(ds))
+	consecutive := 0
+	for i, d := range ds {
+		if AutoResolve(d) {
+			rec.AutoResolved++
+			verdicts[i] = VerdictBase
+			continue
+		}
+		if err := ctx.Err(); err != nil {
+			return domain.Transcription{}, err
+		}
+		before, after := ContextAround(bt, d, window)
+		v, err := adj.Adjudicate(ctx, Request{Before: before, After: after, Base: d.Base, Alt: d.Alt, AltFirst: i%2 == 1})
+		if err != nil {
+			if ctxErr := ctx.Err(); ctxErr != nil {
+				return domain.Transcription{}, ctxErr
+			}
+			rec.Errors++
+			consecutive++
+			if consecutive >= breakAfter {
+				return skipped(base, rec, fmt.Sprintf("the model failed %d times in a row: %v", consecutive, err)), nil
+			}
+			v = VerdictUnsure
+		} else {
+			consecutive = 0
+			rec.Adjudicated++
+		}
+		verdicts[i] = v
+		switch v {
+		case VerdictAlt:
+			rec.Changed++
+			addContested(rec, domain.ContestedSpot{Seconds: seconds(d.BaseStart), Base: d.Base, Alt: d.Alt, Used: "alt"})
+		case VerdictUnsure:
+			rec.Unsure++
+			addContested(rec, domain.ContestedSpot{Seconds: seconds(d.BaseStart), Base: d.Base, Alt: d.Alt, Used: "base", Unsure: true})
+		}
+	}
+
+	var parts []string
+	pos := 0
+	for i, d := range ds {
+		parts = appendTexts(parts, bt[pos:d.BaseStart])
+		if verdicts[i] == VerdictAlt {
+			parts = appendTexts(parts, at[d.AltStart:d.AltEnd])
+		} else {
+			parts = appendTexts(parts, bt[d.BaseStart:d.BaseEnd])
+		}
+		pos = d.BaseEnd
+	}
+	parts = appendTexts(parts, bt[pos:])
+
+	out := base
+	out.Text = strings.Join(parts, " ")
+	out.Reconciliation = rec
+	return out, nil
+}
+
+// skipped is the base text, none of the partial changes, and the reason.
+func skipped(base domain.Transcription, rec *domain.Reconciliation, reason string) domain.Transcription {
+	out := base
+	out.Reconciliation = &domain.Reconciliation{
+		Status: domain.ReconcileSkipped, Reason: reason, Base: rec.Base, Alt: rec.Alt, Reconciler: rec.Reconciler,
+		Disputes: rec.Disputes, AutoResolved: rec.AutoResolved, Errors: rec.Errors,
+	}
+	return out
+}
+
+func appendTexts(dst []string, ts []Token) []string {
+	for _, t := range ts {
+		dst = append(dst, t.Text)
+	}
+	return dst
+}
+
+// positioner maps a token index of the base to the second it starts at: through the
+// segments when there are any, proportionally otherwise.
+func positioner(base domain.Transcription, tokens int) func(int) float64 {
+	if len(base.Segments) == 0 {
+		return func(i int) float64 {
+			if tokens == 0 {
+				return 0
+			}
+			return float64(i) / float64(tokens) * base.AudioSeconds
+		}
+	}
+	starts := make([]int, len(base.Segments))
+	n := 0
+	for i, s := range base.Segments {
+		starts[i] = n
+		n += len(Tokenize(s.Text))
+	}
+	return func(i int) float64 {
+		k := 0
+		for k+1 < len(starts) && starts[k+1] <= i {
+			k++
+		}
+		return base.Segments[k].Start
+	}
+}
+
+func addContested(rec *domain.Reconciliation, c domain.ContestedSpot) {
+	if len(rec.Contested) < MaxContested {
+		rec.Contested = append(rec.Contested, c)
+	}
+}

internal/reconcile/reconcile_test.go

@@ -0,0 +1,293 @@
+package reconcile_test
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"strings"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/reconcile"
+)
+
+// scripted is an Adjudicator that answers from a function and records every request.
+type scripted struct {
+	fn    func(i int, r reconcile.Request) (reconcile.Verdict, error)
+	calls []reconcile.Request
+}
+
+func (s *scripted) Adjudicate(_ context.Context, r reconcile.Request) (reconcile.Verdict, error) {
+	i := len(s.calls)
+	s.calls = append(s.calls, r)
+	return s.fn(i, r)
+}
+
+func always(v reconcile.Verdict) *scripted {
+	return &scripted{fn: func(int, reconcile.Request) (reconcile.Verdict, error) { return v, nil }}
+}
+
+func tr(text string) domain.Transcription {
+	return domain.Transcription{Text: text, AudioSeconds: 100, Endpoint: "e"}
+}
+
+var opts = reconcile.Options{Window: 100, BreakAfter: 3, Reconciler: "iguana/gemma4-26b", BaseEngine: "iguana/pianissimo", AltEngine: "iguana/kb-whisper"}
+
+func merge(t *testing.T, base, alt domain.Transcription, adj reconcile.Adjudicator) domain.Transcription {
+	t.Helper()
+	out, err := reconcile.Reconcile(context.Background(), base, alt, adj, opts)
+	require.NoError(t, err)
+	return out
+}
+
+func TestReconcile_IdenticalTextsNeedNoModelAndComeBackUnchanged(t *testing.T) {
+	adj := always(reconcile.VerdictAlt)
+
+	out := merge(t, tr("Vi går igenom läget."), tr("vi går igenom läget"), adj)
+
+	assert.Equal(t, "Vi går igenom läget.", out.Text, "the base's own spelling and punctuation stand")
+	assert.Empty(t, adj.calls)
+	require.NotNil(t, out.Reconciliation)
+	assert.Equal(t, domain.ReconcileDone, out.Reconciliation.Status)
+	assert.Zero(t, out.Reconciliation.Disputes)
+}
+
+func TestReconcile_TrivialDisputesAreResolvedWithoutAskingTheModel(t *testing.T) {
+	adj := always(reconcile.VerdictAlt)
+
+	out := merge(t, tr("i crossborder delen ska vi ju betala"), tr("i crossborderdelen ska vi betala"), adj)
+
+	assert.Empty(t, adj.calls, "a compound split and a filler particle are not worth a model call")
+	assert.Equal(t, "i crossborder delen ska vi ju betala", out.Text, "the base reading is kept")
+	assert.Equal(t, 2, out.Reconciliation.Disputes)
+	assert.Equal(t, 2, out.Reconciliation.AutoResolved)
+	assert.Zero(t, out.Reconciliation.Adjudicated)
+}
+
+func TestReconcile_AltVerdictReplacesTheSpanWithTheAlternativesOwnWords(t *testing.T) {
+	out := merge(t, tr("Totalt 1015 kronor. Det ska betalas."), tr("Totalt 1050 kronor. Det ska betalas."), always(reconcile.VerdictAlt))
+
+	assert.Equal(t, "Totalt 1050 kronor. Det ska betalas.", out.Text)
+	assert.Equal(t, 1, out.Reconciliation.Changed)
+	assert.Equal(t, 1, out.Reconciliation.Adjudicated)
+}
+
+func TestReconcile_BaseAndUnsureVerdictsKeepTheBaseButOnlyUnsureIsFlagged(t *testing.T) {
+	base := tr("Totalt 1015 kronor till SEB.")
+	alt := tr("Totalt 1050 kronor till SCB.")
+
+	keptBase := merge(t, base, alt, always(reconcile.VerdictBase))
+	unsure := merge(t, base, alt, always(reconcile.VerdictUnsure))
+
+	assert.Equal(t, "Totalt 1015 kronor till SEB.", keptBase.Text)
+	assert.Zero(t, keptBase.Reconciliation.Unsure)
+	assert.Empty(t, keptBase.Reconciliation.Contested, "a confident 'the base is right' is not a spot worth the reader's time")
+	assert.Equal(t, "Totalt 1015 kronor till SEB.", unsure.Text, "unsure keeps the base")
+	assert.Equal(t, 2, unsure.Reconciliation.Unsure)
+	require.Len(t, unsure.Reconciliation.Contested, 2)
+	assert.True(t, unsure.Reconciliation.Contested[0].Unsure)
+	assert.Equal(t, "base", unsure.Reconciliation.Contested[0].Used)
+}
+
+func TestReconcile_ChangedSpotsAreRecordedWithBothReadings(t *testing.T) {
+	out := merge(t, tr("det blir T+1 imorgon"), tr("det blir en T imorgon"), always(reconcile.VerdictAlt))
+
+	require.Len(t, out.Reconciliation.Contested, 1)
+	c := out.Reconciliation.Contested[0]
+	assert.Equal(t, "T+1", c.Base)
+	assert.Equal(t, "en T", c.Alt)
+	assert.Equal(t, "alt", c.Used)
+	assert.False(t, c.Unsure)
+}
+
+func TestReconcile_AWordOnlyTheAlternativeHeardIsInsertedWhereItBelongs(t *testing.T) {
+	out := merge(t, tr("Man får tillbaka ett förslag."), tr("Man får tillbaka ett coolt förslag."), always(reconcile.VerdictAlt))
+
+	assert.Equal(t, "Man får tillbaka ett coolt förslag.", out.Text)
+}
+
+func TestReconcile_AWordOnlyTheBaseHeardCanBeRemoved(t *testing.T) {
+	out := merge(t, tr("Man får tillbaka ett coolt förslag."), tr("Man får tillbaka ett förslag."), always(reconcile.VerdictAlt))
+
+	assert.Equal(t, "Man får tillbaka ett förslag.", out.Text)
+}
+
+func TestReconcile_SeveralDisputesAreAppliedIndependentlyAndInOrder(t *testing.T) {
+	adj := &scripted{fn: func(i int, r reconcile.Request) (reconcile.Verdict, error) {
+		if r.Alt == "omvårdade" || r.Alt == "SCB" {
+			return reconcile.VerdictAlt, nil
+		}
+		return reconcile.VerdictBase, nil
+	}}
+
+	out := merge(t, tr("kunderna onboardade nu och sen betalas till SEB via ombud med 1050 kronor"),
+		tr("kunderna omvårdade nu och sen betalas till SCB via ombud med 2050 kronor"), adj)
+
+	assert.Equal(t, "kunderna omvårdade nu och sen betalas till SCB via ombud med 1050 kronor", out.Text)
+	assert.Equal(t, 3, out.Reconciliation.Disputes)
+	assert.Equal(t, 2, out.Reconciliation.Changed)
+}
+
+func TestReconcile_TheReadingOrderAlternatesFromOneDisputeToTheNext(t *testing.T) {
+	adj := always(reconcile.VerdictBase)
+
+	merge(t, tr("a x s1 s2 y t1 t2 z d"), tr("a p s1 s2 q t1 t2 r d"), adj)
+
+	require.Len(t, adj.calls, 3)
+	assert.False(t, adj.calls[0].AltFirst)
+	assert.True(t, adj.calls[1].AltFirst)
+	assert.False(t, adj.calls[2].AltFirst, "alternating averages out position bias across a transcript")
+}
+
+func TestReconcile_TheModelGetsTheContextWindowFromTheBase(t *testing.T) {
+	adj := always(reconcile.VerdictBase)
+	o := opts
+	o.Window = 2
+
+	_, err := reconcile.Reconcile(context.Background(), tr("ett två tre fyra fem sex sju åtta nio"), tr("ett två tre fyra FEMM sex sju åtta nio"), adj, o)
+
+	require.NoError(t, err)
+	require.Len(t, adj.calls, 1)
+	assert.Equal(t, "tre fyra", adj.calls[0].Before)
+	assert.Equal(t, "sex sju", adj.calls[0].After)
+	assert.Equal(t, "fem", adj.calls[0].Base)
+	assert.Equal(t, "FEMM", adj.calls[0].Alt)
+}
+
+func TestReconcile_OneModelErrorLeavesThatSpotUnsureAndCarriesOn(t *testing.T) {
+	boom := errors.New("timeout")
+	adj := &scripted{fn: func(i int, r reconcile.Request) (reconcile.Verdict, error) {
+		if i == 0 {
+			return reconcile.VerdictUnsure, boom
+		}
+		return reconcile.VerdictAlt, nil
+	}}
+
+	out := merge(t, tr("a x s1 s2 y c"), tr("a p s1 s2 q c"), adj)
+
+	assert.Equal(t, "a x s1 s2 q c", out.Text, "the first spot stayed as the base, the second was decided")
+	assert.Equal(t, 1, out.Reconciliation.Errors)
+	assert.Equal(t, 1, out.Reconciliation.Unsure)
+	assert.Equal(t, domain.ReconcileDone, out.Reconciliation.Status)
+}
+
+func TestReconcile_ConsecutiveModelErrorsStopTheWholeReconciliationInsteadOfWaitingOutEveryTimeout(t *testing.T) {
+	boom := errors.New("gateway down")
+	adj := &scripted{fn: func(int, reconcile.Request) (reconcile.Verdict, error) { return reconcile.VerdictUnsure, boom }}
+	a, b := make([]string, 0, 20), make([]string, 0, 20)
+	for i := 0; i < 10; i++ {
+		a = append(a, fmt.Sprintf("w%d", i), "s", "t")
+		b = append(b, fmt.Sprintf("v%d", i), "s", "t")
+	}
+
+	out := merge(t, tr(strings.Join(a, " ")), tr(strings.Join(b, " ")), adj)
+
+	assert.Len(t, adj.calls, 3, "three consecutive failures and it stops asking")
+	assert.Equal(t, strings.Join(a, " "), out.Text, "the base text, with none of the partial changes")
+	assert.Equal(t, domain.ReconcileSkipped, out.Reconciliation.Status)
+	assert.Contains(t, out.Reconciliation.Reason, "gateway down")
+}
+
+func TestReconcile_ASuccessBetweenFailuresResetsTheBreaker(t *testing.T) {
+	boom := errors.New("blip")
+	adj := &scripted{fn: func(i int, r reconcile.Request) (reconcile.Verdict, error) {
+		if i%3 == 2 {
+			return reconcile.VerdictBase, nil
+		}
+		return reconcile.VerdictUnsure, boom
+	}}
+	a, b := make([]string, 0, 20), make([]string, 0, 20)
+	for i := 0; i < 9; i++ {
+		a = append(a, fmt.Sprintf("w%d", i), "s", "t")
+		b = append(b, fmt.Sprintf("v%d", i), "s", "t")
+	}
+
+	out := merge(t, tr(strings.Join(a, " ")), tr(strings.Join(b, " ")), adj)
+
+	assert.Len(t, adj.calls, 9, "two failures then a success never reach three in a row")
+	assert.Equal(t, domain.ReconcileDone, out.Reconciliation.Status)
+	assert.Equal(t, 6, out.Reconciliation.Errors)
+}
+
+func TestReconcile_ACancelledJobIsAnErrorNotASkippedReconciliation(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	adj := &scripted{fn: func(int, reconcile.Request) (reconcile.Verdict, error) {
+		cancel()
+		return reconcile.VerdictUnsure, ctx.Err()
+	}}
+
+	_, err := reconcile.Reconcile(ctx, tr("a x b"), tr("a y b"), adj, opts)
+
+	assert.ErrorIs(t, err, context.Canceled, "a shutdown must not be recorded as 'the model was unavailable'")
+}
+
+func TestReconcile_KeepsTheBaseSegmentsDurationAndNamesTheEngines(t *testing.T) {
+	base := tr("a x b")
+	base.Segments = []domain.Segment{{Start: 0, End: 4, Text: "a x b"}}
+	base.AudioSeconds = 3821.5
+
+	out := merge(t, base, tr("a y b"), always(reconcile.VerdictAlt))
+
+	assert.Equal(t, base.Segments, out.Segments, "the timed segments are the base engine's, untouched")
+	assert.InDelta(t, 3821.5, out.AudioSeconds, 1e-9)
+	assert.Equal(t, "iguana/pianissimo", out.Reconciliation.Base)
+	assert.Equal(t, "iguana/kb-whisper", out.Reconciliation.Alt)
+	assert.Equal(t, "iguana/gemma4-26b", out.Reconciliation.Reconciler)
+}
+
+func TestReconcile_ASpotCarriesTheSecondItStartsAt(t *testing.T) {
+	base := tr("ett två tre. fyra fem sex. sju åtta nio.")
+	base.Segments = []domain.Segment{
+		{Start: 0, End: 10, Text: "ett två tre."},
+		{Start: 10, End: 20, Text: "fyra fem sex."},
+		{Start: 20, End: 30, Text: "sju åtta nio."},
+	}
+	alt := tr("ett två tre. fyra femm sex. sju åtta nio.")
+
+	out := merge(t, base, alt, always(reconcile.VerdictUnsure))
+
+	require.Len(t, out.Reconciliation.Contested, 1)
+	assert.InDelta(t, 10.0, out.Reconciliation.Contested[0].Seconds, 1e-9, "'fem' is in the second segment, which starts at 10 s")
+}
+
+func TestReconcile_WithNoSegmentsASpotIsPlacedProportionally(t *testing.T) {
+	base := tr("a b c d e f g h i j")
+	base.AudioSeconds = 100
+	alt := tr("a b c d e f g h i J2")
+
+	out := merge(t, base, alt, always(reconcile.VerdictUnsure))
+
+	require.Len(t, out.Reconciliation.Contested, 1)
+	assert.InDelta(t, 90.0, out.Reconciliation.Contested[0].Seconds, 1e-9, "token 9 of 10 in 100 s")
+}
+
+func TestReconcile_ContestedSpotsAreCappedButTheCountsAreExact(t *testing.T) {
+	a, b := make([]string, 0, 600), make([]string, 0, 600)
+	for i := 0; i < 300; i++ {
+		a = append(a, fmt.Sprintf("w%d", i), "same", "same2")
+		b = append(b, fmt.Sprintf("v%d", i), "same", "same2")
+	}
+
+	out := merge(t, tr(strings.Join(a, " ")), tr(strings.Join(b, " ")), always(reconcile.VerdictUnsure))
+
+	assert.Equal(t, 300, out.Reconciliation.Unsure)
+	assert.Equal(t, 300, out.Reconciliation.Disputes)
+	assert.Len(t, out.Reconciliation.Contested, reconcile.MaxContested, "the stored list is bounded; the counts are not")
+}
+
+func TestReconcile_AVerdictReturnedAlongsideAnErrorIsIgnored(t *testing.T) {
+	// An adjudicator that returns a verdict AND an error must not have the verdict used:
+	// the contract is that an error means no decision.
+	adj := &scripted{fn: func(int, reconcile.Request) (reconcile.Verdict, error) {
+		return reconcile.VerdictAlt, errors.New("half a reply")
+	}}
+
+	out := merge(t, tr("a x b"), tr("a y b"), adj)
+
+	assert.Equal(t, "a x b", out.Text, "no decision means the base stands")
+	assert.Zero(t, out.Reconciliation.Changed)
+	assert.Equal(t, 1, out.Reconciliation.Unsure)
+}

internal/reconcile/tokens.go

@@ -0,0 +1,57 @@
+// Package reconcile merges two transcripts of the same speech: it finds where they
+// disagree word by word, resolves the trivial disagreements itself, and asks a LOCAL
+// model to choose between the two readings of the rest (ADR-034). It depends on
+// nothing outside the standard library and the domain; the model, the engines and
+// the store are wired in from outside.
+package reconcile
+
+import (
+	"strings"
+	"unicode"
+)
+
+// Token is one word of a transcript. Text is the word as written, with its casing
+// and punctuation, so the merged transcript reads naturally; Key is what two engines
+// are compared on: lower case, letters and digits only.
+type Token struct {
+	Text string
+	Key  string
+}
+
+// Tokenize splits s on whitespace. A "word" that has no letter or digit (a dash, a
+// lone full stop) is not something two engines can disagree about, so it is joined
+// to the neighbouring word instead of becoming a token of its own.
+func Tokenize(s string) []Token {
+	var out []Token
+	pending := ""
+	for _, w := range strings.Fields(s) {
+		k := key(w)
+		if k == "" {
+			if len(out) > 0 {
+				out[len(out)-1].Text += " " + w
+			} else if pending == "" {
+				pending = w
+			} else {
+				pending += " " + w
+			}
+			continue
+		}
+		text := w
+		if pending != "" {
+			text = pending + " " + w
+			pending = ""
+		}
+		out = append(out, Token{Text: text, Key: k})
+	}
+	return out
+}
+
+func key(w string) string {
+	var b strings.Builder
+	for _, r := range w {
+		if unicode.IsLetter(r) || unicode.IsDigit(r) {
+			b.WriteRune(unicode.ToLower(r))
+		}
+	}
+	return b.String()
+}

internal/reconcile/tokens_test.go

@@ -0,0 +1,66 @@
+package reconcile_test
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+
+	"git.d-ma.be/mathias/tapir/internal/reconcile"
+)
+
+func texts(ts []reconcile.Token) []string {
+	out := make([]string, len(ts))
+	for i, t := range ts {
+		out[i] = t.Text
+	}
+	return out
+}
+
+func keys(ts []reconcile.Token) []string {
+	out := make([]string, len(ts))
+	for i, t := range ts {
+		out[i] = t.Key
+	}
+	return out
+}
+
+func TestTokenize_KeepsTheWordAsWrittenAndKeysItWithoutCaseOrPunctuation(t *testing.T) {
+	got := reconcile.Tokenize("Hej, allihopa! Vi går igenom ÅÄÖ.")
+
+	assert.Equal(t, []string{"Hej,", "allihopa!", "Vi", "går", "igenom", "ÅÄÖ."}, texts(got),
+		"the original spelling and punctuation are kept for the merged text")
+	assert.Equal(t, []string{"hej", "allihopa", "vi", "går", "igenom", "åäö"}, keys(got),
+		"the key is lower case with punctuation removed, Swedish letters lowered correctly")
+}
+
+func TestTokenize_KeysIgnoreHyphensAndPlusSoOnlyWordsDecideEquality(t *testing.T) {
+	got := reconcile.Tokenize("end-to-end T+1 5 000 kr")
+
+	assert.Equal(t, []string{"endtoend", "t1", "5", "000", "kr"}, keys(got))
+}
+
+func TestTokenize_APunctuationOnlyWordJoinsThePreviousWordInsteadOfBecomingOne(t *testing.T) {
+	got := reconcile.Tokenize("Vi går – sen tillbaka . Slut")
+
+	assert.Equal(t, []string{"Vi", "går –", "sen", "tillbaka .", "Slut"}, texts(got),
+		"a dash or a lone full stop is not a word the engines can disagree about")
+	assert.Equal(t, []string{"vi", "går", "sen", "tillbaka", "slut"}, keys(got))
+}
+
+func TestTokenize_LeadingPunctuationOnlyWordJoinsTheNextWord(t *testing.T) {
+	got := reconcile.Tokenize("— Hej då")
+
+	assert.Equal(t, []string{"— Hej", "då"}, texts(got))
+}
+
+func TestTokenize_NewlinesAndRunsOfSpaceAreJustSeparators(t *testing.T) {
+	got := reconcile.Tokenize("  Hej\n  allihopa \t Vi ")
+
+	assert.Equal(t, []string{"Hej", "allihopa", "Vi"}, texts(got))
+}
+
+func TestTokenize_EmptyAndPunctuationOnlyInputGiveNoTokens(t *testing.T) {
+	assert.Empty(t, reconcile.Tokenize(""))
+	assert.Empty(t, reconcile.Tokenize("   \n "))
+	assert.Empty(t, reconcile.Tokenize("— . !"))
+}