Simplify references and output instructions in LLM prompts

cmd/audita/main_integration_test.go

@@ -143,12 +143,12 @@ func (p *fixtureProvider) serve(w http.ResponseWriter, r *http.Request) {
 		}
 		if validation {
 			if p.mode == "span-clean" || p.mode == "span-filter" {
-				content = `{"validations":[{"correction_index":0,"approved":true,"confidence":0.99,"reason":"ok"},{"correction_index":1,"approved":true,"confidence":0.99,"reason":"ok"}]}`
+				content = `{"validations":[{"correction_index":1,"approved":true,"confidence":0.99,"reason":"ok"},{"correction_index":2,"approved":true,"confidence":0.99,"reason":"ok"}]}`
 			} else {
-				content = `{"validations":[{"correction_index":0,"approved":true,"confidence":0.99,"reason":"ok"}]}`
+				content = `{"validations":[{"correction_index":1,"approved":true,"confidence":0.99,"reason":"ok"}]}`
 			}
 			if p.mode == "span-negative" && validationCall == 2 {
-				content = `{"validations":[{"correction_index":0,"approved":false,"confidence":0.99,"reason":"would reverse meaning"}]}`
+				content = `{"validations":[{"correction_index":1,"approved":false,"confidence":0.99,"reason":"would reverse meaning"}]}`
 			}
 		}
 	}
@@ -159,10 +159,8 @@ func (p *fixtureProvider) serve(w http.ResponseWriter, r *http.Request) {
 		}
 	}
 	if validation && !strings.HasPrefix(p.mode, "span-") {
-		content = `{"validations":[{"correction_index":0,"approved":true,"confidence":0.99,"reason":"ok"}]}`
-		if p.mode == "repair-filter-later-failure" {
 		content = `{"validations":[{"correction_index":1,"approved":true,"confidence":0.99,"reason":"ok"}]}`
-		} else if p.mode == "invalid-decision-after-repair" {
+		if p.mode == "invalid-decision-after-repair" {
 			if validationCall == 1 {
 				content = `{"wrong":true}`
 			} else {
@@ -173,12 +171,12 @@ func (p *fixtureProvider) serve(w http.ResponseWriter, r *http.Request) {
 	if validation && strings.HasPrefix(p.mode, "decision-") {
 		decisions := make([]map[string]any, 0, len(requestedIndices)+2)
 		for _, index := range requestedIndices {
-			if p.mode == "decision-mixed" && len(requestedIndices) == 4 && index == 2 {
+			if p.mode == "decision-mixed" && len(requestedIndices) == 4 && index == 3 {
 				continue
 			}
-			approved := !(p.mode == "decision-mixed" && len(requestedIndices) == 4 && index == 1)
+			approved := !(p.mode == "decision-mixed" && len(requestedIndices) == 4 && index == 2)
 			decisions = append(decisions, map[string]any{"correction_index": index, "approved": approved, "confidence": 0.99, "reason": "controlled review"})
-			if p.mode == "decision-mixed" && len(requestedIndices) == 4 && index == 3 {
+			if p.mode == "decision-mixed" && len(requestedIndices) == 4 && index == 4 {
 				decisions = append(decisions, map[string]any{"correction_index": index, "approved": false, "confidence": 0.99, "reason": "conflicting review"})
 			}
 		}
@@ -606,9 +604,9 @@ func TestSubprocessKeepsUsableValidatorDecisionsInMixedResponse(t *testing.T) {
 	full, survivor := 0, 0
 	for _, indices := range requests {
 		switch {
-		case reflect.DeepEqual(indices, []int{0, 1, 2, 3}):
+		case reflect.DeepEqual(indices, []int{1, 2, 3, 4}):
 			full++
-		case reflect.DeepEqual(indices, []int{0}):
+		case reflect.DeepEqual(indices, []int{1}):
 			survivor++
 		default:
 			t.Fatalf("unexpected validation identities: %v", requests)
@@ -662,7 +660,7 @@ func TestSubprocessReportsFilteredExtraValidatorDecision(t *testing.T) {
 	}
 	requests := f.assertDecisionRequestsUseOneAttempt(t)
 	for _, indices := range requests {
-		if !reflect.DeepEqual(indices, []int{0}) {
+		if !reflect.DeepEqual(indices, []int{1}) {
 			t.Fatalf("unexpected identities: %v", requests)
 		}
 	}

docs/internal/modules.md

@@ -42,7 +42,13 @@ transcript section, and trimmed description once for all four modules. PromptKit
 renders their embedded definitions from those variables and validates the
 correction-set schema before Audita maps proposals. Domain validators still
 check confidence, applicability, protected glossary terms, and replacement
-policy. Shared context encoding lives in `internal/framework/promptcontext`.
+policy. Model-facing segment IDs are one-based positions within the proposal
+request. Generation maps returned references back to the current chunk's actual
+segment IDs, preserving response order and multiple proposals per segment.
+Unknown references become invalid-target candidates for the existing domain
+checks; they are never interpreted as actual transcript IDs. Debug request
+metadata records `segment_ids` in local-reference order (reference 1 is the
+first entry). Shared context encoding lives in `internal/framework/promptcontext`.
 An empty `corrected_text` deletes the identified span when the resulting segment
 remains nonblank. Whitespace-only original spans can express spacing corrections;
 matching stays literal under each module's replacement policy. No automatic

docs/internal/prompts.md

@@ -42,9 +42,14 @@ pass these values through the role-bound adapter. See the
 
 Proposal instructions describe literal matching, each module's replacement policy,
 source-based proposal identity, and confidence in whether an edit is warranted.
-Validator instructions distinguish correction indices from segment IDs and confidence
+Proposal `id` and validator `correction_index` values are short, one-based
-in the verdict from the verdict itself. The grammar article exception and glossary
+references scoped to each request, not internal transcript/proposal identities.
-mishearing exception apply consistently across their instructions.
+Mapping belongs to [proposal generation](modules.md) and
+[validator reconciliation](validators.md). Validator instructions distinguish
+confidence in the verdict from the verdict itself. Compact output examples
+illustrate the supplied schema without requiring fixed JSON formatting or key
+order. The grammar article exception and glossary mishearing exception apply
+consistently across their instructions.
 
 Only the three active validator definitions are retained. The unused internal
 `grammar_review` and `spoken_word_review` definitions and type selections were

docs/internal/validators.md

@@ -63,12 +63,14 @@ Oversized single inputs are rejected with a consolidated
 model failures make only their batch candidates unavailable while other
 batches continue. PromptKit validates each complete response against the
 validator-decision JSON Schema and performs configured structural repair before
-Audita receives decoded decisions. Audita then reconciles those decisions with
+Audita receives decoded decisions. Each actual batch presents candidates
-the correction indices requested in that batch: one usable entry keeps its
+numbered from 1, independent of their internal proposal indices. Audita then
+reconciles those local references: one usable entry keeps its
 approval or editorial rejection; a missing or duplicate requested entry makes
 only that candidate unavailable; unexpected indices are ignored. An empty list
-makes all requested candidates unavailable. Results are ordered by request
+makes all requested candidates unavailable. Decisions and unavailable indices
-identity and satisfy strict internal cardinality before the runner advances
+map back to internal proposal identities; results are ordered by request identity
+and satisfy strict internal cardinality before the runner advances
 approved candidates. An unavailable candidate never advances as approved.
 Invalid confidence in a decoded test double also makes only its candidate
 unavailable; production schema validation still rejects out-of-range confidence
@@ -79,6 +81,12 @@ chain without returning unchecked candidates as approved. Reconciliation makes
 no semantic retry or extra model call; guided recovery remains
 [deferred](../roadmap/future.md#domain-aware-response-recovery).
 
+Debug request metadata records `proposal_indices` in local-reference order
+(reference 1 is the first entry). Raw decoded responses and ignored indices use
+local references; reconciled decisions and unavailable indices use internal
+proposal indices. Batch token estimates use the same locally numbered payload
+as prompt encoding.
+
 ## Decision and Rejection Reporting
 
 Runner collects decisions, rejected candidates, and operation outcomes. Ledger
@@ -111,9 +119,11 @@ is covered by [report assembly](diagnostics-reporting.md).
 ## Supplemental Validation Context
 
 Each candidate includes the immediately preceding and following source segments
-available within its fixed chunk, plus glossary entries whose names, aliases, or
+available within its fixed chunk, labeled `preceding` or `following` with speaker,
+text, and categories, plus glossary entries whose names, aliases, or
 recognized plurals appear in its original/resulting text or those neighbors.
 Glossary matching reuses the protected vocabulary; entries retain glossary order.
+Model-facing candidate and neighboring context omit internal segment IDs.
 This supplies evidence such as a troupe definition for a proposed `Jesters`
 correction or an adjacent speaker turn explaining deliberate repetition. A correction
 already clear from its own segment does not require such evidence. No audio,

internal/framework/promptcontext/transcript_section.go

@@ -21,7 +21,8 @@ type transcriptSectionPayload struct {
 }
 
 // MarshalTranscriptSectionJSON builds the standardized transcript-section JSON
-// payload consumed by proposal prompt templates.
+// payload consumed by proposal prompt templates. IDs are one-based positions in
+// this request; the transcript and its actual IDs are unchanged.
 func MarshalTranscriptSectionJSON(transcript *schema.Transcript, sectionIndex int) ([]byte, error) {
 	payload := transcriptSectionPayload{
 		SectionIndex: sectionIndex,
@@ -29,9 +30,9 @@ func MarshalTranscriptSectionJSON(transcript *schema.Transcript, sectionIndex in
 	}
 
 	if transcript != nil {
-		for _, s := range transcript.Segments {
+		for i, s := range transcript.Segments {
 			payload.Segments = append(payload.Segments, transcriptSectionSegment{
-				ID:         s.ID,
+				ID:         i + 1,
 				Speaker:    s.Speaker,
 				Start:      s.Start,
 				End:        s.End,

internal/framework/promptcontext/transcript_section_test.go

@@ -10,7 +10,7 @@ import (
 
 func TestMarshalTranscriptSectionJSONShape(t *testing.T) {
 	transcript := &schema.Transcript{Segments: []schema.Segment{
-		{ID: 1, Speaker: "A", Start: 0.1, End: 1.2, Text: "alpha", Categories: []string{"session", "intro"}},
+		{ID: 817, Speaker: "A", Start: 0.1, End: 1.2, Text: "alpha", Categories: []string{"session", "intro"}},
 	}}
 
 	raw, err := MarshalTranscriptSectionJSON(transcript, 3)

internal/framework/proposal_generation/adapter_integration_test.go

@@ -61,7 +61,7 @@ func TestRepairedProposalStillFacesDomainChecks(t *testing.T) {
 	if err != nil {
 		t.Fatal(err)
 	}
-	generated, err := GenerateCandidates(context.Background(), Request{ModuleKey: "grammar", ModuleInstance: "grammar", ReplacementPolicy: proposals.ReplacementPolicyRequireUnique, Config: &loaded.Config, PromptID: prompts.PromptIDModuleGrammarProposal, PromptVersion: prompts.VersionV1, Variables: vars, LLMClient: client})
+	generated, err := GenerateCandidates(context.Background(), Request{WorkingTranscript: transcript, ModuleKey: "grammar", ModuleInstance: "grammar", ReplacementPolicy: proposals.ReplacementPolicyRequireUnique, Config: &loaded.Config, PromptID: prompts.PromptIDModuleGrammarProposal, PromptVersion: prompts.VersionV1, Variables: vars, LLMClient: client})
 	if err != nil {
 		t.Fatal(err)
 	}

internal/framework/proposal_generation/generate.go

@@ -36,7 +36,7 @@ type InteractionArtifacts struct {
 
 // StructuredCorrectionProposal is one LLM response correction payload.
 type StructuredCorrectionProposal struct {
-	TargetSegmentID int     `json:"id"`
+	SegmentReference int     `json:"id"`
 	OriginalText     string  `json:"original_text"`
 	CorrectedText    string  `json:"corrected_text"`
 	Confidence       float64 `json:"confidence"`
@@ -142,6 +142,7 @@ func GenerateCandidates(ctx context.Context, req Request) (Result, error) {
 		"replacement_policy": req.ReplacementPolicy,
 		"section":            req.Section,
 		"start_index":        req.StartIndex,
+		"segment_ids":        segmentIDs(req.WorkingTranscript),
 		"prompt":             completion.Prompt,
 	}
 	requestMetadata["response_schema"] = responseschema.MustLookup(responseschema.CorrectionSetKey).DiagnosticsMap()
@@ -185,8 +186,13 @@ func GenerateCandidates(ctx context.Context, req Request) (Result, error) {
 	corrections := make([]proposals.CorrectionProposal, 0, len(response.Corrections))
 	enriched := make([]proposals.EnrichedCorrectionProposal, 0, len(response.Corrections))
 	for i, raw := range response.Corrections {
+		// Unknown references remain invalid candidates for the existing domain validators.
+		targetID := 0
+		if req.WorkingTranscript != nil && raw.SegmentReference > 0 && raw.SegmentReference <= len(req.WorkingTranscript.Segments) {
+			targetID = req.WorkingTranscript.Segments[raw.SegmentReference-1].ID
+		}
 		candidate := proposals.CorrectionProposal{
-			TargetSegmentID: raw.TargetSegmentID,
+			TargetSegmentID: targetID,
 			OriginalText:    raw.OriginalText,
 			CorrectedText:   raw.CorrectedText,
 			Confidence:      raw.Confidence,
@@ -247,3 +253,15 @@ func (a diagnosticsWriterAdapter) WriteInteraction(stage string, requestMetadata
 		ErrorPayloadPath:    artifacts.ErrorPayloadPath,
 	}, nil
 }
+
+// segmentIDs records the request-local reference mapping in debug metadata.
+func segmentIDs(transcript *schema.Transcript) []int {
+	if transcript == nil {
+		return nil
+	}
+	ids := make([]int, len(transcript.Segments))
+	for i, segment := range transcript.Segments {
+		ids[i] = segment.ID
+	}
+	return ids
+}

internal/framework/proposal_generation/generate_test.go

@@ -12,6 +12,7 @@ import (
 
 	ck "gitea.maximumdirect.net/eric/ConfigKit"
 	"gitea.maximumdirect.net/eric/audita/internal/core/config"
+	"gitea.maximumdirect.net/eric/audita/internal/core/schema"
 	"gitea.maximumdirect.net/eric/audita/internal/framework/contracts"
 	"gitea.maximumdirect.net/eric/audita/internal/framework/outcomes"
 	"gitea.maximumdirect.net/eric/audita/internal/framework/proposals"
@@ -77,6 +78,7 @@ func defaultRequest(t *testing.T) Request {
 		PromptID:          "modules.grammar.proposal",
 		PromptVersion:     "v1",
 		Variables:         map[string]string{"SectionJSON": "[]"},
+		WorkingTranscript: &schema.Transcript{Segments: []schema.Segment{{ID: 7, Text: "teh"}, {ID: 25, Text: "other"}}},
 		StartIndex:        0,
 	}
 }
@@ -86,7 +88,7 @@ func TestGenerateCandidatesSuccess(t *testing.T) {
 		responses: []StructuredCorrectionSet{
 			{
 				Corrections: []StructuredCorrectionProposal{
-					{TargetSegmentID: 7, OriginalText: "teh", CorrectedText: "the", Confidence: 0.9},
+					{SegmentReference: 1, OriginalText: "teh", CorrectedText: "the", Confidence: 0.9},
 				},
 			},
 		},
@@ -104,6 +106,9 @@ func TestGenerateCandidatesSuccess(t *testing.T) {
 	if len(got.Corrections) != 1 || len(got.Enriched) != 1 {
 		t.Fatalf("unexpected proposal lengths: %+v", got)
 	}
+	if got.Corrections[0].TargetSegmentID != 7 {
+		t.Fatalf("local reference not resolved: %+v", got.Corrections)
+	}
 	if got.Enriched[0].ProposalIndex != 10 {
 		t.Fatalf("expected proposal index 10, got %d", got.Enriched[0].ProposalIndex)
 	}
@@ -118,7 +123,7 @@ func TestGenerateCandidatesSuccess(t *testing.T) {
 func TestGenerateCandidatesUsesCorrectionSetSchema(t *testing.T) {
 	client := &fakeStructuredClient{
 		responses: []StructuredCorrectionSet{
-			{Corrections: []StructuredCorrectionProposal{{TargetSegmentID: 1, OriginalText: "teh", CorrectedText: "the", Confidence: 0.9}}},
+			{Corrections: []StructuredCorrectionProposal{{SegmentReference: 1, OriginalText: "teh", CorrectedText: "the", Confidence: 0.9}}},
 		},
 	}
 	req := defaultRequest(t)
@@ -140,7 +145,7 @@ func TestGenerateCandidatesUsesCorrectionSetSchema(t *testing.T) {
 func TestGenerateCandidatesDiagnosticsIncludeSchemaMetadata(t *testing.T) {
 	client := &fakeStructuredClient{
 		responses: []StructuredCorrectionSet{
-			{Corrections: []StructuredCorrectionProposal{{TargetSegmentID: 1, OriginalText: "teh", CorrectedText: "the", Confidence: 0.9}}},
+			{Corrections: []StructuredCorrectionProposal{{SegmentReference: 1, OriginalText: "teh", CorrectedText: "the", Confidence: 0.9}}},
 		},
 	}
 	diag := &captureDiagnosticsWriter{}
@@ -172,7 +177,7 @@ func TestGenerateCandidatesInvalidCorrectionIsPreservedForLaterValidation(t *tes
 		responses: []StructuredCorrectionSet{
 			{
 				Corrections: []StructuredCorrectionProposal{
-					{TargetSegmentID: 0, OriginalText: "x", CorrectedText: "", Confidence: 1.2},
+					{SegmentReference: 0, OriginalText: "x", CorrectedText: "", Confidence: 1.2},
 				},
 			},
 		},
@@ -235,8 +240,8 @@ func TestGenerateCandidatesValidEmptyResponseIsSuccessful(t *testing.T) {
 func TestGenerateCandidatesDeterministicIndexAssignment(t *testing.T) {
 	baseResponse := StructuredCorrectionSet{
 		Corrections: []StructuredCorrectionProposal{
-			{TargetSegmentID: 1, OriginalText: "a", CorrectedText: "A", Confidence: 0.9},
+			{SegmentReference: 1, OriginalText: "a", CorrectedText: "A", Confidence: 0.9},
-			{TargetSegmentID: 2, OriginalText: "b", CorrectedText: "B", Confidence: 0.9},
+			{SegmentReference: 2, OriginalText: "b", CorrectedText: "B", Confidence: 0.9},
 		},
 	}
 	clientA := &fakeStructuredClient{responses: []StructuredCorrectionSet{baseResponse}}
@@ -266,8 +271,8 @@ func TestGenerateCandidatesDeterministicIndexAssignment(t *testing.T) {
 func TestGenerateCandidatesMultipleSectionsStableMetadata(t *testing.T) {
 	client := &fakeStructuredClient{
 		responses: []StructuredCorrectionSet{
-			{Corrections: []StructuredCorrectionProposal{{TargetSegmentID: 1, OriginalText: "alpha", CorrectedText: "ALPHA", Confidence: 0.9}}},
+			{Corrections: []StructuredCorrectionProposal{{SegmentReference: 1, OriginalText: "alpha", CorrectedText: "ALPHA", Confidence: 0.9}}},
-			{Corrections: []StructuredCorrectionProposal{{TargetSegmentID: 3, OriginalText: "charlie", CorrectedText: "CHARLIE", Confidence: 0.9}}},
+			{Corrections: []StructuredCorrectionProposal{{SegmentReference: 1, OriginalText: "charlie", CorrectedText: "CHARLIE", Confidence: 0.9}}},
 		},
 	}
 	section0 := contracts.SectionMetadata{Index: 0}
@@ -276,6 +281,7 @@ func TestGenerateCandidatesMultipleSectionsStableMetadata(t *testing.T) {
 	req0 := defaultRequest(t)
 	req0.LLMClient = client
 	req0.Section = &section0
+	req0.WorkingTranscript = &schema.Transcript{Segments: []schema.Segment{{ID: 71, Text: "alpha"}}}
 	req0.StartIndex = 0
 	part0, err := GenerateCandidates(context.Background(), req0)
 	if err != nil {
@@ -285,6 +291,7 @@ func TestGenerateCandidatesMultipleSectionsStableMetadata(t *testing.T) {
 	req1 := defaultRequest(t)
 	req1.LLMClient = client
 	req1.Section = &section1
+	req1.WorkingTranscript = &schema.Transcript{Segments: []schema.Segment{{ID: 903, Text: "charlie"}}}
 	req1.StartIndex = len(part0.Enriched)
 	part1, err := GenerateCandidates(context.Background(), req1)
 	if err != nil {
@@ -295,6 +302,9 @@ func TestGenerateCandidatesMultipleSectionsStableMetadata(t *testing.T) {
 	if len(all) != 2 {
 		t.Fatalf("expected 2 proposals, got %d", len(all))
 	}
+	if all[0].TargetSegmentID != 71 || all[1].TargetSegmentID != 903 {
+		t.Fatalf("section references did not reset: %+v", all)
+	}
 	if all[0].ProposalIndex != 0 || all[1].ProposalIndex != 1 {
 		t.Fatalf("unexpected proposal indexes: %d, %d", all[0].ProposalIndex, all[1].ProposalIndex)
 	}
@@ -308,7 +318,7 @@ func TestGenerateCandidatesDiagnosticsWrittenAndRedacted(t *testing.T) {
 	validationSecret := "proposal-validation-secret"
 	client := &fakeStructuredClient{
 		responses: []StructuredCorrectionSet{
-			{Corrections: []StructuredCorrectionProposal{{TargetSegmentID: 1, OriginalText: validationSecret, CorrectedText: "safe", Confidence: 0.9}}},
+			{Corrections: []StructuredCorrectionProposal{{SegmentReference: 1, OriginalText: validationSecret, CorrectedText: "safe", Confidence: 0.9}}},
 		},
 	}
 	t.Setenv("PRIMARY_SECRET", primarySecret)

internal/framework/proposal_generation/module_proposal_test.go

@@ -2,6 +2,7 @@ package proposal_generation
 
 import (
 	"context"
+	"encoding/json"
 	"errors"
 	"testing"
 
@@ -17,7 +18,7 @@ import (
 func TestExecuteModuleProposalBuildsMessagesFromSectionAndDescription(t *testing.T) {
 	client := &fakeStructuredClient{
 		responses: []StructuredCorrectionSet{
-			{Corrections: []StructuredCorrectionProposal{{TargetSegmentID: 1, OriginalText: "teh", CorrectedText: "the", Confidence: 0.9}}},
+			{Corrections: []StructuredCorrectionProposal{{SegmentReference: 1, OriginalText: "teh", CorrectedText: "the", Confidence: 0.9}}},
 		},
 	}
 	section := contracts.SectionMetadata{Index: 7}
@@ -135,3 +136,47 @@ func TestExecuteModuleProposalPreservesFatalOperation(t *testing.T) {
 		t.Fatalf("fatal operation lost: %+v, %v", result, err)
 	}
 }
+
+func TestModuleProposalMapsLocalReferencesWithoutGuessing(t *testing.T) {
+	transcript := &schema.Transcript{Segments: []schema.Segment{{ID: 71, Text: "teh cat"}, {ID: 903, Text: "uh hello"}}}
+	client := &fakeStructuredClient{responses: []StructuredCorrectionSet{{Corrections: []StructuredCorrectionProposal{
+		{SegmentReference: 2, OriginalText: "uh ", CorrectedText: "", Confidence: .9},
+		{SegmentReference: 1, OriginalText: "teh", CorrectedText: "the", Confidence: .9},
+		{SegmentReference: 2, OriginalText: "hello", CorrectedText: "Hello", Confidence: .9},
+		{SegmentReference: 71, OriginalText: "teh", CorrectedText: "the", Confidence: .9},
+	}}}}
+	result, err := ExecuteModuleProposal(context.Background(), ModuleProposalRequest{
+		ProposalRequest: contracts.ProposalRequest{
+			ExecutionContext: contracts.ExecutionContext{WorkingTranscript: transcript},
+			RunSpec:          contracts.ModuleRunSpec{ModuleKey: "grammar", InstanceName: "grammar", ReplacementPolicy: proposals.ReplacementPolicyRequireUnique},
+			LLMClient:        client,
+		},
+		PromptID: prompts.PromptIDModuleGrammarProposal, BuildVariables: BuildProposalVariables,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	var payload struct {
+		Segments []struct {
+			ID   int    `json:"id"`
+			Text string `json:"text"`
+		} `json:"segments"`
+	}
+	if err := json.Unmarshal([]byte(client.calls[0].Variables["SectionJSON"]), &payload); err != nil {
+		t.Fatal(err)
+	}
+	if len(payload.Segments) != 2 || payload.Segments[0].ID != 1 || payload.Segments[1].ID != 2 || payload.Segments[1].Text != "uh hello" {
+		t.Fatalf("request references: %+v", payload)
+	}
+	if len(result.Proposals) != 4 {
+		t.Fatalf("proposals: %+v", result)
+	}
+	for i, id := range []int{903, 71, 903, 0} {
+		if result.Proposals[i].TargetSegmentID != id {
+			t.Fatalf("mapped proposal %d: %+v", i, result.Proposals[i])
+		}
+	}
+	if result.Proposals[0].CorrectedText != "" || transcript.Segments[0].ID != 71 || transcript.Segments[1].ID != 903 {
+		t.Fatal("mapping changed deletion or source identities")
+	}
+}

internal/framework/runner/runner_test.go

@@ -621,7 +621,7 @@ func (c *captureContractStructuredClient) CompleteStructured(ctx context.Context
 	if target, ok := out.(*validators.LLMValidationResponse); ok {
 		*target = validators.LLMValidationResponse{
 			Validations: []validators.LLMValidationDecision{
-				{CorrectionIndex: 0, Approved: true, Confidence: 0.9, Reason: "ok"},
+				{CorrectionIndex: 1, Approved: true, Confidence: 0.9, Reason: "ok"},
 			},
 		}
 	}
@@ -653,7 +653,7 @@ func (c *stageAwareStructuredClient) CompleteStructured(ctx context.Context, req
 	target := out.(*validators.LLMValidationResponse)
 	*target = validators.LLMValidationResponse{
 		Validations: []validators.LLMValidationDecision{
-			{CorrectionIndex: section, Approved: true, Confidence: 0.99, Reason: "ok"},
+			{CorrectionIndex: 1, Approved: true, Confidence: 0.99, Reason: "ok"},
 		},
 	}
 	return contracts.StructuredCompletionResponse{}, nil
@@ -942,7 +942,7 @@ func (f *fakeStructuredClient) CompleteStructured(ctx context.Context, req contr
 }
 
 func TestRunnerLLMValidatorApprovalApplied(t *testing.T) {
-	client := &fakeStructuredClient{responses: []validators.LLMValidationResponse{{Validations: []validators.LLMValidationDecision{{CorrectionIndex: 0, Approved: true, Confidence: 0.9, Reason: "ok"}}}}}
+	client := &fakeStructuredClient{responses: []validators.LLMValidationResponse{{Validations: []validators.LLMValidationDecision{{CorrectionIndex: 1, Approved: true, Confidence: 0.9, Reason: "ok"}}}}}
 	llmValidator, err := validators.NewLLMBackedValidator("spoken_form_plausibility_review", validators.LLMValidatorTypeSpokenFormPlausibility)
 	if err != nil {
 		t.Fatalf("NewLLMBackedValidator: %v", err)
@@ -969,7 +969,7 @@ func TestRunnerLLMValidatorApprovalApplied(t *testing.T) {
 }
 
 func TestRunnerLLMValidatorRejectionPreventsApplication(t *testing.T) {
-	client := &fakeStructuredClient{responses: []validators.LLMValidationResponse{{Validations: []validators.LLMValidationDecision{{CorrectionIndex: 0, Approved: false, Confidence: 0.95, Reason: "reject"}}}}}
+	client := &fakeStructuredClient{responses: []validators.LLMValidationResponse{{Validations: []validators.LLMValidationDecision{{CorrectionIndex: 1, Approved: false, Confidence: 0.95, Reason: "reject"}}}}}
 	llmValidator, _ := validators.NewLLMBackedValidator("spoken_form_plausibility_review", validators.LLMValidatorTypeSpokenFormPlausibility)
 	r := New(fakeFactory{modules: map[string]contracts.TranscriptModule{
 		"m": fakeModule{key: "m", policy: proposals.ReplacementPolicyRequireUnique, validators: []contracts.Validator{llmValidator}, proposeF: func(req contracts.ProposalRequest) ([]proposals.CorrectionProposal, error) {
@@ -1028,8 +1028,8 @@ func TestRunnerLLMValidatorMalformedResponseRejectsBatchAndKeepsPartialProgress(
 
 func TestRunnerReconciledSurvivorsCompleteLaterValidation(t *testing.T) {
 	client := &fakeStructuredClient{responses: []validators.LLMValidationResponse{{Validations: []validators.LLMValidationDecision{
-		{CorrectionIndex: 0, Approved: true, Confidence: .9, Reason: "ok"},
+		{CorrectionIndex: 1, Approved: true, Confidence: .9, Reason: "ok"},
-		{CorrectionIndex: 2, Approved: true, Confidence: .9, Reason: "ok"},
+		{CorrectionIndex: 3, Approved: true, Confidence: .9, Reason: "ok"},
 		{CorrectionIndex: 99, Approved: true, Confidence: .9, Reason: "unexpected"},
 	}}}}
 	first, _ := validators.NewLLMBackedValidator("editorial_review", validators.LLMValidatorTypeEditorialReview)
@@ -1095,8 +1095,8 @@ func TestRunnerLLMValidatorMissingDecisionRejectsBatch(t *testing.T) {
 
 func TestRunnerLLMValidatorDuplicateDecisionRejectsBatch(t *testing.T) {
 	client := &fakeStructuredClient{responses: []validators.LLMValidationResponse{{Validations: []validators.LLMValidationDecision{
-		{CorrectionIndex: 0, Approved: true, Confidence: 0.9, Reason: "ok"},
+		{CorrectionIndex: 1, Approved: true, Confidence: 0.9, Reason: "ok"},
-		{CorrectionIndex: 0, Approved: false, Confidence: 0.9, Reason: "dup"},
+		{CorrectionIndex: 1, Approved: false, Confidence: 0.9, Reason: "dup"},
 	}}}}
 	llmValidator, _ := validators.NewLLMBackedValidator("spoken_form_plausibility_review", validators.LLMValidatorTypeSpokenFormPlausibility)
 	r := New(fakeFactory{modules: map[string]contracts.TranscriptModule{
@@ -1122,7 +1122,7 @@ func (lenEstimator) EstimateTokens(text string) int { return len(text) }
 
 func TestRunnerLLMValidatorBatching(t *testing.T) {
 	client := &fakeStructuredClient{responses: []validators.LLMValidationResponse{
-		{Validations: []validators.LLMValidationDecision{{CorrectionIndex: 0, Approved: true, Confidence: 0.9, Reason: "ok"}}},
+		{Validations: []validators.LLMValidationDecision{{CorrectionIndex: 1, Approved: true, Confidence: 0.9, Reason: "ok"}}},
 		{Validations: []validators.LLMValidationDecision{{CorrectionIndex: 1, Approved: true, Confidence: 0.9, Reason: "ok"}}},
 	}}
 	llmValidator, _ := validators.NewLLMBackedValidator("spoken_form_plausibility_review", validators.LLMValidatorTypeSpokenFormPlausibility)
@@ -1159,7 +1159,7 @@ func TestRunnerLLMValidatorBatching(t *testing.T) {
 func TestRunnerLLMValidatorDiagnosticsWrittenAndRedacted(t *testing.T) {
 	primarySecret := "runner-primary-secret"
 	validationSecret := "runner-validation-secret"
-	client := &fakeStructuredClient{responses: []validators.LLMValidationResponse{{Validations: []validators.LLMValidationDecision{{CorrectionIndex: 0, Approved: true, Confidence: 0.9, Reason: validationSecret}}}}}
+	client := &fakeStructuredClient{responses: []validators.LLMValidationResponse{{Validations: []validators.LLMValidationDecision{{CorrectionIndex: 1, Approved: true, Confidence: 0.9, Reason: validationSecret}}}}}
 	llmValidator, _ := validators.NewLLMBackedValidator("spoken_form_plausibility_review", validators.LLMValidatorTypeSpokenFormPlausibility)
 	t.Setenv("RUNNER_PRIMARY_KEY", primarySecret)
 	t.Setenv("RUNNER_VALIDATION_KEY", validationSecret)
@@ -1265,7 +1265,7 @@ func TestRunnerProposalGenerationHelperFlowsThroughPipeline(t *testing.T) {
 		responses: []proposal_generation.StructuredCorrectionSet{
 			{
 				Corrections: []proposal_generation.StructuredCorrectionProposal{
-					{TargetSegmentID: 1, OriginalText: "teh", CorrectedText: "the", Confidence: 0.9},
+					{SegmentReference: 1, OriginalText: "teh", CorrectedText: "the", Confidence: 0.9},
 				},
 			},
 		},
@@ -1308,7 +1308,7 @@ func TestRunnerGrammarModuleUsesConfidenceThreshold(t *testing.T) {
 		responses: []proposal_generation.StructuredCorrectionSet{
 			{
 				Corrections: []proposal_generation.StructuredCorrectionProposal{
-					{TargetSegmentID: 1, OriginalText: "hello ,world", CorrectedText: "Hello, world", Confidence: 0.5},
+					{SegmentReference: 1, OriginalText: "hello ,world", CorrectedText: "Hello, world", Confidence: 0.5},
 				},
 			},
 		},
@@ -1342,7 +1342,7 @@ func TestRunnerGlossaryModuleUsesConfidenceThreshold(t *testing.T) {
 		responses: []proposal_generation.StructuredCorrectionSet{
 			{
 				Corrections: []proposal_generation.StructuredCorrectionProposal{
-					{TargetSegmentID: 1, OriginalText: "gestures", CorrectedText: "Jesters", Confidence: 0.5},
+					{SegmentReference: 1, OriginalText: "gestures", CorrectedText: "Jesters", Confidence: 0.5},
 				},
 			},
 		},

internal/framework/validators/llm_batching.go

@@ -70,7 +70,7 @@ func ChunkLLMValidationItems(items []LLMValidationItem, maxPromptTokens int, est
 }
 
 func estimateBatchTokens(estimator chunking.TokenEstimator, batch []LLMValidationItem) (int, error) {
-	payload, err := json.Marshal(batch)
+	payload, err := json.Marshal(localValidationItems(batch))
 	if err != nil {
 		return 0, fmt.Errorf("marshal batch payload: %w", err)
 	}

internal/framework/validators/llm_context.go

@@ -8,8 +8,8 @@ import (
 // validationContext supplies adjacent source segments within the current chunk
 // and glossary definitions mentioned in the candidate or that local context.
 // These fields travel in each item so batching accounts for their size.
-func validationContext(req Request, segmentID int, corrected string) ([]schema.Segment, []schema.GlossaryEntry) {
+func validationContext(req Request, segmentID int, corrected string) ([]NeighboringSegment, []schema.GlossaryEntry) {
-	var neighbors []schema.Segment
+	var neighbors []NeighboringSegment
 	var source string
 	if req.WorkingTranscript != nil {
 		segments := req.WorkingTranscript.Segments
@@ -19,10 +19,10 @@ func validationContext(req Request, segmentID int, corrected string) ([]schema.S
 			}
 			source = segment.Text
 			if i > 0 {
-				neighbors = append(neighbors, segments[i-1])
+				neighbors = append(neighbors, neighboringSegment("preceding", segments[i-1]))
 			}
 			if i+1 < len(segments) {
-				neighbors = append(neighbors, segments[i+1])
+				neighbors = append(neighbors, neighboringSegment("following", segments[i+1]))
 			}
 			break
 		}
@@ -66,3 +66,7 @@ func fitValidationContext(item LLMValidationItem, budget int, estimator chunking
 		}
 	}
 }
+
+func neighboringSegment(position string, segment schema.Segment) NeighboringSegment {
+	return NeighboringSegment{Position: position, Speaker: segment.Speaker, Text: segment.Text, Categories: segment.Categories}
+}

internal/framework/validators/llm_context_test.go

@@ -29,7 +29,7 @@ func TestValidationContextCarriesAdjacentEvidenceAndRelevantDefinitions(t *testi
 		t.Fatalf("payload: %+v immediate: %+v", payload, immediate)
 	}
 	item := payload.Items[0]
-	if len(item.NeighboringSegments) != 2 || item.NeighboringSegments[0].ID != 1 || item.NeighboringSegments[1].ID != 3 || len(item.GlossaryEntries) != 2 || item.GlossaryEntries[0].Name != "Jesters" || item.GlossaryEntries[1].Name != "Lyra" {
+	if len(item.NeighboringSegments) != 2 || item.NeighboringSegments[0].Position != "preceding" || item.NeighboringSegments[1].Position != "following" || len(item.GlossaryEntries) != 2 || item.GlossaryEntries[0].Name != "Jesters" || item.GlossaryEntries[1].Name != "Lyra" {
 		t.Fatalf("context: %+v", item)
 	}
 	variables, err := BuildValidationVariables(payload.Items, "")
@@ -40,6 +40,9 @@ func TestValidationContextCarriesAdjacentEvidenceAndRelevantDefinitions(t *testi
 	if err := json.Unmarshal([]byte(variables["PayloadJSON"]), &rendered); err != nil {
 		t.Fatal(err)
 	}
+	if strings.Contains(variables["PayloadJSON"], `"id":`) || rendered[0].CorrectionIndex != 1 || payload.Items[0].CorrectionIndex != 7 {
+		t.Fatalf("leaked internal identity or mutated request: %s", variables["PayloadJSON"])
+	}
 	if len(rendered[0].NeighboringSegments) != 2 || rendered[0].GlossaryEntries[0].Summary != "A traveling troupe" {
 		t.Fatalf("rendered context: %+v", rendered)
 	}
@@ -60,7 +63,7 @@ func TestValidationContextFitsBudgetWithoutDiscardingCandidate(t *testing.T) {
 		t.Fatal(err)
 	}
 	enriched := base
-	enriched.NeighboringSegments = []schema.Segment{{ID: 2, Text: strings.Repeat("long context ", 300)}}
+	enriched.NeighboringSegments = []NeighboringSegment{{Position: "following", Text: strings.Repeat("long context ", 300)}}
 	enriched.GlossaryEntries = []schema.GlossaryEntry{{Name: "cat", Summary: strings.Repeat("long definition ", 300)}}
 	fitted := fitValidationContext(enriched, required, estimator)
 	tokens, err := estimateBatchTokens(estimator, []LLMValidationItem{fitted})

internal/framework/validators/llm_integration_test.go

@@ -18,7 +18,7 @@ import (
 )
 
 func TestValidationRepairAndRuntimeFailuresThroughDomainBoundary(t *testing.T) {
-	valid := `{"validations":[{"correction_index":0,"approved":true,"confidence":0.9,"reason":"ok"}]}`
+	valid := `{"validations":[{"correction_index":1,"approved":true,"confidence":0.9,"reason":"ok"}]}`
 	for _, tc := range []struct {
 		name         string
 		responses    []string

internal/framework/validators/llm_models.go

@@ -17,16 +17,17 @@ type StructuredCompletionRequest = structuredoutput.Request
 type StructuredCompletionResponse = structuredoutput.Response
 type StructuredLLMClient = structuredoutput.Client
 
-// LLMValidationItem is one candidate correction payload passed to LLM validators.
+// LLMValidationItem retains the internal proposal index until prompt encoding.
+// BuildValidationVariables substitutes a one-based reference for each batch.
 type LLMValidationItem struct {
 	CorrectionIndex      int                    `json:"correction_index"`
-	SegmentID            int                    `json:"id"`
+	SegmentID            int                    `json:"-"`
 	OriginalText         string                 `json:"original_text"`
 	CorrectedText        string                 `json:"corrected_text"`
 	OriginalSegmentText  string                 `json:"original_segment_text"`
 	CorrectedSegmentText string                 `json:"corrected_segment_text"`
 	Categories           []string               `json:"categories,omitempty"`
-	NeighboringSegments  []schema.Segment       `json:"neighboring_segments,omitempty"`
+	NeighboringSegments  []NeighboringSegment   `json:"neighboring_segments,omitempty"`
 	GlossaryEntries      []schema.GlossaryEntry `json:"glossary_entries,omitempty"`
 }
 
@@ -41,6 +42,7 @@ type LLMValidationRequest struct {
 	Items             []LLMValidationItem `json:"corrections"`
 }
 
+// LLMValidationDecision contains a request-local correction_index, not a proposal index.
 type LLMValidationDecision struct {
 	CorrectionIndex int     `json:"correction_index"`
 	Approved        bool    `json:"approved"`
@@ -51,3 +53,11 @@ type LLMValidationDecision struct {
 type LLMValidationResponse struct {
 	Validations []LLMValidationDecision `json:"validations"`
 }
+
+// NeighboringSegment carries semantic context without internal transcript identity.
+type NeighboringSegment struct {
+	Position   string   `json:"position"`
+	Speaker    string   `json:"speaker"`
+	Text       string   `json:"text"`
+	Categories []string `json:"categories,omitempty"`
+}

internal/framework/validators/llm_partial_test.go

@@ -2,6 +2,7 @@ package validators
 
 import (
 	"context"
+	"encoding/json"
 	"errors"
 	"strings"
 	"testing"
@@ -47,9 +48,9 @@ func TestValidationKeepsSuccessfulBatchesAroundUnavailableBatch(t *testing.T) {
 	}
 	failed := &structuredoutput.Failure{Class: structuredoutput.FailureRecoverable, Reason: outcomes.ReasonTransportFailed, Message: "model transport failed"}
 	client := &scriptedValidationClient{replies: []validationReply{
-		{response: LLMValidationResponse{Validations: []LLMValidationDecision{{CorrectionIndex: 0, Approved: true, Confidence: .9, Reason: "ok"}}}},
+		{response: LLMValidationResponse{Validations: []LLMValidationDecision{{CorrectionIndex: 1, Approved: true, Confidence: .9, Reason: "ok"}}}},
 		{completion: StructuredCompletionResponse{Execution: outcomes.ExecutionCounts{Available: true, ProviderAttempts: 1}}, err: failed},
-		{response: LLMValidationResponse{Validations: []LLMValidationDecision{{CorrectionIndex: 2, Approved: true, Confidence: .9, Reason: "ok"}}}},
+		{response: LLMValidationResponse{Validations: []LLMValidationDecision{{CorrectionIndex: 1, Approved: true, Confidence: .9, Reason: "ok"}}}},
 	}}
 	validator, _ := NewLLMBackedValidator("editorial_review", LLMValidatorTypeEditorialReview)
 	validator.SetTokenEstimator(perItemEstimator{})
@@ -72,8 +73,8 @@ func TestValidationKeepsSuccessfulBatchesAroundUnavailableBatch(t *testing.T) {
 
 	survivors := []proposals.EnrichedCorrectionProposal{candidates[0], candidates[2]}
 	nextClient := &scriptedValidationClient{replies: []validationReply{
-		{response: LLMValidationResponse{Validations: []LLMValidationDecision{{CorrectionIndex: 0, Approved: true, Confidence: .9, Reason: "ok"}}}},
+		{response: LLMValidationResponse{Validations: []LLMValidationDecision{{CorrectionIndex: 1, Approved: true, Confidence: .9, Reason: "ok"}}}},
-		{response: LLMValidationResponse{Validations: []LLMValidationDecision{{CorrectionIndex: 2, Approved: true, Confidence: .9, Reason: "ok"}}}},
+		{response: LLMValidationResponse{Validations: []LLMValidationDecision{{CorrectionIndex: 1, Approved: true, Confidence: .9, Reason: "ok"}}}},
 	}}
 	nextValidator, _ := NewLLMBackedValidator("meaning_reversal_review", LLMValidatorTypeMeaningReversal)
 	nextValidator.SetTokenEstimator(perItemEstimator{})
@@ -84,6 +85,18 @@ func TestValidationKeepsSuccessfulBatchesAroundUnavailableBatch(t *testing.T) {
 	if err != nil || len(next.Decisions) != 2 || len(nextClient.calls) != 2 || *nextClient.calls[0].Context.BatchIndex != 0 || *nextClient.calls[1].Context.BatchIndex != 1 {
 		t.Fatalf("later validator rebatching: %+v, calls=%+v, err=%v", next, nextClient.calls, err)
 	}
+	for _, call := range append(client.calls, nextClient.calls...) {
+		var items []LLMValidationItem
+		if err := json.Unmarshal([]byte(call.Variables["PayloadJSON"]), &items); err != nil {
+			t.Fatal(err)
+		}
+		if len(items) != 1 || items[0].CorrectionIndex != 1 {
+			t.Fatalf("batch references must restart: %+v", items)
+		}
+	}
+	if !next.Decisions[0].Approved || !next.Decisions[1].Approved || next.Decisions[0].ProposalIndex != 0 || next.Decisions[1].ProposalIndex != 2 {
+		t.Fatalf("rebatching lost identities: %+v", next.Decisions)
+	}
 }
 
 func TestValidationDomainInvalidAfterRepairIsAnError(t *testing.T) {
@@ -103,7 +116,7 @@ func TestValidationDomainInvalidAfterRepairIsAnError(t *testing.T) {
 func TestValidationExtraDecisionsMergeActualRecoveries(t *testing.T) {
 	client := &scriptedValidationClient{replies: []validationReply{{
 		response: LLMValidationResponse{Validations: []LLMValidationDecision{
-			{CorrectionIndex: 0, Approved: true, Confidence: .9, Reason: "ok"},
+			{CorrectionIndex: 1, Approved: true, Confidence: .9, Reason: "ok"},
 			{CorrectionIndex: 99, Approved: true, Confidence: .9, Reason: "extra"},
 		}},
 		completion: StructuredCompletionResponse{Execution: outcomes.ExecutionCounts{Available: true, ProviderAttempts: 3, TransientRetries: 1, StructuralRepairs: 1}},
@@ -124,9 +137,9 @@ func TestValidationExtraDecisionsMergeActualRecoveries(t *testing.T) {
 func TestValidationFatalFailureKeepsEarlierFactsAndStops(t *testing.T) {
 	cause := errors.New("broken embedded schema")
 	client := &scriptedValidationClient{replies: []validationReply{
-		{response: LLMValidationResponse{Validations: []LLMValidationDecision{{CorrectionIndex: 0, Approved: true, Confidence: .9, Reason: "ok"}}}},
+		{response: LLMValidationResponse{Validations: []LLMValidationDecision{{CorrectionIndex: 1, Approved: true, Confidence: .9, Reason: "ok"}}}},
 		{err: &structuredoutput.Failure{Class: structuredoutput.FailureFatal, Reason: outcomes.ReasonSetupFailure, Message: "model setup failed", Cause: cause}},
-		{response: LLMValidationResponse{Validations: []LLMValidationDecision{{CorrectionIndex: 2, Approved: true, Confidence: .9, Reason: "ok"}}}},
+		{response: LLMValidationResponse{Validations: []LLMValidationDecision{{CorrectionIndex: 1, Approved: true, Confidence: .9, Reason: "ok"}}}},
 	}}
 	validator, _ := NewLLMBackedValidator("editorial_review", LLMValidatorTypeEditorialReview)
 	validator.SetTokenEstimator(perItemEstimator{})

internal/framework/validators/llm_prompt_variables.go

@@ -8,7 +8,7 @@ import (
 )
 
 func BuildValidationVariables(validationPayload []LLMValidationItem, transcriptDescription string) (map[string]string, error) {
-	payloadJSON, err := json.MarshalIndent(validationPayload, "", "  ")
+	payloadJSON, err := json.MarshalIndent(localValidationItems(validationPayload), "", "  ")
 	if err != nil {
 		return nil, fmt.Errorf("marshal validation payload: %w", err)
 	}
@@ -17,3 +17,12 @@ func BuildValidationVariables(validationPayload []LLMValidationItem, transcriptD
 		"PayloadJSON":                string(payloadJSON),
 	}, nil
 }
+
+// localValidationItems gives each actual batch its own short references.
+func localValidationItems(items []LLMValidationItem) []LLMValidationItem {
+	local := append([]LLMValidationItem(nil), items...)
+	for i := range local {
+		local[i].CorrectionIndex = i + 1
+	}
+	return local
+}

internal/framework/validators/llm_validators.go

@@ -163,6 +163,7 @@ func (v *LLMBackedValidator) Validate(ctx context.Context, req Request) (Result,
 					"validator_name":   v.name,
 					"validator_type":   v.validatorType,
 					"batch_index":      batch.BatchIndex,
+					"proposal_indices": batchProposalIndices(batch.Items),
 					"prompt":           completion.Prompt,
 					"response_schema":  responseSchema.DiagnosticsMap(),
 				},
@@ -303,22 +304,18 @@ type validationReconciliation struct {
 }
 
 func reconcileLLMResponse(items []LLMValidationItem, response LLMValidationResponse) validationReconciliation {
-	expected := make(map[int]struct{}, len(items))
-	for _, item := range items {
-		expected[item.CorrectionIndex] = struct{}{}
-	}
 	result := validationReconciliation{Decisions: make([]Decision, 0, len(items))}
 	seen := make(map[int][]LLMValidationDecision, len(items))
 	for _, d := range response.Validations {
-		if _, ok := expected[d.CorrectionIndex]; !ok {
+		if d.CorrectionIndex < 1 || d.CorrectionIndex > len(items) {
 			result.IgnoredIndices = append(result.IgnoredIndices, d.CorrectionIndex)
 			continue
 		}
 		seen[d.CorrectionIndex] = append(seen[d.CorrectionIndex], d)
 	}
 
-	for _, item := range items {
+	for i, item := range items {
-		entries := seen[item.CorrectionIndex]
+		entries := seen[i+1]
 		if len(entries) == 0 {
 			result.Decisions = append(result.Decisions, unavailable(item.CorrectionIndex, ReasonMissingValidationDecision, "required model validation decision was missing"))
 			result.UnavailableIndices = append(result.UnavailableIndices, item.CorrectionIndex)

internal/framework/validators/llm_validators_test.go

@@ -152,8 +152,8 @@ func TestChunkLLMValidationItemsMultipleBatchesStableNoDropNoDup(t *testing.T) {
 
 func TestLLMBackedValidatorApprovalAndRejection(t *testing.T) {
 	client := &fakeStructuredLLMClient{responses: []LLMValidationResponse{{Validations: []LLMValidationDecision{
-		{CorrectionIndex: 0, Approved: true, Confidence: 0.9, Reason: "ok"},
+		{CorrectionIndex: 1, Approved: true, Confidence: 0.9, Reason: "ok"},
-		{CorrectionIndex: 1, Approved: false, Confidence: 0.95, Reason: "bad"},
+		{CorrectionIndex: 2, Approved: false, Confidence: 0.95, Reason: "bad"},
 	}}}}
 	v, err := NewLLMBackedValidator("spoken_form_plausibility_review", LLMValidatorTypeSpokenFormPlausibility)
 	if err != nil {
@@ -224,13 +224,14 @@ func TestLLMBackedValidatorReconcilesDecisionIdentities(t *testing.T) {
 		unavailable []int
 		ignored     []int
 	}{
-		{"missing and extra", []LLMValidationDecision{approve(7), reject(18), approve(99)}, []expectedDecision{{7, true, ReasonApproved}, {11, false, ReasonMissingValidationDecision}, {18, false, "llm_rejected"}}, []int{11}, []int{99}},
+		{"internal index is not a local reference", []LLMValidationDecision{approve(7), approve(2), reject(3)}, []expectedDecision{{7, false, ReasonMissingValidationDecision}, {11, true, ReasonApproved}, {18, false, "llm_rejected"}}, []int{7}, []int{7}},
-		{"conflicting duplicate", []LLMValidationDecision{approve(7), approve(11), reject(11), reject(18)}, []expectedDecision{{7, true, ReasonApproved}, {11, false, ReasonDuplicateValidationDecision}, {18, false, "llm_rejected"}}, []int{11}, nil},
+		{"missing and extra", []LLMValidationDecision{approve(1), reject(3), approve(99)}, []expectedDecision{{7, true, ReasonApproved}, {11, false, ReasonMissingValidationDecision}, {18, false, "llm_rejected"}}, []int{11}, []int{99}},
-		{"identical duplicate", []LLMValidationDecision{approve(7), approve(11), approve(11), reject(18)}, []expectedDecision{{7, true, ReasonApproved}, {11, false, ReasonDuplicateValidationDecision}, {18, false, "llm_rejected"}}, []int{11}, nil},
+		{"conflicting duplicate", []LLMValidationDecision{approve(1), approve(2), reject(2), reject(3)}, []expectedDecision{{7, true, ReasonApproved}, {11, false, ReasonDuplicateValidationDecision}, {18, false, "llm_rejected"}}, []int{11}, nil},
-		{"extra only", []LLMValidationDecision{approve(99), reject(18), approve(7), approve(99), approve(11)}, []expectedDecision{{7, true, ReasonApproved}, {11, true, ReasonApproved}, {18, false, "llm_rejected"}}, nil, []int{99, 99}},
+		{"identical duplicate", []LLMValidationDecision{approve(1), approve(2), approve(2), reject(3)}, []expectedDecision{{7, true, ReasonApproved}, {11, false, ReasonDuplicateValidationDecision}, {18, false, "llm_rejected"}}, []int{11}, nil},
+		{"extra only", []LLMValidationDecision{approve(99), reject(3), approve(1), approve(99), approve(2)}, []expectedDecision{{7, true, ReasonApproved}, {11, true, ReasonApproved}, {18, false, "llm_rejected"}}, nil, []int{99, 99}},
 		{"empty", nil, []expectedDecision{{7, false, ReasonMissingValidationDecision}, {11, false, ReasonMissingValidationDecision}, {18, false, ReasonMissingValidationDecision}}, []int{7, 11, 18}, nil},
-		{"reordered", []LLMValidationDecision{reject(18), approve(11), approve(7)}, []expectedDecision{{7, true, ReasonApproved}, {11, true, ReasonApproved}, {18, false, "llm_rejected"}}, nil, nil},
+		{"reordered", []LLMValidationDecision{reject(3), approve(2), approve(1)}, []expectedDecision{{7, true, ReasonApproved}, {11, true, ReasonApproved}, {18, false, "llm_rejected"}}, nil, nil},
-		{"invalid confidence", []LLMValidationDecision{approve(7), {CorrectionIndex: 11, Approved: true, Confidence: 1.2}, reject(18)}, []expectedDecision{{7, true, ReasonApproved}, {11, false, ReasonInvalidValidationDecision}, {18, false, "llm_rejected"}}, []int{11}, nil},
+		{"invalid confidence", []LLMValidationDecision{approve(1), {CorrectionIndex: 2, Approved: true, Confidence: 1.2}, reject(3)}, []expectedDecision{{7, true, ReasonApproved}, {11, false, ReasonInvalidValidationDecision}, {18, false, "llm_rejected"}}, []int{11}, nil},
 	} {
 		t.Run(tc.name, func(t *testing.T) {
 			client := &fakeStructuredLLMClient{responses: []LLMValidationResponse{{Validations: tc.entries}}}
@@ -337,7 +338,7 @@ func TestLLMBackedValidatorOversizedSingleProposalRejectsOnlyThatProposal(t *tes
 
 func TestLLMBackedValidatorDiagnosticsIncludeSchemaMetadata(t *testing.T) {
 	client := &fakeStructuredLLMClient{responses: []LLMValidationResponse{{Validations: []LLMValidationDecision{
-		{CorrectionIndex: 0, Approved: true, Confidence: 0.9, Reason: "ok"},
+		{CorrectionIndex: 1, Approved: true, Confidence: 0.9, Reason: "ok"},
 	}}}}
 	writer := &captureValidationDiagnosticsWriter{}
 	v, err := NewLLMBackedValidator("spoken_form_plausibility_review", LLMValidatorTypeSpokenFormPlausibility)
@@ -376,7 +377,7 @@ func TestEveryValidationTypeSelectsEmbeddedPromptAndVariables(t *testing.T) {
 	}
 	for _, tc := range cases {
 		t.Run(string(tc.kind), func(t *testing.T) {
-			client := &fakeStructuredLLMClient{responses: []LLMValidationResponse{{Validations: []LLMValidationDecision{{CorrectionIndex: 0, Approved: true, Confidence: .9, Reason: "ok"}}}}}
+			client := &fakeStructuredLLMClient{responses: []LLMValidationResponse{{Validations: []LLMValidationDecision{{CorrectionIndex: 1, Approved: true, Confidence: .9, Reason: "ok"}}}}}
 			validator, err := NewLLMBackedValidator(string(tc.kind), tc.kind)
 			if err != nil {
 				t.Fatal(err)
@@ -401,7 +402,7 @@ func TestEveryValidationTypeSelectsEmbeddedPromptAndVariables(t *testing.T) {
 func TestValidationDiagnosticFailureDoesNotChangeDecision(t *testing.T) {
 	cause := errors.New("diagnostic storage unavailable")
 	req := makeReq([]proposals.EnrichedCorrectionProposal{mk(0, "gestures", "Jesters")})
-	req.LLMClient = &fakeStructuredLLMClient{responses: []LLMValidationResponse{{Validations: []LLMValidationDecision{{CorrectionIndex: 0, Approved: true, Confidence: 1, Reason: "ok"}}}}}
+	req.LLMClient = &fakeStructuredLLMClient{responses: []LLMValidationResponse{{Validations: []LLMValidationDecision{{CorrectionIndex: 1, Approved: true, Confidence: 1, Reason: "ok"}}}}}
 	req.DiagnosticsWriter = &captureValidationDiagnosticsWriter{err: cause}
 	var reported error
 	req.ReportDiagnosticError = func(err error) { reported = err }

internal/modules/glossary/module_test.go

@@ -93,7 +93,7 @@ func TestGlossaryModuleProposeMapsCorrectionsAndWritesDiagnostics(t *testing.T)
 		responses: []proposal_generation.StructuredCorrectionSet{
 			{
 				Corrections: []proposal_generation.StructuredCorrectionProposal{
-					{TargetSegmentID: 1, OriginalText: "gestures", CorrectedText: "Jesters", Confidence: 0.94},
+					{SegmentReference: 1, OriginalText: "gestures", CorrectedText: "Jesters", Confidence: 0.94},
 				},
 			},
 		},

internal/modules/grammar/module_test.go

@@ -42,7 +42,7 @@ func (f *fakeLLMClient) CompleteStructured(ctx context.Context, req contracts.St
 	items, _ := payload["corrections"].([]map[string]any)
 	for _, item := range items {
 		target.Corrections = append(target.Corrections, proposal_generation.StructuredCorrectionProposal{
-			TargetSegmentID: item["id"].(int),
+			SegmentReference: item["id"].(int),
 			OriginalText:     item["original_text"].(string),
 			CorrectedText:    item["corrected_text"].(string),
 			Confidence:       item["confidence"].(float64),

internal/modules/homophones/module_test.go

@@ -94,7 +94,7 @@ func TestHomophonesModuleProposeMapsCorrectionsAndWritesDiagnostics(t *testing.T
 		responses: []proposal_generation.StructuredCorrectionSet{
 			{
 				Corrections: []proposal_generation.StructuredCorrectionProposal{
-					{TargetSegmentID: 1, OriginalText: "gestures", CorrectedText: "Jesters", Confidence: 0.94},
+					{SegmentReference: 1, OriginalText: "gestures", CorrectedText: "Jesters", Confidence: 0.94},
 				},
 			},
 		},

internal/modules/spoken_word/module_test.go

@@ -94,7 +94,7 @@ func TestSpokenWordModuleProposeMapsCorrectionsAndWritesDiagnostics(t *testing.T
 		responses: []proposal_generation.StructuredCorrectionSet{
 			{
 				Corrections: []proposal_generation.StructuredCorrectionProposal{
-					{TargetSegmentID: 1, OriginalText: "I I think", CorrectedText: "I think", Confidence: 0.94},
+					{SegmentReference: 1, OriginalText: "I I think", CorrectedText: "I think", Confidence: 0.94},
 				},
 			},
 		},

internal/prompts/assets/prompts/modules.glossary.proposal.yaml

@@ -19,16 +19,14 @@ messages:
       - Preserve canonical glossary capitalization for protected names and aliases, even if they look unusual.
       - If a segment includes categories, treat them as additional transcript context.
       - Plural forms of glossary names and aliases are allowed targets when spoken similarity and context support them, even if the plural is not explicitly listed in the glossary.
-      - Use the exact id from the input segment.
+      - Use the segment's local id (1, 2, 3, ...) from this request to identify the correction target.
       - Use a nonempty original_text copied exactly from the supplied segment; matching is literal and case-sensitive, including whitespace and punctuation. Include unchanged surrounding text when needed to identify the intended edit, preserving that context in corrected_text.
       - Audita replaces every literal occurrence of original_text within the target segment. If only one occurrence should change, include enough unchanged context to distinguish it and preserve that context in the replacement.
       - corrected_text must be only the replacement text for that span, not the full corrected segment text unless the whole segment is the replacement span.
-      - Each returned correction must contain only id, original_text, corrected_text, and confidence.
       - Do not return corrections where original_text and corrected_text are identical.
-      - Do not return speaker, start, or end fields.
       - Return only proposed corrections. Multiple corrections may use the same segment id; each original_text refers to the supplied segment, not the imagined result of another proposal.
       - confidence must be between 0.0 and 1.0 and express confidence that the proposed correction is warranted.
-      - Return a JSON object with a corrections array, matching the supplied schema. If no corrections are needed, return {"corrections": []}.
+      - Follow the supplied JSON schema. Output example: {"corrections": [{"id": 1, "original_text": "gestures", "corrected_text": "Jesters", "confidence": 0.95}]}. If no corrections are needed, return {"corrections": []}.
 
       {{ .HardeningPolicy }}
 

internal/prompts/assets/prompts/modules.grammar.proposal.yaml

@@ -20,18 +20,16 @@ messages:
       - Do not replace, Anglicize, normalize, lowercase, or otherwise alter protected glossary names or aliases away from their glossary spelling.
       - Preserve canonical glossary capitalization for protected names and aliases, even if they look unusual.
       - If a segment includes categories, treat them as additional transcript context.
-      - Use the exact id from the input segment.
+      - Use the segment's local id (1, 2, 3, ...) from this request to identify the correction target.
       - Use a nonempty original_text copied exactly from the supplied segment; matching is literal and case-sensitive, including whitespace and punctuation. Include unchanged surrounding text when needed to identify the intended edit, preserving that context in corrected_text.
       - Choose an original_text span that appears exactly once in the current segment text.
       - corrected_text must be only the replacement text for that span, not the full corrected segment text unless the whole segment is the replacement span.
       - A whitespace-only original_text can identify a spacing change: replacing the two spaces in "hello  world" with one space yields "hello world". An empty corrected_text deletes its identified span when punctuation or spacing should be removed.
       - Omit a correction that would leave the segment blank; do not invent padding or content to keep it nonblank.
-      - Each returned correction must contain only id, original_text, corrected_text, and confidence.
       - Do not return corrections where original_text and corrected_text are identical.
-      - Do not return speaker, start, or end fields.
       - Return only proposed corrections. Multiple corrections may use the same segment id; each original_text refers to the supplied segment, not the imagined result of another proposal.
       - confidence must be between 0.0 and 1.0 and express confidence that the proposed correction is warranted.
-      - Return a JSON object with a corrections array, matching the supplied schema. If no corrections are needed, return {"corrections": []}.
+      - Follow the supplied JSON schema. Output example: {"corrections": [{"id": 1, "original_text": "hello  world", "corrected_text": "hello world", "confidence": 0.95}]}. If no corrections are needed, return {"corrections": []}.
 
       {{ .HardeningPolicy }}
 

internal/prompts/assets/prompts/modules.homophones.proposal.yaml

@@ -19,16 +19,14 @@ messages:
       - Do not replace, Anglicize, normalize, lowercase, or otherwise alter protected glossary names or aliases that already appear correctly in the transcript.
       - Preserve canonical glossary capitalization for protected names and aliases, even if they look unusual.
       - If a segment includes categories, treat them as additional transcript context.
-      - Use the exact id from the input segment.
+      - Use the segment's local id (1, 2, 3, ...) from this request to identify the correction target.
       - Use a nonempty original_text copied exactly from the supplied segment; matching is literal and case-sensitive, including whitespace and punctuation. Include unchanged surrounding text when needed to identify the intended edit, preserving that context in corrected_text.
       - Choose an original_text span that appears exactly once in the current segment text.
       - corrected_text must be only the replacement text for that span, not the full corrected segment text unless the whole segment is the replacement span.
-      - Each returned correction must contain only id, original_text, corrected_text, and confidence.
       - Do not return corrections where original_text and corrected_text are identical.
-      - Do not return speaker, start, or end fields.
       - Return only proposed corrections. Multiple corrections may use the same segment id; each original_text refers to the supplied segment, not the imagined result of another proposal.
       - confidence must be between 0.0 and 1.0 and express confidence that the proposed correction is warranted.
-      - Return a JSON object with a corrections array, matching the supplied schema. If no corrections are needed, return {"corrections": []}.
+      - Follow the supplied JSON schema. Output example: {"corrections": [{"id": 1, "original_text": "dam", "corrected_text": "damn", "confidence": 0.95}]}. If no corrections are needed, return {"corrections": []}.
 
       {{ .HardeningPolicy }}
 

internal/prompts/assets/prompts/modules.spoken_word.proposal.yaml

@@ -22,18 +22,16 @@ messages:
       - Do not replace, Anglicize, normalize, lowercase, or otherwise alter protected glossary names or aliases that already appear correctly in the transcript.
       - Preserve canonical glossary capitalization for protected names and aliases, even if they look unusual.
       - If a segment includes categories, treat them as additional transcript context.
-      - Use the exact id from the input segment.
+      - Use the segment's local id (1, 2, 3, ...) from this request to identify the correction target.
       - Use a nonempty original_text copied exactly from the supplied segment; matching is literal and case-sensitive, including whitespace and punctuation. Include unchanged surrounding text when needed to identify the intended edit, preserving that context in corrected_text.
       - Choose an original_text span that appears exactly once in the current segment text.
       - corrected_text must be only the replacement text for that span, not the full corrected segment text unless the whole segment is the replacement span.
       - An empty corrected_text deletes the identified span. For example, removing "uh " from "uh hello" leaves "hello"; include adjacent whitespace or punctuation in the span when needed for the intended literal result.
       - Omit a correction that would leave the segment blank; do not invent padding or content to keep it nonblank.
-      - Each returned correction must contain only id, original_text, corrected_text, and confidence.
       - Do not return corrections where original_text and corrected_text are identical.
-      - Do not return speaker, start, or end fields.
       - Return only proposed corrections. Multiple corrections may use the same segment id; each original_text refers to the supplied segment, not the imagined result of another proposal.
       - confidence must be between 0.0 and 1.0 and express confidence that the proposed correction is warranted.
-      - Return a JSON object with a corrections array, matching the supplied schema. If no corrections are needed, return {"corrections": []}.
+      - Follow the supplied JSON schema. Output example: {"corrections": [{"id": 1, "original_text": "uh ", "corrected_text": "", "confidence": 0.95}]}. If no corrections are needed, return {"corrections": []}.
 
       {{ .HardeningPolicy }}
 

internal/prompts/assets/prompts/validators.editorial_review.yaml

@@ -9,7 +9,7 @@ messages:
       Review these proposed transcript corrections and decide whether each one is an acceptable editorial revision.
 
       Rules:
-      - Return exactly one validation decision for each supplied correction_index, including rejected corrections. Copy correction_index exactly; do not substitute the segment id, renumber entries, or invent indices. Return a JSON object with a validations array matching the supplied schema.
+      - Return one decision for each numbered candidate, including rejections. correction_index identifies the candidate within this request (1, 2, 3, ...). Follow the supplied JSON schema. Output example: {"validations": [{"correction_index": 1, "approved": true, "confidence": 0.95, "reason": "The edit preserves meaning."}]}.
       - Approve editorial revisions that preserve the underlying meaning of the segment.
       - Approve dysfluency cleanup, including cleanup of repeated words, repeated short phrases, filler words, hesitation artifacts, and similar common spoken dysfluencies.
       - Approve punctuation, spacing, capitalization, and article cleanup when they preserve meaning.
@@ -21,8 +21,7 @@ messages:
       - Reject broad paraphrase, substantive semantic changes, substantive meaning changes, and revisions that materially change, obscure, distort, or reverse the segment's meaning.
       - Evaluate the full original_segment_text and corrected_segment_text, not only the replacement span.
       - An empty corrected_text deletes the identified span; judge the actual before/after segment, including adjacent spacing and punctuation. Do not approve or reject solely because the replacement is empty.
-      - If supplied, neighboring_segments are adjacent source segments for context, not additional correction requests. glossary_entries provide reference definitions for terms mentioned in the before/after text or those neighbors. Use this evidence and categories where relevant; do not assume access to audio or other transcript passages.
+      - If supplied, neighboring_segments identify preceding/following source context, not additional correction requests. glossary_entries provide reference definitions for terms mentioned in the before/after text or those neighbors. Use this evidence and categories where relevant; do not assume access to audio or other transcript passages.
-      - Each returned validation must contain only correction_index, approved, confidence, and reason.
       - confidence must be between 0.0 and 1.0 and express confidence in your verdict. A confident rejection may have high confidence; approved, not confidence, states the verdict.
 
       {{ .HardeningPolicy }}

internal/prompts/assets/prompts/validators.meaning_reversal_review.yaml

@@ -9,7 +9,7 @@ messages:
       Review these proposed transcript corrections and decide whether each one avoids reversing the segment meaning.
 
       Rules:
-      - Return exactly one validation decision for each supplied correction_index, including rejected corrections. Copy correction_index exactly; do not substitute the segment id, renumber entries, or invent indices. Return a JSON object with a validations array matching the supplied schema.
+      - Return one decision for each numbered candidate, including rejections. correction_index identifies the candidate within this request (1, 2, 3, ...). Follow the supplied JSON schema. Output example: {"validations": [{"correction_index": 1, "approved": true, "confidence": 0.95, "reason": "The edit does not reverse meaning."}]}.
       - Reject corrections that introduce antonyms or otherwise reverse the meaning of the original segment.
       - Reject examples like changing "visible" to "invisible" or "up" to "down" when that reverses the segment meaning.
       - Evaluate the full original_segment_text and corrected_segment_text, not only the replacement span.
@@ -17,8 +17,7 @@ messages:
       - For example, deleting "not " from "The door is not open" to produce "The door is open" reverses meaning. Judge the actual before/after context, not the presence of a particular token alone.
       - Do not reject a correction merely because the literal written word changes.
       - Do not act as a general semantic-style reviewer; this validator is only a guard against antonyms and meaning reversals.
-      - If supplied, neighboring_segments are adjacent source segments for context, not additional correction requests. glossary_entries provide reference definitions for terms mentioned in the before/after text or those neighbors. Use this evidence and categories where relevant; do not assume access to audio or other transcript passages.
+      - If supplied, neighboring_segments identify preceding/following source context, not additional correction requests. glossary_entries provide reference definitions for terms mentioned in the before/after text or those neighbors. Use this evidence and categories where relevant; do not assume access to audio or other transcript passages.
-      - Each returned validation must contain only correction_index, approved, confidence, and reason.
       - confidence must be between 0.0 and 1.0 and express confidence in your verdict. A confident rejection may have high confidence; approved, not confidence, states the verdict.
 
       {{ .HardeningPolicy }}

internal/prompts/assets/prompts/validators.spoken_form_plausibility.yaml

@@ -9,15 +9,14 @@ messages:
       Review these proposed transcript corrections and decide whether each one is a plausible spoken-form correction.
 
       Rules:
-      - Return exactly one validation decision for each supplied correction_index, including rejected corrections. Copy correction_index exactly; do not substitute the segment id, renumber entries, or invent indices. Return a JSON object with a validations array matching the supplied schema.
+      - Return one decision for each numbered candidate, including rejections. correction_index identifies the candidate within this request (1, 2, 3, ...). Follow the supplied JSON schema. Output example: {"validations": [{"correction_index": 1, "approved": true, "confidence": 0.95, "reason": "The replacement is a plausible mishearing."}]}.
       - Approve when the original text and corrected text are plausibly related by homophone confusion, phonetic similarity, or a common spoken-word mistranscription, and the surrounding segment context supports the correction.
       - Examples that may be approved when context supports them: changing "gestures" to "Jesters", "rank" to "Hrank", or "dam" to "damn".
       - Reject unrelated substitutions like changing "Lyra" to "Jesters".
       - Judge spoken-form plausibility, not whether the correction is cleaner, more formal, or more grammatical.
       - For every correction, evaluate the full original_segment_text and corrected_segment_text, not only the replacement span. When a span is deleted, an empty corrected_text alone does not establish or rule out a spoken-form mistranscription.
       - Do not approve paraphrases, stylistic rewrites, or arbitrary semantic substitutions.
-      - If supplied, neighboring_segments are adjacent source segments for context, not additional correction requests. glossary_entries provide reference definitions for terms mentioned in the before/after text or those neighbors. Use this evidence and categories where relevant; do not assume access to audio or other transcript passages.
+      - If supplied, neighboring_segments identify preceding/following source context, not additional correction requests. glossary_entries provide reference definitions for terms mentioned in the before/after text or those neighbors. Use this evidence and categories where relevant; do not assume access to audio or other transcript passages.
-      - Each returned validation must contain only correction_index, approved, confidence, and reason.
       - confidence must be between 0.0 and 1.0 and express confidence in your verdict. A confident rejection may have high confidence; approved, not confidence, states the verdict.
 
       {{ .HardeningPolicy }}