Complete Phase 14 homophones module

docs/architecture.md

@@ -35,12 +35,13 @@ Implemented today:
 - Production module registry scaffolding with known-key recognition and explicit unsupported/unimplemented errors.
 - Production `grammar` module implementation in `internal/modules/grammar`.
 - Production `glossary` module implementation in `internal/modules/glossary`.
+- Production `homophones` module implementation in `internal/modules/homophones`.
 - Explicit runtime support for `--modules grammar` through the production runner path.
 - Explicit runtime support for `--modules glossary`, including repeated stages such as `--modules glossary,glossary`.
+- Explicit runtime support for `--modules homophones` through the production runner path.
 
 Not implemented in CLI runtime path today:
-- Real module execution pipeline for `homophones` and `spoken_word`.
+- Real module execution pipeline for `spoken_word`.
-- Real domain proposal prompts for remaining production modules.
 - End-to-end transcript polishing with the full default module sequence.
 
 Phase sequencing note:
@@ -49,7 +50,8 @@ Phase sequencing note:
 - Phase 11 shared proposal-generation framework and module-registry scaffolding are complete;
 - Phase 12 grammar module implementation and explicit runtime wiring are complete;
 - Phase 13 glossary module and protected-term behavior are complete;
-- next recommended phase is Phase 14 (homophones module).
+- Phase 14 homophones module implementation and explicit runtime wiring are complete;
+- next recommended phase is Phase 15 (spoken-word module).
 
 ## Actual Go package layout
 
@@ -116,6 +118,10 @@ internal/modules/glossary/
   module.go
   prompt.go
 
+internal/modules/homophones/
+  module.go
+  prompt.go
+
 internal/framework/validators/
   models.go
   deterministic.go
@@ -151,17 +157,17 @@ Current runtime flow (`internal/cli/run.go`):
 10. Chunk normalized transcript and compute chunk summaries.
 11. Write chunking summary artifact.
 12. Execute runner modules sequentially when:
-    - `--modules` is explicitly provided (production grammar/glossary paths); or
+    - `--modules` is explicitly provided (production grammar/glossary/homophones paths); or
     - a test/injected module factory is provided.
 13. Output working transcript to `--output` file or stdout.
-14. Build process report (`phase` currently set to `phase13-glossary-module`).
+14. Build process report (`phase` currently set to `phase14-homophones-module`).
 15. Optionally write `--report-json`; always write run-dir `report.json`.
 16. Apply work-dir retention.
 
 Important behavior details:
-- Glossary is validated and is used for explicit glossary/grammar module correction paths.
+- Glossary is validated and is used for explicit glossary/grammar/homophones module correction paths.
 - Default production CLI behavior remains deterministic normalization/chunking/reporting unless modules are explicitly selected with `--modules`.
-- Explicit `--modules grammar` and `--modules glossary` run production module paths with LLM-backed proposal generation and validator-chain execution.
+- Explicit `--modules grammar`, `--modules glossary`, and `--modules homophones` run production module paths with LLM-backed proposal generation and validator-chain execution.
 - Default runs (without explicit module selection) do not perform LLM calls.
 - Success path is generally quiet on stderr.
 - Source IDs are preserved into a canonical transcript before normalization; normalization then reassigns output IDs sequentially from `1`.
@@ -210,7 +216,7 @@ Implemented config surfaces include:
 - work-dir and retention mode
 
 Current caveat:
-- LLM/module-related settings are active for explicit grammar/glossary runs; the default non-explicit path remains deterministic.
+- LLM/module-related settings are active for explicit grammar/glossary/homophones runs; the default non-explicit path remains deterministic.
 
 ## Implemented structured LLM infrastructure
 `internal/framework/contracts` now defines a typed structured-completion contract:
@@ -227,7 +233,7 @@ Current caveat:
 
 Current runtime boundary:
 - the default CLI runtime path (without explicit module selection) still does not instantiate the full production module sequence.
-- LLM calls are exercised in production when `--modules grammar` or `--modules glossary` is explicitly requested and in tests when fake/injected clients are used.
+- LLM calls are exercised in production when `--modules grammar`, `--modules glossary`, or `--modules homophones` is explicitly requested and in tests when fake/injected clients are used.
 
 `internal/framework/llm` also provides:
 - a bounded `Scheduler` for controlled concurrent LLM calls with reliable permit release;
@@ -334,7 +340,7 @@ This helper only produces candidate proposals; validator-chain execution and pro
   - diagnostics directory context
 - returns explicit errors for unknown keys (`unsupported_module`) and recognized-but-unimplemented keys (`unimplemented_module`).
 
-The `grammar` and `glossary` module keys are now registered and constructible. `homophones` and `spoken_word` remain recognized-but-unimplemented.
+The `grammar`, `glossary`, and `homophones` module keys are now registered and constructible. `spoken_word` remains recognized-but-unimplemented.
 
 ## Implemented grammar production module
 `internal/modules/grammar` now provides the first production module:
@@ -368,6 +374,19 @@ The `grammar` and `glossary` module keys are now registered and constructible. `
 
 This vocabulary is used by deterministic validators for both glossary-stage and non-glossary-stage protection checks, keeping protected-term guardrails active across modules.
 
+## Implemented homophones production module
+`internal/modules/homophones` now provides the third production module:
+- prompt builder aligned to Python homophones-module intent, constrained to conservative homophone/near-homophone/mistranscription corrections;
+- prompt context includes protected glossary names/aliases/plurals to avoid damaging known terms;
+- explicit guardrails against punctuation cleanup, grammar cleanup, style rewriting, summarization, and content invention;
+- proposal generation through `internal/framework/proposal_generation` and `contracts.StructuredLLMClient`;
+- scheduler-aware proposal calls through existing `contracts.LLMScheduler` hooks;
+- replacement policy `require_unique` (matching Python homophones behavior);
+- validator chain integration using existing deterministic + LLM-backed validators;
+- homophones confidence threshold enforcement through existing validator/config infrastructure;
+- protected-term guardrails for non-glossary modules remain active and are exercised through the homophones path;
+- module-level reporting and diagnostics capture through existing runner/reporting paths.
+
 ## Reports and diagnostics (implemented)
 Current per-run artifacts include:
 - `source-transcript.json`
@@ -407,7 +426,7 @@ Retention modes implemented in `ApplyRetention`:
 
 Current runtime note:
 - default non-explicit runs usually have no module-level skipped corrections, so `auto` commonly removes clean successful run directories.
-- explicit grammar/glossary runs can produce validator rejections and application skips, which are reflected in reports and retention input.
+- explicit grammar/glossary/homophones runs can produce validator rejections and application skips, which are reflected in reports and retention input.
 
 Intentionally deferred to module/LLM phases:
 - real domain proposal prompts and production module implementations remain tied to later module phases.
@@ -430,14 +449,15 @@ Implemented tests currently cover:
 - production module-registry known-key recognition and unsupported/unimplemented error behavior (`internal/framework/modules/*_test.go`, `internal/cli/run_test.go`)
 - production grammar module prompt constraints, proposal mapping, validator-chain behavior, confidence-threshold enforcement, diagnostics redaction, and explicit CLI/runtime integration (`internal/modules/grammar/*_test.go`, `internal/cli/run_test.go`, `internal/framework/runner/*_test.go`)
 - production glossary module prompt constraints, proposal mapping, validator-chain behavior, confidence-threshold enforcement, diagnostics redaction, repeated-stage behavior, and explicit CLI/runtime integration (`internal/modules/glossary/*_test.go`, `internal/cli/run_test.go`, `internal/framework/runner/*_test.go`)
+- production homophones module prompt constraints, proposal mapping, validator-chain behavior, confidence-threshold enforcement, diagnostics redaction, protected-term behavior, and explicit CLI/runtime integration (`internal/modules/homophones/*_test.go`, `internal/cli/run_test.go`, `internal/framework/runner/*_test.go`)
 - glossary-derived protected-term extraction and stable behavior (`internal/framework/validators/protected_terms_test.go`)
 
-Not covered yet (because not implemented): production `homophones` and `spoken_word` modules plus full default-sequence transcript-polishing runtime behavior.
+Not covered yet (because not implemented): production `spoken_word` module plus full default-sequence transcript-polishing runtime behavior.
 
 ## Intended final architecture (not yet implemented)
 The intended end-state still matches the rewrite plan:
 - sequential module pipeline over a mutable working transcript
-- real module implementations (`homophones`, `spoken_word`)
+- real module implementation (`spoken_word`)
 - structured LLM proposal generation
 - deterministic and LLM validators
 - validator cardinality enforcement in pipeline execution

docs/rewrite-notes.md

@@ -65,6 +65,9 @@ Implemented:
 - Glossary-derived deterministic protected-term extraction and validator integration.
 - Explicit `--modules glossary` runtime path through runner, shared proposal generation, validators, application, reporting, and diagnostics.
 - Repeated glossary stage support with deterministic instance names (`glossary_1`, `glossary_2`), including mutable working-transcript handoff.
+- Production homophones module package with Python-aligned prompt intent and guardrails.
+- Explicit `--modules homophones` runtime path through runner, shared proposal generation, validators, application, reporting, and diagnostics.
+- Focused multi-module runtime tests for already-implemented interoperability (for example `glossary,homophones`) without claiming full default-pipeline completion.
 - Broad deterministic and CLI/subprocess test coverage for implemented phases through `go test ./...`.
 - Internal typed structured LLM contract (`StructuredLLMClient.CompleteStructured(ctx, req, out)`).
 - `internal/framework/llm` instructor-go-backed adapter with:
@@ -80,8 +83,8 @@ Implemented:
 - Generic JSON diagnostics primitives for LLM interactions (request metadata, request payload, response payload, optional error payload) with secret redaction.
 
 Not yet implemented in runtime pipeline:
-- Real correction modules for `homophones` and `spoken_word`.
+- Real correction module for `spoken_word`.
-- Domain proposal prompts for remaining real modules.
+- Domain proposal prompts for remaining real module.
 - End-to-end transcript polishing behavior with the full default module sequence.
 
 ## Completed phases
@@ -220,7 +223,7 @@ Not implemented in Phase 8 (by design):
 
 ## Remaining work plan
 
-Next recommended phase: **Phase 14 (homophones module)**.
+Next recommended phase: **Phase 15 (spoken-word module)**.
 
 ## Phase 9: Structured LLM client and scheduler infrastructure
 
@@ -391,41 +394,27 @@ Not implemented in Phase 13 (by design):
 
 ## Phase 14: Homophones module
 
-### Purpose
+Completed.
 
-Implement conservative homophone and mistranscription correction behavior.
+Implemented:
+- Production homophones module package in `internal/modules/homophones`.
+- Homophones prompt builder aligned to Python intent and constrained to conservative homophone/near-homophone/mistranscription corrections.
+- Prompt context using glossary/protected-term information (names, aliases, plurals where present) to avoid damaging known domain terms.
+- Homophones proposal generation through shared `internal/framework/proposal_generation` using `contracts.StructuredLLMClient`.
+- Scheduler-aware homophones proposal generation through existing scheduler hooks.
+- Homophones replacement policy `require_unique` (matching Python behavior).
+- Homophones validator chain using existing deterministic and LLM-backed validators.
+- Homophones confidence threshold enforcement through existing config + confidence-threshold validator behavior.
+- Protected-term guardrails remaining active for homophones via existing deterministic validators.
+- Explicit runtime support for `--modules homophones` through normalization, chunking, runner, proposal generation, validation, application, and reporting.
+- Prompt/response diagnostics artifacts for homophones proposal + validator interactions with secret redaction.
+- Module-level reports for homophones including generated proposals, validator decisions/rejections, applied changes, and application skips.
+- CLI/runtime fake-client tests for approved proposals, validator rejection, application skips, diagnostics, protected-term rejection behavior, failure/error.log behavior, and report outputs (`--report-json` and run-dir `report.json`).
+- Focused interoperability tests for already-implemented module combinations (for example `glossary,homophones`) to verify working-transcript handoff and guardrails without claiming full default-sequence parity.
 
-### Scope
+Not implemented in Phase 14 (by design):
-
+- Production `spoken_word` module.
-Implement:
+- Full default module sequence execution as a feature-complete claim.
-- `homophones` module package.
-- Homophones prompt builder ported from Python.
-- Homophones structured response model.
-- Homophones replacement policy.
-- Homophones confidence threshold handling.
-- Homophones validator chain.
-- Prompt/response diagnostics.
-- CLI support for `--modules homophones`.
-- Fake LLM tests.
-- Tests for interaction with glossary/protected terms where relevant.
-
-Do not implement:
-- Spoken-word module.
-- Default full pipeline parity claim unless spoken-word is already complete.
-
-### Expected behavior at end of phase
-
-Running `audita process ... --modules homophones` should perform real conservative homophone/mistranscription corrections using the configured LLM endpoint.
-
-### Definition of done
-
-- Homophones module runs in the production runner.
-- Homophones proposals are generated through structured LLM calls.
-- Validator chain is enforced.
-- Protected-term behavior is respected where applicable.
-- Applied/skipped homophone changes appear in reports.
-- Prompt/response diagnostics are written.
-- `go test ./...` passes without requiring external LLM credentials.
 
 ## Phase 15: Spoken-word module
 

internal/cli/run.go

@@ -460,7 +460,7 @@ func extractErrorPhase(err error) (phase string, message string) {
 
 func buildProcessReport(status string, inv processInvocation, runDir *diagnostics.RunDirectory, startedAt, completedAt time.Time, errorMessage string, errorPhase string, normalizationSummary *normalization.NormalizationSummary, chunkingSummary *chunking.Summary, runOutput *runner.RunOutput) reporting.ProcessReport {
 	report := reporting.ProcessReport{
-		Phase:          "phase13-glossary-module",
+		Phase:          "phase14-homophones-module",
 		Status:         status,
 		Operation:      "process",
 		TranscriptPath: inv.TranscriptPath,

internal/cli/run_test.go

@@ -616,8 +616,8 @@ func TestRunProcessReportJSONIncludesChunkingSummary(t *testing.T) {
 	if report.Chunking.MaxSectionTokens == 0 {
 		t.Errorf("expected max_section_tokens in report")
 	}
-	if report.Phase != "phase13-glossary-module" {
+	if report.Phase != "phase14-homophones-module" {
-		t.Errorf("expected phase 'phase13-glossary-module', got %q", report.Phase)
+		t.Errorf("expected phase 'phase14-homophones-module', got %q", report.Phase)
 	}
 }
 
@@ -924,7 +924,7 @@ func TestRunProcessProductionRegistryUnimplementedModuleFailsCleanly(t *testing.
 	exitCode := Run([]string{
 		"process", transcriptPath,
 		"--glossary", fixturePath("tiny_glossary.yaml"),
-		"--modules", "homophones",
+		"--modules", "spoken_word",
 		"--work-dir", workDir,
 		"--work-dir-retention", "always",
 		"--report-json", reportPath,
@@ -955,7 +955,7 @@ func TestRunProcessProductionRegistryUnimplementedModuleFailsCleanly(t *testing.
 }
 
 func TestRunProcessExplicitUnimplementedModulesFailClearly(t *testing.T) {
-	for _, moduleKey := range []string{"homophones", "spoken_word"} {
+	for _, moduleKey := range []string{"spoken_word"} {
 		t.Run(moduleKey, func(t *testing.T) {
 			var stdout, stderr bytes.Buffer
 			transcriptPath := writeFile(t, "transcript.json", `[
@@ -1472,6 +1472,324 @@ func TestRunProcessExplicitGlossaryMalformedLLMOutputFailsWithErrorLog(t *testin
 	}
 }
 
+func TestRunProcessExplicitHomophonesAppliesCorrectionAndReportsDiagnostics(t *testing.T) {
+	secret := "phase14-secret"
+	proposalClient := &fakeStructuredLLMClient{
+		proposalResponses: []proposal_generation.StructuredCorrectionSet{
+			{
+				Corrections: []proposal_generation.StructuredCorrectionProposal{
+					{TargetSegmentID: 1, OriginalText: "gestures", CorrectedText: "Jesters", Confidence: 0.95},
+				},
+			},
+		},
+	}
+	validationClient := &fakeStructuredLLMClient{
+		validationResponses: []validators.LLMValidationResponse{
+			{Validations: []validators.LLMValidationDecision{{CorrectionIndex: 0, Approved: true, Confidence: 0.99, Reason: "spoken plausible"}}},
+			{Validations: []validators.LLMValidationDecision{{CorrectionIndex: 0, Approved: true, Confidence: 0.99, Reason: secret}}},
+		},
+	}
+	processProposalLLMClient = proposalClient
+	processValidationLLMClient = validationClient
+	t.Setenv("AUDITA_LLM_API_KEY", secret)
+	t.Setenv("AUDITA_VALIDATION_LLM_API_KEY", secret)
+	t.Cleanup(func() {
+		processProposalLLMClient = nil
+		processValidationLLMClient = nil
+	})
+
+	var stdout, stderr bytes.Buffer
+	workDir := t.TempDir()
+	reportPath := filepath.Join(t.TempDir(), "report.json")
+	outputPath := filepath.Join(t.TempDir(), "out.json")
+	transcriptPath := writeFile(t, "transcript.json", `[
+  {"id":1,"speaker":"Alice","start":0.0,"end":1.0,"text":"There were gestures in the hall."}
+]`)
+
+	exitCode := Run([]string{
+		"process", transcriptPath,
+		"--glossary", fixturePath("tiny_glossary.yaml"),
+		"--modules", "homophones",
+		"--output", outputPath,
+		"--report-json", reportPath,
+		"--work-dir", workDir,
+		"--work-dir-retention", "always",
+	}, &stdout, &stderr)
+	if exitCode != 0 {
+		t.Fatalf("expected success, got %d stderr=%q", exitCode, stderr.String())
+	}
+	if stdout.Len() != 0 {
+		t.Fatalf("expected empty stdout with --output, got %q", stdout.String())
+	}
+
+	parsed, err := schema.ParseTranscriptJSON(readFile(t, outputPath))
+	if err != nil {
+		t.Fatalf("parse output: %v", err)
+	}
+	if parsed.Segments[0].Text != "There were Jesters in the hall." {
+		t.Fatalf("expected homophones correction applied, got %q", parsed.Segments[0].Text)
+	}
+
+	report := readProcessReport(t, reportPath)
+	if len(report.ModuleResults) != 1 || report.ModuleResults[0].ModuleKey != "homophones" {
+		t.Fatalf("expected one homophones module result, got %+v", report.ModuleResults)
+	}
+	if len(report.ModuleResults[0].AppliedChanges) != 1 {
+		t.Fatalf("expected one applied homophones change, got %+v", report.ModuleResults[0].AppliedChanges)
+	}
+	if len(report.ModuleResults[0].ValidatorDecisions) == 0 {
+		t.Fatalf("expected validator decisions in report")
+	}
+
+	runDir := onlyRunDir(t, workDir)
+	runReport := readProcessReport(t, filepath.Join(runDir, "report.json"))
+	if len(runReport.ModuleResults) != 1 || runReport.ModuleResults[0].ModuleKey != "homophones" {
+		t.Fatalf("expected homophones module results in run-dir report, got %+v", runReport.ModuleResults)
+	}
+
+	diagFiles, globErr := filepath.Glob(filepath.Join(runDir, "homophones", "*response-payload.json"))
+	if globErr != nil {
+		t.Fatalf("glob diagnostics: %v", globErr)
+	}
+	if len(diagFiles) == 0 {
+		t.Fatalf("expected homophones diagnostics payload files in %s", filepath.Join(runDir, "homophones"))
+	}
+	for _, f := range diagFiles {
+		raw := string(readFile(t, f))
+		if strings.Contains(raw, secret) {
+			t.Fatalf("secret leaked in diagnostics %q: %s", f, raw)
+		}
+	}
+}
+
+func TestRunProcessExplicitHomophonesRejectedAndApplicationSkipAreDistinct(t *testing.T) {
+	processProposalLLMClient = &fakeStructuredLLMClient{
+		proposalResponses: []proposal_generation.StructuredCorrectionSet{
+			{
+				Corrections: []proposal_generation.StructuredCorrectionProposal{
+					{TargetSegmentID: 1, OriginalText: "hall", CorrectedText: "temple", Confidence: 0.99},
+					{TargetSegmentID: 1, OriginalText: "there were gestures in the hall", CorrectedText: "there were gestures in the temple", Confidence: 0.99},
+					{TargetSegmentID: 1, OriginalText: "and there were gestures.", CorrectedText: "and there were jesters.", Confidence: 0.99},
+				},
+			},
+		},
+	}
+	processValidationLLMClient = &fakeStructuredLLMClient{
+		validationResponses: []validators.LLMValidationResponse{
+			{
+				Validations: []validators.LLMValidationDecision{
+					{CorrectionIndex: 0, Approved: true, Confidence: 0.9, Reason: "ok"},
+					{CorrectionIndex: 1, Approved: true, Confidence: 0.9, Reason: "ok"},
+					{CorrectionIndex: 2, Approved: false, Confidence: 0.9, Reason: "reject"},
+				},
+			},
+			{
+				Validations: []validators.LLMValidationDecision{
+					{CorrectionIndex: 0, Approved: true, Confidence: 0.9, Reason: "ok"},
+					{CorrectionIndex: 1, Approved: true, Confidence: 0.9, Reason: "ok"},
+				},
+			},
+		},
+	}
+	t.Cleanup(func() {
+		processProposalLLMClient = nil
+		processValidationLLMClient = nil
+	})
+
+	var stdout, stderr bytes.Buffer
+	reportPath := filepath.Join(t.TempDir(), "report.json")
+	outputPath := filepath.Join(t.TempDir(), "out.json")
+	transcriptPath := writeFile(t, "transcript.json", `[
+  {"id":1,"speaker":"Alice","start":0.0,"end":1.0,"text":"there were gestures in the hall and there were gestures."}
+]`)
+	exitCode := Run([]string{
+		"process", transcriptPath,
+		"--glossary", fixturePath("tiny_glossary.yaml"),
+		"--modules", "homophones",
+		"--output", outputPath,
+		"--report-json", reportPath,
+	}, &stdout, &stderr)
+	if exitCode != 0 {
+		t.Fatalf("expected success, got %d stderr=%q", exitCode, stderr.String())
+	}
+
+	report := readProcessReport(t, reportPath)
+	if len(report.ModuleResults) != 1 {
+		t.Fatalf("expected one module result")
+	}
+	module := report.ModuleResults[0]
+	if len(module.ValidatorRejected) != 1 {
+		t.Fatalf("expected one validator rejection, got %+v", module.ValidatorRejected)
+	}
+	if len(module.SkippedChanges) != 1 {
+		t.Fatalf("expected one application skip, got %+v", module.SkippedChanges)
+	}
+	if module.ValidatorRejected[0].ReasonCode == string(module.SkippedChanges[0].SkipReason) {
+		t.Fatalf("validator rejection and application skip should remain distinct")
+	}
+}
+
+func TestRunProcessExplicitHomophonesProtectedGlossaryTermRejected(t *testing.T) {
+	processProposalLLMClient = &fakeStructuredLLMClient{
+		proposalResponses: []proposal_generation.StructuredCorrectionSet{
+			{
+				Corrections: []proposal_generation.StructuredCorrectionProposal{
+					{TargetSegmentID: 1, OriginalText: "Audita", CorrectedText: "audita", Confidence: 0.99},
+				},
+			},
+		},
+	}
+	t.Cleanup(func() {
+		processProposalLLMClient = nil
+		processValidationLLMClient = nil
+	})
+
+	var stdout, stderr bytes.Buffer
+	reportPath := filepath.Join(t.TempDir(), "report.json")
+	outputPath := filepath.Join(t.TempDir(), "out.json")
+	transcriptPath := writeFile(t, "transcript.json", `[
+  {"id":1,"speaker":"Alice","start":0.0,"end":1.0,"text":"Audita held the line."}
+]`)
+	exitCode := Run([]string{
+		"process", transcriptPath,
+		"--glossary", fixturePath("tiny_glossary.yaml"),
+		"--modules", "homophones",
+		"--output", outputPath,
+		"--report-json", reportPath,
+	}, &stdout, &stderr)
+	if exitCode != 0 {
+		t.Fatalf("expected success, got %d stderr=%q", exitCode, stderr.String())
+	}
+
+	parsed, err := schema.ParseTranscriptJSON(readFile(t, outputPath))
+	if err != nil {
+		t.Fatalf("parse output: %v", err)
+	}
+	if parsed.Segments[0].Text != "Audita held the line." {
+		t.Fatalf("expected protected glossary term to remain unchanged, got %q", parsed.Segments[0].Text)
+	}
+
+	report := readProcessReport(t, reportPath)
+	if len(report.ModuleResults) != 1 {
+		t.Fatalf("expected one module result, got %+v", report.ModuleResults)
+	}
+	module := report.ModuleResults[0]
+	if len(module.AppliedChanges) != 0 {
+		t.Fatalf("expected no applied changes, got %+v", module.AppliedChanges)
+	}
+	if len(module.ValidatorRejected) != 1 {
+		t.Fatalf("expected one validator rejection, got %+v", module.ValidatorRejected)
+	}
+	if module.ValidatorRejected[0].ReasonCode != validators.ReasonProtectedGlossaryTerm {
+		t.Fatalf("expected protected glossary term rejection, got %+v", module.ValidatorRejected[0])
+	}
+}
+
+func TestRunProcessExplicitHomophonesMalformedLLMOutputFailsWithErrorLog(t *testing.T) {
+	processProposalLLMClient = &fakeStructuredLLMClient{err: errors.New("malformed structured output")}
+	t.Cleanup(func() { processProposalLLMClient = nil })
+
+	var stdout, stderr bytes.Buffer
+	workDir := t.TempDir()
+	reportPath := filepath.Join(t.TempDir(), "report.json")
+	transcriptPath := writeFile(t, "transcript.json", `[
+  {"id":1,"speaker":"Alice","start":0.0,"end":1.0,"text":"hello"}
+]`)
+
+	exitCode := Run([]string{
+		"process", transcriptPath,
+		"--glossary", fixturePath("tiny_glossary.yaml"),
+		"--modules", "homophones",
+		"--work-dir", workDir,
+		"--work-dir-retention", "always",
+		"--report-json", reportPath,
+	}, &stdout, &stderr)
+	if exitCode == 0 {
+		t.Fatal("expected failure")
+	}
+	if stdout.Len() != 0 {
+		t.Fatalf("expected empty stdout on failure, got %q", stdout.String())
+	}
+	if !strings.Contains(stderr.String(), "runner_execution") {
+		t.Fatalf("expected runner_execution error, got %q", stderr.String())
+	}
+	runDir := onlyRunDir(t, workDir)
+	if _, err := os.Stat(filepath.Join(runDir, "error.log")); err != nil {
+		t.Fatalf("expected error.log on failed homophones run: %v", err)
+	}
+	report := readProcessReport(t, reportPath)
+	if report.Status != "failed" || report.ErrorPhase != "runner_execution" {
+		t.Fatalf("expected failed runner_execution report, got %+v", report)
+	}
+}
+
+func TestRunProcessExplicitGlossaryThenHomophonesSeesWorkingTranscriptChanges(t *testing.T) {
+	proposalClient := &fakeStructuredLLMClient{
+		proposalResponses: []proposal_generation.StructuredCorrectionSet{
+			{
+				Corrections: []proposal_generation.StructuredCorrectionProposal{
+					{TargetSegmentID: 1, OriginalText: "gestures", CorrectedText: "Jesters", Confidence: 0.99},
+				},
+			},
+			{
+				Corrections: []proposal_generation.StructuredCorrectionProposal{
+					{TargetSegmentID: 1, OriginalText: "Jesters", CorrectedText: "Jester", Confidence: 0.99},
+				},
+			},
+		},
+	}
+	validationClient := &fakeStructuredLLMClient{
+		validationResponses: []validators.LLMValidationResponse{
+			{Validations: []validators.LLMValidationDecision{{CorrectionIndex: 0, Approved: true, Confidence: 0.99, Reason: "ok"}}},
+			{Validations: []validators.LLMValidationDecision{{CorrectionIndex: 0, Approved: true, Confidence: 0.99, Reason: "ok"}}},
+			{Validations: []validators.LLMValidationDecision{{CorrectionIndex: 0, Approved: true, Confidence: 0.99, Reason: "ok"}}},
+			{Validations: []validators.LLMValidationDecision{{CorrectionIndex: 0, Approved: true, Confidence: 0.99, Reason: "ok"}}},
+		},
+	}
+	processProposalLLMClient = proposalClient
+	processValidationLLMClient = validationClient
+	t.Cleanup(func() {
+		processProposalLLMClient = nil
+		processValidationLLMClient = nil
+	})
+
+	var stdout, stderr bytes.Buffer
+	reportPath := filepath.Join(t.TempDir(), "report.json")
+	outputPath := filepath.Join(t.TempDir(), "out.json")
+	transcriptPath := writeFile(t, "transcript.json", `[
+  {"id":1,"speaker":"Alice","start":0.0,"end":1.0,"text":"There were gestures in the hall."}
+]`)
+	exitCode := Run([]string{
+		"process", transcriptPath,
+		"--glossary", fixturePath("tiny_glossary.yaml"),
+		"--modules", "glossary,homophones",
+		"--output", outputPath,
+		"--report-json", reportPath,
+	}, &stdout, &stderr)
+	if exitCode != 0 {
+		t.Fatalf("expected success, got %d stderr=%q", exitCode, stderr.String())
+	}
+
+	parsed, err := schema.ParseTranscriptJSON(readFile(t, outputPath))
+	if err != nil {
+		t.Fatalf("parse output: %v", err)
+	}
+	if parsed.Segments[0].Text != "There were Jester in the hall." {
+		t.Fatalf("expected homophones stage to see glossary output, got %q", parsed.Segments[0].Text)
+	}
+
+	report := readProcessReport(t, reportPath)
+	if len(report.ModuleResults) != 2 {
+		t.Fatalf("expected two module results, got %+v", report.ModuleResults)
+	}
+	if report.ModuleResults[0].ModuleKey != "glossary" || report.ModuleResults[1].ModuleKey != "homophones" {
+		t.Fatalf("expected glossary then homophones results, got %+v", report.ModuleResults)
+	}
+	if len(report.ModuleResults[0].AppliedChanges) != 1 || len(report.ModuleResults[1].AppliedChanges) != 1 {
+		t.Fatalf("expected one applied change per module, got %+v", report.ModuleResults)
+	}
+}
+
 func TestRunProcessChunkingSummaryArtifactWritten(t *testing.T) {
 	var stdout bytes.Buffer
 	var stderr bytes.Buffer

internal/core/reporting/report_test.go

@@ -11,7 +11,7 @@ import (
 func TestProcessReportModuleResultsJSONSuccessAndSkipped(t *testing.T) {
 	now := time.Now().UTC()
 	report := ProcessReport{
-		Phase:  "phase13-glossary-module",
+		Phase:  "phase14-homophones-module",
 		Status: "success",
 		ModuleResults: []ModuleReport{
 			{
@@ -80,7 +80,7 @@ func TestProcessReportModuleResultsJSONSuccessAndSkipped(t *testing.T) {
 func TestProcessReportModuleResultsJSONFailedModule(t *testing.T) {
 	now := time.Now().UTC()
 	report := ProcessReport{
-		Phase:  "phase13-glossary-module",
+		Phase:  "phase14-homophones-module",
 		Status: "failed",
 		ModuleResults: []ModuleReport{
 			{

internal/framework/modules/registry.go

@@ -10,6 +10,7 @@ import (
 	"gitea.maximumdirect.net/eric/audita/internal/framework/contracts"
 	glossarymodule "gitea.maximumdirect.net/eric/audita/internal/modules/glossary"
 	grammarmodule "gitea.maximumdirect.net/eric/audita/internal/modules/grammar"
+	homophonesmodule "gitea.maximumdirect.net/eric/audita/internal/modules/homophones"
 )
 
 const (
@@ -72,6 +73,7 @@ func NewFactory(deps Dependencies) *Factory {
 		constructors: make(map[string]Constructor, len(knownModuleKeys)),
 	}
 	_ = factory.RegisterConstructor(ModuleKeyGlossary, constructGlossaryModule)
+	_ = factory.RegisterConstructor(ModuleKeyHomophones, constructHomophonesModule)
 	_ = factory.RegisterConstructor(ModuleKeyGrammar, constructGrammarModule)
 	return factory
 }
@@ -104,6 +106,12 @@ func constructGlossaryModule(ctx context.Context, req ConstructRequest) (contrac
 	return glossarymodule.New()
 }
 
+func constructHomophonesModule(ctx context.Context, req ConstructRequest) (contracts.TranscriptModule, error) {
+	_ = ctx
+	_ = req
+	return homophonesmodule.New()
+}
+
 // ModuleForSpec resolves one configured run spec into a module instance.
 func (f *Factory) ModuleForSpec(spec contracts.ModuleRunSpec) (contracts.TranscriptModule, error) {
 	if f == nil {

internal/framework/modules/registry_test.go

@@ -66,7 +66,7 @@ func TestUnsupportedUnknownModuleKeyFailsCleanly(t *testing.T) {
 
 func TestRecognizedButUnimplementedModuleKeyFailsCleanly(t *testing.T) {
 	factory := NewFactory(Dependencies{})
-	for _, key := range []string{ModuleKeyHomophones, ModuleKeySpokenWord} {
+	for _, key := range []string{ModuleKeySpokenWord} {
 		t.Run(key, func(t *testing.T) {
 			_, err := factory.ModuleForSpec(contracts.ModuleRunSpec{ModuleKey: key, InstanceName: key})
 			if err == nil {
@@ -106,6 +106,17 @@ func TestGlossaryIsRegisteredAndConstructibleByDefault(t *testing.T) {
 	}
 }
 
+func TestHomophonesIsRegisteredAndConstructibleByDefault(t *testing.T) {
+	factory := NewFactory(Dependencies{})
+	module, err := factory.ModuleForSpec(contracts.ModuleRunSpec{ModuleKey: ModuleKeyHomophones, InstanceName: ModuleKeyHomophones})
+	if err != nil {
+		t.Fatalf("ModuleForSpec error: %v", err)
+	}
+	if module.Key() != ModuleKeyHomophones {
+		t.Fatalf("expected homophones module key, got %q", module.Key())
+	}
+}
+
 func TestRegisterConstructorAndConstruct(t *testing.T) {
 	factory := NewFactory(Dependencies{})
 	if err := factory.RegisterConstructor(ModuleKeyGlossary, func(ctx context.Context, req ConstructRequest) (contracts.TranscriptModule, error) {

internal/modules/homophones/module.go

@@ -0,0 +1,76 @@
+package homophones
+
+import (
+	"context"
+	"fmt"
+
+	"gitea.maximumdirect.net/eric/audita/internal/framework/contracts"
+	"gitea.maximumdirect.net/eric/audita/internal/framework/proposal_generation"
+	"gitea.maximumdirect.net/eric/audita/internal/framework/proposals"
+	"gitea.maximumdirect.net/eric/audita/internal/framework/validators"
+)
+
+type Module struct {
+	validators []contracts.Validator
+}
+
+func New() (*Module, error) {
+	spokenForm, err := validators.NewLLMBackedValidator("spoken_form_plausibility_review", validators.LLMValidatorTypeSpokenFormPlausibility, "")
+	if err != nil {
+		return nil, err
+	}
+	meaningReversal, err := validators.NewLLMBackedValidator("meaning_reversal_review", validators.LLMValidatorTypeMeaningReversal, "")
+	if err != nil {
+		return nil, err
+	}
+
+	return &Module{
+		validators: []contracts.Validator{
+			validators.NoEffectValidator{},
+			validators.OriginalTextPresenceValidator{},
+			validators.ConfidenceThresholdValidator{},
+			validators.ProtectedGlossaryTermValidator{},
+			validators.NonEmptyCorrectionValidator{},
+			spokenForm,
+			meaningReversal,
+		},
+	}, nil
+}
+
+func (m *Module) Key() string { return "homophones" }
+
+func (m *Module) ReplacementPolicy() proposals.ReplacementPolicy {
+	// Python homophones module uses require_unique for conservative single-span replacement.
+	return proposals.ReplacementPolicyRequireUnique
+}
+
+func (m *Module) Validators() []contracts.Validator {
+	return append([]contracts.Validator(nil), m.validators...)
+}
+
+func (m *Module) Propose(ctx context.Context, req contracts.ProposalRequest) ([]proposals.CorrectionProposal, error) {
+	messages, err := BuildProposalMessages(req.WorkingTranscript, req.Glossary)
+	if err != nil {
+		return nil, err
+	}
+
+	generated, err := proposal_generation.GenerateCandidates(ctx, proposal_generation.Request{
+		ModuleKey:         req.RunSpec.ModuleKey,
+		ModuleInstance:    req.RunSpec.InstanceName,
+		ReplacementPolicy: req.RunSpec.ReplacementPolicy,
+		WorkingTranscript: req.WorkingTranscript,
+		Section:           req.Section,
+		Glossary:          req.Glossary,
+		Config:            req.Config,
+		Messages:          messages,
+		StageName:         fmt.Sprintf("%s:proposal", req.RunSpec.InstanceName),
+		StartIndex:        0,
+		LLMClient:         req.LLMClient,
+		Scheduler:         req.LLMScheduler,
+		DiagnosticsDir:    req.DiagnosticsDir,
+	})
+	if err != nil {
+		return nil, err
+	}
+	return generated.Corrections, nil
+}

internal/modules/homophones/module_test.go

@@ -0,0 +1,296 @@
+package homophones
+
+import (
+	"context"
+	"errors"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"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/proposal_generation"
+	"gitea.maximumdirect.net/eric/audita/internal/framework/proposals"
+	"gitea.maximumdirect.net/eric/audita/internal/framework/validators"
+)
+
+type fakeLLMClient struct {
+	responses []proposal_generation.StructuredCorrectionSet
+	err       error
+	calls     []contracts.StructuredCompletionRequest
+}
+
+func (f *fakeLLMClient) CompleteStructured(ctx context.Context, req contracts.StructuredCompletionRequest, out any) (contracts.StructuredCompletionResponse, error) {
+	_ = ctx
+	f.calls = append(f.calls, req)
+	if f.err != nil {
+		return contracts.StructuredCompletionResponse{}, f.err
+	}
+	target, ok := out.(*proposal_generation.StructuredCorrectionSet)
+	if !ok {
+		return contracts.StructuredCompletionResponse{}, errors.New("unexpected output type")
+	}
+	if len(f.responses) == 0 {
+		return contracts.StructuredCompletionResponse{}, errors.New("unexpected call")
+	}
+	*target = f.responses[0]
+	f.responses = f.responses[1:]
+	return contracts.StructuredCompletionResponse{}, nil
+}
+
+type countingScheduler struct {
+	runs int
+}
+
+func (s *countingScheduler) Run(ctx context.Context, fn func(context.Context) error) error {
+	s.runs++
+	return fn(ctx)
+}
+
+func tinyTranscript() *schema.Transcript {
+	return &schema.Transcript{Segments: []schema.Segment{
+		{ID: 1, Speaker: "A", Start: 0, End: 1, Text: "There were gestures in the hall.", Categories: []string{"session"}},
+	}}
+}
+
+func tinyGlossary() *schema.Glossary {
+	return &schema.Glossary{Entries: []schema.GlossaryEntry{
+		{Name: "Jesters", Aliases: []string{"Jester"}, Category: "faction", Summary: "Guild members", Plural: "Jesters"},
+	}}
+}
+
+func TestBuildProposalMessagesContainsContextAndConservativeConstraints(t *testing.T) {
+	msgs, err := BuildProposalMessages(tinyTranscript(), tinyGlossary())
+	if err != nil {
+		t.Fatalf("BuildProposalMessages error: %v", err)
+	}
+	if len(msgs) != 2 {
+		t.Fatalf("expected 2 messages, got %d", len(msgs))
+	}
+	combined := msgs[0].Content + "\n" + msgs[1].Content
+	for _, want := range []string{
+		"Protected glossary/context:",
+		"Transcript section:",
+		`"Aliases":`,
+		`"Category": "faction"`,
+		`"Summary": "Guild members"`,
+		`"Plural": "Jesters"`,
+		"homophone correction assistant",
+		"mistranscriptions of spoken English",
+		"Do not add or remove punctuation",
+		"Do not return speaker, start, or end fields",
+		`"id": 1`,
+	} {
+		if !strings.Contains(combined, want) {
+			t.Fatalf("expected prompt to contain %q", want)
+		}
+	}
+	for _, forbidden := range []string{
+		"summarize the transcript",
+		"style rewrite",
+		"grammar cleanup",
+		"punctuation cleanup",
+	} {
+		if strings.Contains(strings.ToLower(combined), forbidden) {
+			t.Fatalf("prompt should not invite %q", forbidden)
+		}
+	}
+}
+
+func TestHomophonesModuleReplacementPolicy(t *testing.T) {
+	m, err := New()
+	if err != nil {
+		t.Fatalf("New error: %v", err)
+	}
+	if m.ReplacementPolicy() != proposals.ReplacementPolicyRequireUnique {
+		t.Fatalf("unexpected replacement policy: %q", m.ReplacementPolicy())
+	}
+}
+
+func TestHomophonesModuleValidatorChain(t *testing.T) {
+	m, err := New()
+	if err != nil {
+		t.Fatalf("New error: %v", err)
+	}
+	got := make([]string, 0, len(m.Validators()))
+	for _, v := range m.Validators() {
+		got = append(got, v.Name())
+	}
+	want := []string{
+		"no_effect",
+		"original_text_presence",
+		"confidence_threshold",
+		"protected_glossary_terms",
+		"non_empty_correction",
+		"spoken_form_plausibility_review",
+		"meaning_reversal_review",
+	}
+	if strings.Join(got, ",") != strings.Join(want, ",") {
+		t.Fatalf("unexpected validator chain\n got: %v\nwant: %v", got, want)
+	}
+}
+
+func TestHomophonesModuleProposeMapsCorrectionsAndWritesDiagnostics(t *testing.T) {
+	secret := "homophones-secret"
+	client := &fakeLLMClient{
+		responses: []proposal_generation.StructuredCorrectionSet{
+			{
+				Corrections: []proposal_generation.StructuredCorrectionProposal{
+					{TargetSegmentID: 1, OriginalText: "gestures", CorrectedText: "Jesters", Confidence: 0.94},
+				},
+			},
+		},
+	}
+	scheduler := &countingScheduler{}
+	cfg := config.Default()
+	cfg.PrimaryLLM.APIKey = secret
+	m, err := New()
+	if err != nil {
+		t.Fatalf("New error: %v", err)
+	}
+	diagDir := t.TempDir()
+	out, err := m.Propose(context.Background(), contracts.ProposalRequest{
+		ExecutionContext: contracts.ExecutionContext{
+			Config:            &cfg,
+			WorkingTranscript: tinyTranscript(),
+			Glossary:          tinyGlossary(),
+			DiagnosticsDir:    diagDir,
+		},
+		RunSpec: contracts.ModuleRunSpec{
+			ModuleKey:         "homophones",
+			InstanceName:      "homophones",
+			ReplacementPolicy: proposals.ReplacementPolicyRequireUnique,
+		},
+		LLMClient:    client,
+		LLMScheduler: scheduler,
+	})
+	if err != nil {
+		t.Fatalf("Propose error: %v", err)
+	}
+	if scheduler.runs != 1 {
+		t.Fatalf("expected scheduler run count 1, got %d", scheduler.runs)
+	}
+	if len(client.calls) != 1 || client.calls[0].StageName != "homophones:proposal" {
+		t.Fatalf("expected one homophones:proposal call, got %+v", client.calls)
+	}
+	if len(out) != 1 || out[0].CorrectedText != "Jesters" {
+		t.Fatalf("unexpected proposals: %+v", out)
+	}
+	diagFiles, globErr := filepath.Glob(filepath.Join(diagDir, "homophones", "*proposal*response-payload.json"))
+	if globErr != nil {
+		t.Fatalf("glob error: %v", globErr)
+	}
+	if len(diagFiles) == 0 {
+		t.Fatalf("expected diagnostics response payload under %s", filepath.Join(diagDir, "homophones"))
+	}
+	for _, f := range diagFiles {
+		raw, readErr := os.ReadFile(f)
+		if readErr != nil {
+			t.Fatalf("read diag %q: %v", f, readErr)
+		}
+		if strings.Contains(string(raw), secret) {
+			t.Fatalf("secret leaked in diagnostics: %s", string(raw))
+		}
+	}
+}
+
+func TestHomophonesConfidenceThresholdIsUsed(t *testing.T) {
+	module, err := New()
+	if err != nil {
+		t.Fatalf("New error: %v", err)
+	}
+	var thresholdValidator contracts.Validator
+	for _, v := range module.Validators() {
+		if v.Name() == "confidence_threshold" {
+			thresholdValidator = v
+			break
+		}
+	}
+	if thresholdValidator == nil {
+		t.Fatal("expected confidence_threshold validator")
+	}
+
+	cfg := config.Default()
+	cfg.Thresholds.Homophones = 0.95
+	result, err := thresholdValidator.Validate(context.Background(), validators.Request{
+		Config:    &cfg,
+		ModuleKey: "homophones",
+		CandidateProposal: []proposals.EnrichedCorrectionProposal{
+			{
+				CorrectionProposal: proposals.CorrectionProposal{
+					TargetSegmentID: 1,
+					OriginalText:    "gestures",
+					CorrectedText:   "Jesters",
+					Confidence:      0.94,
+				},
+				ProposalMetadata: proposals.ProposalMetadata{ProposalIndex: 0, ModuleKey: "homophones", ModuleInstance: "homophones"},
+			},
+		},
+	})
+	if err != nil {
+		t.Fatalf("validate error: %v", err)
+	}
+	if len(result.Decisions) != 1 {
+		t.Fatalf("expected one decision, got %+v", result.Decisions)
+	}
+	if result.Decisions[0].Approved {
+		t.Fatalf("expected rejection under homophones threshold, got %+v", result.Decisions[0])
+	}
+	if result.Decisions[0].ReasonCode != validators.ReasonLowConfidence {
+		t.Fatalf("expected low confidence reason, got %+v", result.Decisions[0])
+	}
+}
+
+func TestHomophonesProtectedGlossaryTermBehaviorRejectsUnsafeCorrection(t *testing.T) {
+	module, err := New()
+	if err != nil {
+		t.Fatalf("New error: %v", err)
+	}
+	var protectedValidator contracts.Validator
+	for _, v := range module.Validators() {
+		if v.Name() == "protected_glossary_terms" {
+			protectedValidator = v
+			break
+		}
+	}
+	if protectedValidator == nil {
+		t.Fatal("expected protected_glossary_terms validator")
+	}
+
+	working := &schema.Transcript{Segments: []schema.Segment{
+		{ID: 1, Speaker: "A", Start: 0, End: 1, Text: "Jesters are here."},
+	}}
+	glossary := &schema.Glossary{Entries: []schema.GlossaryEntry{
+		{Name: "Jesters", Category: "faction", Summary: "Faction"},
+	}}
+	result, err := protectedValidator.Validate(context.Background(), validators.Request{
+		WorkingTranscript: working,
+		Glossary:          glossary,
+		ModuleKey:         "homophones",
+		CandidateProposal: []proposals.EnrichedCorrectionProposal{
+			{
+				CorrectionProposal: proposals.CorrectionProposal{
+					TargetSegmentID: 1,
+					OriginalText:    "Jesters",
+					CorrectedText:   "gestures",
+					Confidence:      0.99,
+				},
+				ProposalMetadata: proposals.ProposalMetadata{ProposalIndex: 0, ModuleKey: "homophones", ModuleInstance: "homophones"},
+			},
+		},
+	})
+	if err != nil {
+		t.Fatalf("validate error: %v", err)
+	}
+	if len(result.Decisions) != 1 {
+		t.Fatalf("expected one decision, got %+v", result.Decisions)
+	}
+	if result.Decisions[0].Approved {
+		t.Fatalf("expected protected-term rejection, got %+v", result.Decisions[0])
+	}
+	if result.Decisions[0].ReasonCode != validators.ReasonProtectedGlossaryTerm {
+		t.Fatalf("expected protected glossary-term reason, got %+v", result.Decisions[0])
+	}
+}

internal/modules/homophones/prompt.go

@@ -0,0 +1,83 @@
+package homophones
+
+import (
+	"encoding/json"
+	"fmt"
+
+	"gitea.maximumdirect.net/eric/audita/internal/core/schema"
+	"gitea.maximumdirect.net/eric/audita/internal/framework/contracts"
+)
+
+type promptSegment struct {
+	ID         int      `json:"id"`
+	Speaker    string   `json:"speaker"`
+	Start      float64  `json:"start"`
+	End        float64  `json:"end"`
+	Text       string   `json:"text"`
+	Categories []string `json:"categories,omitempty"`
+}
+
+type promptTranscriptSection struct {
+	SectionIndex int             `json:"section_index"`
+	Segments     []promptSegment `json:"segments"`
+}
+
+// BuildProposalMessages mirrors the Python homophones-module prompt intent:
+// conservative homophone and mistranscription correction only.
+func BuildProposalMessages(transcript *schema.Transcript, glossary *schema.Glossary) ([]contracts.LLMMessage, error) {
+	glossaryJSON, err := json.MarshalIndent(glossary, "", "  ")
+	if err != nil {
+		return nil, fmt.Errorf("marshal glossary prompt context: %w", err)
+	}
+
+	sectionPayload := promptTranscriptSection{
+		SectionIndex: 0,
+		Segments:     make([]promptSegment, 0),
+	}
+	if transcript != nil {
+		for _, s := range transcript.Segments {
+			sectionPayload.Segments = append(sectionPayload.Segments, promptSegment{
+				ID:         s.ID,
+				Speaker:    s.Speaker,
+				Start:      s.Start,
+				End:        s.End,
+				Text:       s.Text,
+				Categories: append([]string(nil), s.Categories...),
+			})
+		}
+	}
+	sectionJSON, err := json.MarshalIndent(sectionPayload, "", "  ")
+	if err != nil {
+		return nil, fmt.Errorf("marshal transcript prompt context: %w", err)
+	}
+
+	system := "You are Audita, a conservative homophone correction assistant. Identify only transcript changes that plausibly reflect homophones, phonetic similarity, or common mistranscriptions of spoken English. Do not make punctuation, capitalization, spacing, filler-word, repetition, style, or grammar edits. Do not paraphrase, summarize, or rewrite content."
+	user := "Review this transcript section and return only homophone or spoken-form corrections that should be applied.\n\n" +
+		"Rules:\n" +
+		"- Approve only corrections where the original text is plausibly a mistaken homophone, phonetic rendering, or mistranscription of what was likely spoken.\n" +
+		"- Allow examples such as changing \"dam\" to \"damn\", \"rank\" to \"Hrank\", or \"gestures\" to \"Jesters\" when local context supports the correction.\n" +
+		"- Reject unrelated substitutions like changing \"Lyra\" to \"Jesters\".\n" +
+		"- Reject antonyms or reversals such as changing \"visible\" to \"invisible\".\n" +
+		"- Do not add or remove punctuation, alter capitalization only, normalize spacing, remove filler words, collapse repetitions, or make general readability edits.\n" +
+		"- Treat glossary names and aliases as protected spellings and context.\n" +
+		"- You may correct toward glossary names, aliases, or their plural forms when the correction is acoustically plausible and supported by local context.\n" +
+		"- Do not replace, Anglicize, normalize, lowercase, or otherwise alter protected glossary names or aliases that already appear correctly in the transcript.\n" +
+		"- Preserve canonical glossary capitalization for protected names and aliases, even if they look unusual.\n" +
+		"- If a segment includes categories, treat them as additional transcript context.\n" +
+		"- Use the exact id from the input segment.\n" +
+		"- For returned corrections, original_text must be only the exact text span that needs replacement, not the full segment text unless the whole segment is the replacement span.\n" +
+		"- Choose an original_text span that appears exactly once in the current segment text.\n" +
+		"- 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.\n" +
+		"- Each returned correction must contain only id, original_text, corrected_text, and confidence.\n" +
+		"- Do not return corrections where original_text and corrected_text are identical.\n" +
+		"- Do not return speaker, start, or end fields.\n" +
+		"- Return only changed segments; do not return entries for unchanged segments.\n" +
+		"- confidence must be between 0.0 and 1.0.\n" +
+		"- If no corrections are needed, return an empty corrections list.\n\n" +
+		fmt.Sprintf("Protected glossary/context:\n%s\n\nTranscript section:\n%s", string(glossaryJSON), string(sectionJSON))
+
+	return []contracts.LLMMessage{
+		{Role: "system", Content: system},
+		{Role: "user", Content: user},
+	}, nil
+}