Improve semantic reconciliation retries

docs/adr/0016-feedback-aware-module-requested-retries.md

@@ -0,0 +1,76 @@
+# ADR-0016: Use feedback-aware module-requested retries
+
+**Status:** Accepted
+**Date:** 2026-08-28
+
+## Context
+
+An LLM-backed module can discover that a structurally valid model response is
+unsafe while translating it into a typed candidate, before the ordinary
+validator chain runs. Semantic registry reconciliation is the first such case:
+the shared core can identify invalid duplicate-group proposals, and typed
+application can reject a domain-incompatible group while preserving a safe
+partial result. Repeating the original request without the rejected response or
+corrective explanation gives the model no information with which to improve.
+
+The existing feedback-aware validation mechanism already establishes the safe
+correction protocol, but module-owned retry directives also carry internal
+reason codes, operator messages, and fallback diagnostics. Those values are not
+model instructions. Some module retry conditions, including exhausted
+structured-output repair, also have no valid exact response to append.
+
+## Decision
+
+An LLM-backed normalizer may attach optional, bounded model-facing correction
+guidance to a module-requested retry. Guidance is a separate contract field; the
+framework never derives it from a reason code, operator message, diagnostic, or
+error. A feedback-capable directive must include the exact model candidate that
+controlled the safe fallback and must use `single_response_v1`.
+
+The artifact-neutral producer-attempt state machine constructs the correction
+from that exact latest response and the supplied guidance. The producer rebuilds
+its complete ordinary request and appends the response as one assistant message
+and the guidance as one user message. Earlier attempts do not accumulate, and
+the attempt consumes the same configured stage retry budget as operational,
+structural, validation, and feedback-free module retries.
+
+A module retry without guidance remains valid and starts a fresh attempt. This
+is the required behavior when no valid model candidate exists, including after
+structured-output repair is exhausted. If feedback is supplied without a valid
+supported candidate, the framework fails the module contract rather than
+retrying blindly or inventing guidance.
+
+After retry exhaustion, the normalizer's safe fallback continues through its
+validator chain. Safe semantic groups may remain applied, unsafe groups remain
+separate, and bounded fallback diagnostics may explain the process degradation.
+Exact responses and correction text remain attempt-local and are excluded from
+ordinary errors, warnings, manifests, receipts, caches, and checkpoints.
+
+This decision extends, rather than supersedes,
+[ADR-0014](0014-feedback-aware-validation-retries.md): both correction sources
+use the same transport-neutral payload, replacement-request semantics, outer
+retry budget, and sensitive-data boundary.
+
+## Alternatives considered
+
+- Continue blind module retries. This preserves a smaller contract but wastes
+  the module's deterministic diagnosis and commonly repeats the same defect.
+- Convert module safety checks into validators. Typed reconciliation must apply
+  only safe proposal groups and retain a fallback before validation; moving
+  artifact-owned translation and application policy into validators would blur
+  stage ownership.
+- Copy the retry reason or operator message into the model request. Those values
+  are written for provenance and humans, can contain opaque internal labels,
+  and do not reliably describe a correct replacement.
+- Require feedback for every module retry. Structural failures may have no
+  valid exact candidate, so this would either prevent useful fresh retries or
+  fabricate prior-response material.
+
+## Consequences
+
+The normalize retry contract and generic producer-attempt directive gain an
+optional correction-guidance field and candidate-pairing validation. Modules
+that use it must provide semantically meaningful bounded prose and an exact
+candidate. Registry reconciliation maintains separate operator and model
+renderers, and policy fingerprints change so checkpoints created under blind
+retry behavior are not reused.

docs/internal/dnd.md

@@ -29,13 +29,17 @@ The D&D registrar registers the family’s artifact codecs, extractors, typed
 append-order mergers, normalizers, validators, prompt assets, fallback LLM
 profile asset, and default validator chains. Each extractor and normalizer has
 a stable module spec, explicit execution class, strict option decoding, and a
-typed builder. Scene chunking, every extractor, and NPC, location, and item-registry
-normalization are registered as `llm_backed`; the remaining current D&D mergers
-and normalizers are `deterministic`. The metadata is available to catalog inspection and
-resolved-pipeline debug data and determines which selected bindings inherit the
-pipeline profile. The registry normalizers use `single_response_v1`, forwarding
-corrections to their reconciliation completion and retaining the accepted raw
-proposal only as an owned model candidate. Configuration remains the canonical owner of the exact keys,
+typed builder. Scene chunking, every extractor, and NPC, location, and
+item-registry normalization are registered as `llm_backed`; the remaining
+current D&D mergers and normalizers are `deterministic`. The metadata is
+available to catalog inspection and resolved-pipeline debug data and determines
+which selected bindings inherit the pipeline profile. The registry normalizers
+use `single_response_v1`, forwarding corrections to their reconciliation
+completion and retaining the accepted raw proposal only as an owned model
+candidate. When deterministic proposal safety or typed application rejects a
+group, they provide separate model-facing prose for a corrective module retry;
+internal issue categories, reason codes, and operator messages remain
+diagnostic-only. Configuration remains the canonical owner of the exact keys,
 profile precedence, and validator order.
 
 Private structured-LLM response schemas are deliberately minimal. They reject
@@ -117,6 +121,17 @@ presentation, and final ephemeral generic transcript windows. These orders and
 cache controls are prompt behavior; change them only through the owning
 manifest and prompt declaration.
 
+NPC, item, and location registry reconciliation translate shared proposal
+safety categories into bounded prose that references only the response-local
+duplicate-group ordinals and candidate handles. Item reconciliation appends its
+typed rule that currency may be consolidated only with aliases of the same
+denomination and never with non-currency items. The next normalize attempt
+receives that prose with the exact defective proposal under the shared
+replacement-request protocol. Structurally invalid output has no valid proposal
+candidate and receives a fresh feedback-free attempt instead. If the stage
+budget is exhausted, safe groups stay applied, unsafe groups stay separate, and
+one fallback warning summarizes the final defect without raw model content.
+
 ## Evidence, Candidates, And Normalization
 
 The current transcript is the only durable evidence source. Extractors assign

docs/internal/llm.md

@@ -247,8 +247,12 @@ redaction boundary.
 Structural repair does not replace pipeline retry behavior: a binding's
 configured retry count reruns its complete stage attempt after an operational
 or structural error, module-requested retry, or actionable semantic rejection.
-The pipeline owns attempt lifecycle, validation chains, and retry diagnostics;
-see [Pipeline Internals](pipeline.md#validation-retries-and-output) and the
+An actionable module-requested retry and a validator rejection both use the
+same correction payload when the producer exposes an exact latest response;
+feedback-free module retries reconstruct the ordinary request without appended
+messages. The pipeline owns attempt lifecycle, validation chains, and retry
+diagnostics; see
+[Pipeline Internals](pipeline.md#validation-retries-and-output) and the
 [binding reference](../config.md#module-bindings-and-validators).
 
 ## Timeout Ownership

docs/internal/modules.md

@@ -115,6 +115,16 @@ eligible producers. Deterministic skip, limit, and fallback outcomes carry no
 model candidate, so a later rejection applies terminal policy without spending
 an ineffective semantic retry.
 
+When proposal assessment or typed group application rejects a structurally
+valid group, the normalizer may return its safe partial value with a
+module-requested retry. A feedback-capable directive supplies bounded
+model-facing correction guidance separately from operator diagnostics and
+retains the exact proposal response as its candidate. The shared stage retry
+mechanism appends that response and guidance to a fresh complete request. A
+structurally invalid completion has no valid candidate and therefore requests a
+feedback-free fresh attempt. On exhaustion, only the final safe fallback and
+its bounded process diagnostic advance to validation.
+
 The core supplies a conservative generic prompt and the single private
 response schema. A domain prompt may substitute its semantic instructions but
 mounts the core-owned protocol and candidate/transcript presentation assets.

docs/internal/pipeline.md

@@ -154,18 +154,25 @@ complete validation chain. It preserves terminal diagnostics only from the final
 or rejected attempt, plus one fixed validation-incomplete warning per validator whose execution
 budget was exhausted under `warn_continue`. Cancellation stops retries.
 Normalizer-specific retry directives consume this same budget and validate any
-final safe fallback through the normalizer chain.
+final safe fallback through the normalizer chain. A directive may carry bounded
+correction guidance only when it also exposes the exact latest
+`single_response_v1` candidate. The state machine then uses the same replacement
+request shape as validator correction. A directive without guidance clears any
+prior correction and starts a fresh attempt, which preserves structural retry
+behavior when no valid response exists.
 
 The artifact-neutral producer-attempt state machine owns that shared budget,
 attempt provenance, semantic-correction material, and terminal-policy
 selection. It accepts producer and complete-validation closures, so artifact
 materialization, cache handling, checkpoints, and debug output stay at the
 operation boundary. It distinguishes operational, structural, module-requested,
-and semantic retries. A semantic retry is available only for a valid latest
-`single_response_v1` candidate; a deterministic or no-model rejection instead
-settles the semantic policy immediately. Structural-output errors alone use the
-structural policy, and validation failure without rejection settles the
-validator-failure policy without regenerating the producer.
+and validator-semantic retries. Model feedback from either semantic source is
+available only for a valid latest `single_response_v1` candidate. A
+deterministic or no-model validator rejection instead settles the semantic
+policy immediately, while a feedback-free module directive remains an ordinary
+fresh retry. Structural-output errors alone use the structural policy, and
+validation failure without rejection settles the validator-failure policy
+without regenerating the producer.
 
 Chunk planning uses this state machine for generated plans. A rejected or
 validation-incomplete automatic cache hit is not model material and therefore

docs/policy/architecture.md

@@ -201,6 +201,13 @@ they are not model instructions. The framework constructs model-facing retry
 text only from the semantic guidance and fails the contract rather than
 inventing or truncating missing guidance.
 
+An LLM-backed module may also request a feedback-aware retry when its own
+deterministic translation or typed safety policy rejects a structurally valid
+model response. It must supply model-facing guidance separately from its
+reason code, operator message, and fallback diagnostics, together with the
+exact `single_response_v1` candidate. A feedback-free module retry remains
+valid when no exact candidate exists.
+
 Default validator chains are production composition policy and are registered
 centrally by stage and module. Configuration may replace a stage-local default,
 including with an explicitly empty chain. Configured validator order is
@@ -222,12 +229,14 @@ the two budgets must remain separate.
 
 An LLM-backed producer can participate in semantic correction only when it
 declares `single_response_v1` and returns the exact one response that directly
-controlled its candidate. On an actionable rejection, the framework rebuilds
-the ordinary request and appends only the latest defective response as an
-`assistant` message plus one aggregated `user` correction message. This is a
-fresh replacement request, not a growing conversation. The retry budgets,
-terminal policy, and sensitive-data rationale are recorded in
-[ADR-0014](../adr/0014-feedback-aware-validation-retries.md).
+controlled its candidate. On an actionable validator rejection or
+feedback-capable module retry, the framework rebuilds the ordinary request and
+appends only the latest defective response as an `assistant` message plus one
+aggregated `user` correction message. This is a fresh replacement request, not
+a growing conversation. The retry budgets, terminal policy, and sensitive-data
+rationale are recorded in
+[ADR-0014](../adr/0014-feedback-aware-validation-retries.md) and
+[ADR-0016](../adr/0016-feedback-aware-module-requested-retries.md).
 
 When a model selects an application entity, callers must supply a contextual
 selection and deterministically attach the opaque application identity whenever

docs/roadmap/future.md

@@ -15,7 +15,7 @@ diagnostics. The remaining near-term work applies those completed foundations
 to domain review and empirical evaluation.
 
 The active D&D reliability work is defined by
-[D&D Item-Occurrence Holder Reliability](item-occurrence-holder-reliability.md).
+[Feedback-Aware Semantic-Reconciliation Retries](semantic-reconciliation-retry-feedback.md).
 
 ## Near-Term D&D Pipeline
 

docs/roadmap/semantic-reconciliation-retry-feedback.md

@@ -0,0 +1,117 @@
+# Feedback-Aware Semantic-Reconciliation Retries
+
+## Purpose
+
+Improve semantic registry reconciliation for smaller models by turning the
+existing module-requested retry into a bounded corrective exchange. Preserve
+the current deterministic safety checks, shared retry budget, and safe fallback
+when correction is unsuccessful.
+
+## Problem
+
+NPC, item, and location registry normalizers can reject unsafe duplicate-group
+proposals and retain a safe partial result. They currently ask the pipeline for
+another attempt, but the next attempt receives the original prompt without the
+defective response or an explanation of the defect. The normalizers also build
+their retry messages from internal issue-category strings, which are suitable
+for diagnostics but not useful model instructions. If all attempts are
+exhausted, the warning reports the omitted-group count without explaining the
+final defect or confirming that affected candidates were preserved separately.
+
+## Target State
+
+- A normalizer retry directive may carry model-facing correction guidance
+  separately from its stable reason code, operator message, and fallback
+  diagnostics.
+- When the directive includes correction guidance and a valid
+  `single_response_v1` model candidate, the shared producer-attempt state
+  machine reconstructs the complete original prompt and appends the exact
+  defective response as an assistant message followed by one user correction
+  message. Earlier failed attempts do not accumulate.
+- This correction consumes the existing normalize-stage retry budget. It does
+  not add a module-local loop or alter PromptKit's independent structured-output
+  repair budget.
+- A retry caused by structurally invalid output remains a fresh stage attempt
+  without semantic feedback when no valid model candidate is available.
+- Internal issue categories remain available for operator and debug
+  diagnostics, but are never sent to a model. A shared renderer translates
+  every known category into concise prose that explains the defect and the
+  required valid shape.
+- Item reconciliation appends its domain rule when a proposal attempts to
+  combine different currency denominations or currency with a non-currency
+  item. NPC and location reconciliation use only the shared protocol guidance.
+- On final exhaustion, safe proposal groups remain applied, unsafe groups
+  remain separate, and one process warning gives a bounded summary of the final
+  defect and explicitly states that the affected candidates were preserved
+  separately.
+
+## Required Work
+
+1. Extend `contracts.NormalizeRetry` and its clone, validation, and debug paths
+   with an optional, bounded model-facing correction-guidance field. Keep
+   reason codes and operator messages out of model input. Require guidance to
+   be paired with a valid `single_response_v1` model candidate; preserve the
+   existing feedback-free directive for structural retry cases.
+2. Extend the generic producer-attempt state machine so a feedback-capable
+   module retry constructs `SemanticCorrection` from the latest exact response
+   and the supplied guidance. Preserve attempt provenance, cancellation,
+   terminal policy, diagnostic handling, byte-faithful response provenance,
+   and the rule that only the latest response is included.
+3. Add a model-facing renderer to `internal/framework/semanticreconcile`.
+   Implement it with a typed lookup table keyed by `IssueCategory` and an
+   authoritative list of all categories, rather than a positional array.
+   Render response-local duplicate-group numbers in human-friendly form,
+   de-duplicate repeated guidance, enforce a size bound, and return an error for
+   any unmapped category instead of exposing its raw value. Retain
+   `IssueDetails` for operator/debug diagnostics.
+4. Update the NPC, item, and location registry normalizers to populate the new
+   correction guidance for semantically invalid proposals. Add the item-only
+   currency guidance after the shared prose when typed application rejects a
+   group. Do not add guidance to the invalid-structured-output path when no
+   exact valid candidate exists.
+5. Improve exhausted-fallback diagnostics across all three registries. Describe
+   the final proposal failure accurately, state that affected candidates were
+   preserved separately, and keep warning samples bounded and free of raw model
+   responses.
+6. Update the shared semantic-reconciliation and affected normalizer policy
+   fingerprints so checkpoints produced under blind retry behavior are not
+   reused. Record the durable module-retry correction decision in a new ADR,
+   and update the architecture, pipeline, LLM-runtime, and D&D internal docs in
+   their canonical scopes when the behavior is implemented.
+7. Add lean offline behavioral coverage for corrective module retries, exact
+   latest-response use, feedback-free structural retries, contract validation,
+   complete category-to-prose coverage, domain-specific item guidance, bounded
+   messages, safe final fallback, and warning content. Avoid exact prose
+   snapshots, prompt-length assertions, live providers, and redundant tests of
+   PromptKit internals.
+
+## Non-Goals
+
+- Changing normalize retry counts, PromptKit repair settings, or pipeline
+  terminal policies.
+- Changing the reconciliation prompt protocol, response schema, registry
+  artifact schemas, or request-local candidate IDs.
+- Reintroducing provider-sensitive schema constraints such as `uniqueItems`.
+- Automatically applying an unsafe duplicate proposal or treating a fallback
+  warning as an extraction-quality warning.
+- Persisting raw model responses or correction text in ordinary warnings,
+  receipts, manifests, caches, or checkpoints.
+
+## Acceptance Criteria
+
+- A semantic proposal rejected by an NPC, item, or location registry normalizer
+  receives the exact defective response and actionable prose on the next
+  normalize attempt, using only the configured stage retry budget.
+- Every `IssueCategory` has non-empty model-facing guidance, no raw category or
+  reason-code string reaches the model, and an unmapped future category fails
+  closed.
+- An invalid item consolidation additionally tells the model to keep currency
+  denominations distinct and not mix currency with non-currency items.
+- Structural-output failures without a valid model candidate retain the current
+  fresh-retry behavior.
+- After retry exhaustion, safe consolidations remain applied, affected
+  candidates remain separate, and the single fallback warning explains both
+  the final cause and preservation behavior without including raw model output.
+- Checkpoint identities change where policy changed, canonical documentation
+  and a new ADR describe the implemented behavior, and focused tests plus
+  `go test ./...`, `go vet ./...`, and `go build ./cmd/notarius` pass.

internal/framework/contracts/typed_pipeline.go

@@ -102,11 +102,12 @@ type TypedNormalizeResult[T any] struct {
 	ModelCandidate *ModelCandidate
 }
 
-// Normalize retry diagnostic limits bound module-provided values before the
-// framework persists them in debug artifacts.
+// Normalize retry limits bound module-provided control and diagnostic text
+// before the framework consumes or records it.
 const (
 	MaxNormalizeRetryReasonCodeBytes         = 128
 	MaxNormalizeRetryMessageBytes            = 4096
+	MaxNormalizeRetryCorrectionGuidanceBytes = 4096
 )
 
 // NormalizeRetry asks the framework to retry normalization while retaining a
@@ -114,6 +115,7 @@ const (
 type NormalizeRetry struct {
 	ReasonCode          string
 	Message             string
+	CorrectionGuidance  string
 	FallbackDiagnostics []ProducerDiagnostic
 }
 

internal/framework/pipeline/normalize_retry.go

@@ -30,5 +30,14 @@ func validateNormalizeRetry(retry *contracts.NormalizeRetry) error {
 	if len(retry.Message) > contracts.MaxNormalizeRetryMessageBytes {
 		return errors.New("normalize retry directive message exceeds maximum length")
 	}
+	if !utf8.ValidString(retry.CorrectionGuidance) {
+		return errors.New("normalize retry directive correction guidance has invalid UTF-8")
+	}
+	if retry.CorrectionGuidance != "" && strings.TrimSpace(retry.CorrectionGuidance) == "" {
+		return errors.New("normalize retry directive correction guidance is blank")
+	}
+	if len(retry.CorrectionGuidance) > contracts.MaxNormalizeRetryCorrectionGuidanceBytes {
+		return errors.New("normalize retry directive correction guidance exceeds maximum length")
+	}
 	return nil
 }

internal/framework/pipeline/normalizer_registry.go

@@ -101,6 +101,7 @@ func cloneNormalizeRetry(retry *contracts.NormalizeRetry) *contracts.NormalizeRe
 	return &contracts.NormalizeRetry{
 		ReasonCode:          retry.ReasonCode,
 		Message:             retry.Message,
+		CorrectionGuidance:  retry.CorrectionGuidance,
 		FallbackDiagnostics: contracts.CloneProducerDiagnostics(retry.FallbackDiagnostics),
 	}
 }

internal/framework/pipeline/normalizer_registry_test.go

@@ -21,6 +21,7 @@ func TestNormalizerRegistryErasureClonesRetryDirective(t *testing.T) {
 	retry := &contracts.NormalizeRetry{
 		ReasonCode:         "retryable",
 		Message:            "safe fallback available",
+		CorrectionGuidance: "return a complete corrected proposal",
 	}
 	registry := NewNormalizerRegistry()
 	if err := RegisterNormalizer(registry, ModuleSpec{Key: "test/retry-normalize", Stage: StageNormalize, ExecutionClass: contracts.ExecutionClassDeterministic, ArtifactKind: "test/notes"}, func() (contracts.Normalizer[codecNotes], error) {
@@ -41,7 +42,8 @@ func TestNormalizerRegistryErasureClonesRetryDirective(t *testing.T) {
 		t.Fatalf("normalize() error = %v", err)
 	}
 	retry.Message = "mutated"
-	if result.Retry == nil || result.Retry.Message != "safe fallback available" {
+	retry.CorrectionGuidance = "mutated guidance"
+	if result.Retry == nil || result.Retry.Message != "safe fallback available" || result.Retry.CorrectionGuidance != "return a complete corrected proposal" {
 		t.Fatalf("erased retry result = %#v, want independent retry data", result)
 	}
 }

internal/framework/pipeline/producer_attempts.go

@@ -52,6 +52,7 @@ type producerAttemptRequest struct {
 // the current value as a safe fallback if its shared budget is exhausted.
 // Artifact-specific adapters are responsible for validating and populating it.
 type producerRetryDirective struct {
+	CorrectionGuidance  string
 	FallbackDiagnostics []contracts.ProducerDiagnostic
 }
 
@@ -59,7 +60,10 @@ func (directive *producerRetryDirective) clone() *producerRetryDirective {
 	if directive == nil {
 		return nil
 	}
-	return &producerRetryDirective{FallbackDiagnostics: contracts.CloneProducerDiagnostics(directive.FallbackDiagnostics)}
+	return &producerRetryDirective{
+		CorrectionGuidance:  directive.CorrectionGuidance,
+		FallbackDiagnostics: contracts.CloneProducerDiagnostics(directive.FallbackDiagnostics),
+	}
 }
 
 // producerAttemptOutput is intentionally artifact-neutral. Value remains
@@ -204,10 +208,21 @@ func runProducerAttempts(ctx context.Context, config producerAttemptConfig, prod
 			return failedProducerAttempt(provenance), err
 		}
 		if output.Retry != nil && number < attemptLimit {
+			correction, err = moduleRetryCorrection(output)
+			if err != nil {
+				provenance = append(provenance, producerAttemptProvenance{Number: number, Kind: kind, Outcome: producerAttemptFailed})
+				return failedProducerAttempt(provenance), err
+			}
 			provenance = append(provenance, producerAttemptProvenance{Number: number, Kind: kind, Outcome: producerAttemptRetried})
-			kind, correction = producerAttemptModuleRetry, nil
+			kind = producerAttemptModuleRetry
 			continue
 		}
+		if output.Retry != nil && output.Retry.CorrectionGuidance != "" {
+			if _, err := moduleRetryCorrection(output); err != nil {
+				provenance = append(provenance, producerAttemptProvenance{Number: number, Kind: kind, Outcome: producerAttemptFailed})
+				return failedProducerAttempt(provenance), err
+			}
+		}
 		if output.Retry != nil {
 			output.Diagnostics = append(output.Diagnostics, contracts.CloneProducerDiagnostics(output.Retry.FallbackDiagnostics)...)
 		}
@@ -269,6 +284,23 @@ func runProducerAttempts(ctx context.Context, config producerAttemptConfig, prod
 	return failedProducerAttempt(provenance), errors.New("producer attempt budget was not exhausted deterministically")
 }
 
+func moduleRetryCorrection(output producerAttemptOutput) (*contracts.SemanticCorrection, error) {
+	if output.Retry == nil || output.Retry.CorrectionGuidance == "" {
+		return nil, nil
+	}
+	if output.Candidate == nil {
+		return nil, errors.New("module retry correction guidance requires a model candidate")
+	}
+	if output.Candidate.Protocol != contracts.CorrectionProtocolSingleResponseV1 {
+		return nil, fmt.Errorf("module retry correction guidance requires protocol %q", contracts.CorrectionProtocolSingleResponseV1)
+	}
+	correction, err := contracts.NewSemanticCorrection(output.Candidate.Response, output.Retry.CorrectionGuidance)
+	if err != nil {
+		return nil, fmt.Errorf("construct module retry semantic correction: %w", err)
+	}
+	return correction, nil
+}
+
 func validateProducerAttemptDiagnostics(output producerAttemptOutput) error {
 	if err := contracts.ValidateProducerDiagnostics(output.Diagnostics); err != nil {
 		return fmt.Errorf("producer returned invalid diagnostics: %w", err)

internal/framework/pipeline/producer_attempts_test.go

@@ -218,6 +218,9 @@ func TestRunProducerAttemptsUsesModuleRetryBudgetAndFallback(t *testing.T) {
 		calls := 0
 		terminal, err := runProducerAttempts(context.Background(), producerAttemptConfig{Retries: 1, Policy: DefaultValidationPolicy()}, func(_ context.Context, request producerAttemptRequest) (producerAttemptOutput, error) {
 			calls++
+			if request.Correction != nil {
+				t.Fatalf("feedback-free module retry correction = %#v, want nil", request.Correction)
+			}
 			if request.Number == 1 {
 				return producerAttemptOutput{Value: "fallback", Retry: &producerRetryDirective{}}, nil
 			}
@@ -234,6 +237,58 @@ func TestRunProducerAttemptsUsesModuleRetryBudgetAndFallback(t *testing.T) {
 		}
 	})
 
+	t.Run("feedback retry", func(t *testing.T) {
+		const (
+			defective = `{"duplicate_groups":[{"candidate_ids":[1,99],"canonical_candidate_id":1}]}`
+			guidance  = "Use only candidate IDs from the supplied candidate list. Return one complete corrected response."
+		)
+		var observed *contracts.SemanticCorrection
+		terminal, err := runProducerAttempts(context.Background(), producerAttemptConfig{Retries: 1, Policy: DefaultValidationPolicy()}, func(_ context.Context, request producerAttemptRequest) (producerAttemptOutput, error) {
+			if request.Number == 1 {
+				return producerAttemptOutput{Value: "safe fallback", Candidate: attemptCandidate(t, defective), Retry: &producerRetryDirective{CorrectionGuidance: guidance}}, nil
+			}
+			observed = request.Correction
+			return producerAttemptOutput{Value: "corrected"}, nil
+		}, approveAttempt)
+		if err != nil {
+			t.Fatalf("runProducerAttempts() error = %v", err)
+		}
+		if terminal.Action != producerTerminalAccepted || terminal.Value != "corrected" {
+			t.Fatalf("terminal = %#v, want corrected accepted value", terminal)
+		}
+		if observed == nil || string(observed.AssistantResponse) != defective || observed.UserGuidance != guidance {
+			t.Fatalf("module retry correction = %#v, want exact latest response and guidance", observed)
+		}
+		if got := attemptKinds(terminal.Provenance); !reflect.DeepEqual(got, []producerAttemptKind{producerAttemptInitial, producerAttemptModuleRetry}) {
+			t.Fatalf("attempt kinds = %v", got)
+		}
+	})
+
+	t.Run("feedback-free retry clears prior correction", func(t *testing.T) {
+		terminal, err := runProducerAttempts(context.Background(), producerAttemptConfig{Retries: 2, Policy: DefaultValidationPolicy()}, func(_ context.Context, request producerAttemptRequest) (producerAttemptOutput, error) {
+			switch request.Number {
+			case 1:
+				return producerAttemptOutput{Value: "first fallback", Candidate: attemptCandidate(t, "first defective response"), Retry: &producerRetryDirective{CorrectionGuidance: "Correct the first response."}}, nil
+			case 2:
+				if request.Correction == nil || string(request.Correction.AssistantResponse) != "first defective response" {
+					t.Fatalf("second attempt correction = %#v", request.Correction)
+				}
+				return producerAttemptOutput{Value: "second fallback", Retry: &producerRetryDirective{}}, nil
+			case 3:
+				if request.Correction != nil {
+					t.Fatalf("third attempt retained stale correction %#v", request.Correction)
+				}
+				return producerAttemptOutput{Value: "accepted"}, nil
+			default:
+				t.Fatalf("unexpected producer attempt %d", request.Number)
+				return producerAttemptOutput{}, nil
+			}
+		}, approveAttempt)
+		if err != nil || terminal.Action != producerTerminalAccepted || terminal.Value != "accepted" {
+			t.Fatalf("terminal = %#v, error = %v", terminal, err)
+		}
+	})
+
 	t.Run("fallback", func(t *testing.T) {
 		fallbackDiagnostic := contracts.ProducerDiagnostic{Disposition: contracts.DiagnosticDispositionWarning, Category: contracts.DiagnosticCategoryFallback, ReasonCode: "fallback", OccurrenceCount: 1, Samples: []contracts.DiagnosticSample{{Scope: "fallback", Message: "fallback warning"}}}
 		terminal, err := runProducerAttempts(context.Background(), producerAttemptConfig{Policy: DefaultValidationPolicy()}, func(context.Context, producerAttemptRequest) (producerAttemptOutput, error) {
@@ -248,6 +303,24 @@ func TestRunProducerAttemptsUsesModuleRetryBudgetAndFallback(t *testing.T) {
 	})
 }
 
+func TestRunProducerAttemptsRejectsModuleCorrectionWithoutModelCandidate(t *testing.T) {
+	producerCalls := 0
+	validatorCalls := 0
+	terminal, err := runProducerAttempts(context.Background(), producerAttemptConfig{Retries: 1, Policy: DefaultValidationPolicy()}, func(context.Context, producerAttemptRequest) (producerAttemptOutput, error) {
+		producerCalls++
+		return producerAttemptOutput{Value: "safe", Retry: &producerRetryDirective{CorrectionGuidance: "Return a complete corrected response."}}, nil
+	}, func(context.Context, producerAttemptOutput) (validationReport, error) {
+		validatorCalls++
+		return validationReport{}, nil
+	})
+	if err == nil || !strings.Contains(err.Error(), "requires a model candidate") {
+		t.Fatalf("runProducerAttempts() error = %v, want model-candidate contract failure", err)
+	}
+	if terminal.Action != producerTerminalFailed || producerCalls != 1 || validatorCalls != 0 {
+		t.Fatalf("terminal = %#v, producer calls = %d, validator calls = %d", terminal, producerCalls, validatorCalls)
+	}
+}
+
 func TestRunProducerAttemptsRejectsInvalidDiagnosticsWithoutRetry(t *testing.T) {
 	tests := []struct {
 		name   string

internal/framework/pipeline/runner_normalize_retry_test.go

@@ -153,6 +153,60 @@ func TestRunnerHandlesRetryableNormalizeFallbacks(t *testing.T) {
 	}
 }
 
+func TestRunnerForwardsModuleRequestedNormalizeCorrection(t *testing.T) {
+	const (
+		defective = `{"duplicate_groups":[{"candidate_ids":[1,99],"canonical_candidate_id":1}]}`
+		guidance  = "Duplicate group 1 must use only supplied candidate IDs. Return one complete corrected response."
+	)
+	prepared := preparedAttemptDebugPipeline(t)
+	lane := &prepared.Steps[0].lanes[0]
+	lane.resolved.Normalize.Retries = 1
+	var observed *contracts.SemanticCorrection
+	calls := 0
+	lane.typed.normalize = func(_ context.Context, _ any, request contracts.TypedNormalizeRequest[any]) (erasedTypedResult, error) {
+		calls++
+		if request.Correction != nil {
+			clone, err := contracts.CloneSemanticCorrection(request.Correction)
+			if err != nil {
+				return erasedTypedResult{}, err
+			}
+			observed = clone
+		}
+		if calls == 1 {
+			return erasedTypedResult{
+				Value:          codecNotes{Items: []string{"safe fallback"}},
+				ModelCandidate: attemptCandidate(t, defective),
+				Retry: &contracts.NormalizeRetry{
+					ReasonCode:         "semantic_proposal_invalid",
+					Message:            "operator-facing proposal diagnostic",
+					CorrectionGuidance: guidance,
+				},
+			}, nil
+		}
+		return erasedTypedResult{Value: codecNotes{Items: []string{"corrected"}}}, nil
+	}
+	debug := newCapturedDebugRecorder()
+	output, err := New().Run(context.Background(), RunInput{Prepared: prepared, RawInput: []byte("input"), Debug: debug})
+	if err != nil {
+		t.Fatalf("Run() error = %v", err)
+	}
+	if calls != 2 || observed == nil || string(observed.AssistantResponse) != defective || observed.UserGuidance != guidance {
+		t.Fatalf("normalize calls = %d correction = %#v", calls, observed)
+	}
+	if len(output.NormalizeOutputs) != 1 || len(output.Rejected) != 0 {
+		t.Fatalf("run output = %#v, want corrected accepted output", output)
+	}
+	var retryDebug strings.Builder
+	for _, name := range debug.names() {
+		if strings.HasPrefix(name, "normalize/notes/attempt-") && strings.HasSuffix(name, ".json") {
+			retryDebug.Write(debug.json[name])
+		}
+	}
+	if !strings.Contains(retryDebug.String(), `"correction_available":true`) || strings.Contains(retryDebug.String(), guidance) || strings.Contains(retryDebug.String(), defective) {
+		t.Fatalf("retry debug = %s, want safe correction metadata without content", retryDebug.String())
+	}
+}
+
 func TestRunnerValidatesNormalizeRetryDiagnostics(t *testing.T) {
 	const (
 		reasonSentinel  = "reason-diagnostic-sentinel"
@@ -160,9 +214,11 @@ func TestRunnerValidatesNormalizeRetryDiagnostics(t *testing.T) {
 	)
 	reasonOverLimit := strings.Repeat("r", contracts.MaxNormalizeRetryReasonCodeBytes-len(reasonSentinel)) + reasonSentinel + "x"
 	messageOverLimit := strings.Repeat("m", contracts.MaxNormalizeRetryMessageBytes-len(messageSentinel)) + messageSentinel + "x"
+	guidanceOverLimit := strings.Repeat("g", contracts.MaxNormalizeRetryCorrectionGuidanceBytes-len(messageSentinel)) + messageSentinel + "x"
 	tests := []struct {
 		name         string
 		retry        contracts.NormalizeRetry
+		candidate    *contracts.ModelCandidate
 		wantError    string
 		hiddenValues []string
 	}{
@@ -173,6 +229,15 @@ func TestRunnerValidatesNormalizeRetryDiagnostics(t *testing.T) {
 				Message:    strings.Repeat("m", contracts.MaxNormalizeRetryMessageBytes),
 			},
 		},
+		{
+			name: "accepts correction guidance byte limit with candidate",
+			retry: contracts.NormalizeRetry{
+				ReasonCode:         reasonSentinel,
+				Message:            messageSentinel,
+				CorrectionGuidance: strings.Repeat("g", contracts.MaxNormalizeRetryCorrectionGuidanceBytes),
+			},
+			candidate: attemptCandidate(t, `{"duplicate_groups":[]}`),
+		},
 		{
 			name: "rejects oversized reason code",
 			retry: contracts.NormalizeRetry{
@@ -209,6 +274,46 @@ func TestRunnerValidatesNormalizeRetryDiagnostics(t *testing.T) {
 			wantError:    "message has invalid UTF-8",
 			hiddenValues: []string{reasonSentinel, messageSentinel},
 		},
+		{
+			name: "rejects oversized correction guidance",
+			retry: contracts.NormalizeRetry{
+				ReasonCode:         reasonSentinel,
+				Message:            messageSentinel,
+				CorrectionGuidance: guidanceOverLimit,
+			},
+			wantError:    "correction guidance exceeds maximum length",
+			hiddenValues: []string{reasonSentinel, messageSentinel},
+		},
+		{
+			name: "rejects invalid correction guidance UTF-8",
+			retry: contracts.NormalizeRetry{
+				ReasonCode:         reasonSentinel,
+				Message:            messageSentinel,
+				CorrectionGuidance: messageSentinel + string([]byte{0xff}),
+			},
+			wantError:    "correction guidance has invalid UTF-8",
+			hiddenValues: []string{reasonSentinel, messageSentinel},
+		},
+		{
+			name: "rejects blank correction guidance",
+			retry: contracts.NormalizeRetry{
+				ReasonCode:         reasonSentinel,
+				Message:            messageSentinel,
+				CorrectionGuidance: " \t\n ",
+			},
+			wantError:    "correction guidance is blank",
+			hiddenValues: []string{reasonSentinel, messageSentinel},
+		},
+		{
+			name: "rejects correction guidance without candidate",
+			retry: contracts.NormalizeRetry{
+				ReasonCode:         reasonSentinel,
+				Message:            messageSentinel,
+				CorrectionGuidance: "Return one complete corrected response.",
+			},
+			wantError:    "requires a model candidate",
+			hiddenValues: []string{reasonSentinel, messageSentinel},
+		},
 		{
 			name: "rejects blank reason code",
 			retry: contracts.NormalizeRetry{
@@ -233,7 +338,7 @@ func TestRunnerValidatesNormalizeRetryDiagnostics(t *testing.T) {
 		t.Run(tc.name, func(t *testing.T) {
 			prepared := preparedAttemptDebugPipeline(t)
 			prepared.Steps[0].lanes[0].typed.normalize = func(context.Context, any, contracts.TypedNormalizeRequest[any]) (erasedTypedResult, error) {
-				return erasedTypedResult{Value: codecNotes{Items: []string{"safe"}}, Retry: &tc.retry}, nil
+				return erasedTypedResult{Value: codecNotes{Items: []string{"safe"}}, Retry: &tc.retry, ModelCandidate: tc.candidate}, nil
 			}
 			debug := newCapturedDebugRecorder()
 			output, err := New().Run(context.Background(), RunInput{Prepared: prepared, RawInput: []byte("input"), Debug: debug})

internal/framework/pipeline/runner_typed.go

@@ -451,8 +451,8 @@ func (r *Runner) runNormalizeStage(ctx context.Context, input RunInput, checkpoi
 					return producerAttemptOutput{}, terminal.record(map[string]any{"output": debugCheckpointArtifact(serializedCandidate)}, fmt.Errorf("normalize lane %q returned invalid retry directive: %w", lane.ID, err))
 				}
 				anotherAttempt := request.Number <= lane.Normalize.Retries
-				attemptValue.retry = map[string]any{"reason_code": result.Retry.ReasonCode, "message": result.Retry.Message, "another_attempt": anotherAttempt, "fallback_accepted": !anotherAttempt}
-				directive = &producerRetryDirective{FallbackDiagnostics: contracts.CloneProducerDiagnostics(result.Retry.FallbackDiagnostics)}
+				attemptValue.retry = map[string]any{"reason_code": result.Retry.ReasonCode, "message": result.Retry.Message, "correction_available": result.Retry.CorrectionGuidance != "", "another_attempt": anotherAttempt, "fallback_accepted": !anotherAttempt}
+				directive = &producerRetryDirective{CorrectionGuidance: result.Retry.CorrectionGuidance, FallbackDiagnostics: contracts.CloneProducerDiagnostics(result.Retry.FallbackDiagnostics)}
 				if anotherAttempt {
 					payload := map[string]any{"output": debugCheckpointArtifact(serializedCandidate), "retry": attemptValue.retry}
 					if debugErr := terminal.record(payload, nil); debugErr != nil {

internal/framework/semanticreconcile/identity.go

@@ -13,7 +13,7 @@ import (
 )
 
 // Policy identifies the framework-owned reconciliation and assessment rules.
-const Policy = "semantic_reconciliation.v1"
+const Policy = "semantic_reconciliation.v2"
 
 var _ contracts.ManifestMetadataProvider = (*Engine)(nil)
 var _ pipeline.CheckpointFingerprintProvider = (*Engine)(nil)

internal/framework/semanticreconcile/proposal.go

@@ -3,6 +3,10 @@ package semanticreconcile
 import (
 	"fmt"
 	"sort"
+	"strings"
+	"unicode/utf8"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/contracts"
 )
 
 // ProposalResponse is the complete private structured response contract.
@@ -31,6 +35,28 @@ const (
 	IssueOverlappingMember    IssueCategory = "overlapping_member"
 )
 
+var allIssueCategories = []IssueCategory{
+	IssueMemberNonPositive,
+	IssueMemberUnknown,
+	IssueRepeatedMember,
+	IssueFewerThanTwoMembers,
+	IssueCanonicalNonPositive,
+	IssueCanonicalUnknown,
+	IssueCanonicalNotMember,
+	IssueOverlappingMember,
+}
+
+var issueCorrectionProse = map[IssueCategory]string{
+	IssueMemberNonPositive:    "Use only positive candidate IDs from the supplied candidate list.",
+	IssueMemberUnknown:        "Remove every candidate ID that is not present in the supplied candidate list.",
+	IssueRepeatedMember:       "List each candidate ID at most once within the duplicate group.",
+	IssueFewerThanTwoMembers:  "Include at least two distinct candidate IDs, or omit the duplicate group.",
+	IssueCanonicalNonPositive: "Choose a positive canonical_candidate_id from the supplied candidate list.",
+	IssueCanonicalUnknown:     "Choose canonical_candidate_id from the supplied candidate list.",
+	IssueCanonicalNotMember:   "Make canonical_candidate_id one of the candidate_ids in the same duplicate group.",
+	IssueOverlappingMember:    "Place each candidate ID in at most one duplicate group.",
+}
+
 // Issue identifies an unsafe proposal category at its original response group
 // index without prescribing caller diagnostic text.
 type Issue struct {
@@ -38,8 +64,8 @@ type Issue struct {
 	Category   IssueCategory
 }
 
-// IssueDetails renders stable, domain-neutral proposal diagnostics for an
-// adapter's retry message.
+// IssueDetails renders stable, domain-neutral proposal diagnostics for
+// operators and debug records. Its internal categories are not model guidance.
 func IssueDetails(issues []Issue) []string {
 	details := make([]string, len(issues))
 	for index, issue := range issues {
@@ -48,6 +74,94 @@ func IssueDetails(issues []Issue) []string {
 	return details
 }
 
+// CorrectionDetails translates proposal issues into stable model-facing prose.
+// The response-local group ordinals help the model find the defective group in
+// the exact response appended to the correction request.
+func CorrectionDetails(issues []Issue) ([]string, error) {
+	groupsByCategory := make(map[IssueCategory][]int)
+	seen := make(map[IssueCategory]map[int]struct{})
+	for _, issue := range issues {
+		if issue.GroupIndex < 0 {
+			return nil, fmt.Errorf("semantic reconciliation issue group index must not be negative")
+		}
+		if _, exists := issueCorrectionProse[issue.Category]; !exists {
+			return nil, fmt.Errorf("semantic reconciliation issue category %q has no correction guidance", issue.Category)
+		}
+		if seen[issue.Category] == nil {
+			seen[issue.Category] = make(map[int]struct{})
+		}
+		if _, exists := seen[issue.Category][issue.GroupIndex]; exists {
+			continue
+		}
+		seen[issue.Category][issue.GroupIndex] = struct{}{}
+		groupsByCategory[issue.Category] = append(groupsByCategory[issue.Category], issue.GroupIndex)
+	}
+
+	details := make([]string, 0, len(groupsByCategory))
+	for _, category := range allIssueCategories {
+		groups := groupsByCategory[category]
+		if len(groups) == 0 {
+			continue
+		}
+		sort.Ints(groups)
+		details = append(details, fmt.Sprintf("%s: %s", correctionGroupLabel(groups), issueCorrectionProse[category]))
+	}
+	return details, nil
+}
+
+// CorrectionGuidance builds one bounded request for a complete corrected
+// proposal. Additional details let a typed owner append a domain rule without
+// weakening or duplicating the shared protocol guidance.
+func CorrectionGuidance(issues []Issue, additionalDetails ...string) (string, error) {
+	details, err := CorrectionDetails(issues)
+	if err != nil {
+		return "", err
+	}
+	for _, detail := range additionalDetails {
+		if !utf8.ValidString(detail) {
+			return "", fmt.Errorf("semantic reconciliation additional correction detail has invalid UTF-8")
+		}
+		detail = strings.TrimSpace(detail)
+		if detail == "" {
+			return "", fmt.Errorf("semantic reconciliation additional correction detail must not be blank")
+		}
+		details = append(details, detail)
+	}
+	if len(details) == 0 {
+		return "", fmt.Errorf("semantic reconciliation correction guidance requires at least one detail")
+	}
+
+	parts := make([]string, 0, len(details)+2)
+	parts = append(parts, "The previous semantic-duplicate proposal was invalid.")
+	parts = append(parts, details...)
+	parts = append(parts, "Return one complete corrected JSON response that follows the original instructions; do not return a patch or commentary.")
+	guidance := strings.Join(parts, " ")
+	if len(guidance) > contracts.MaxNormalizeRetryCorrectionGuidanceBytes {
+		return "", fmt.Errorf("semantic reconciliation correction guidance exceeds maximum length")
+	}
+	return guidance, nil
+}
+
+func correctionGroupLabel(groupIndexes []int) string {
+	const maximumDisplayedGroups = 12
+	displayed := groupIndexes
+	if len(displayed) > maximumDisplayedGroups {
+		displayed = displayed[:maximumDisplayedGroups]
+	}
+	ordinals := make([]string, len(displayed))
+	for index, groupIndex := range displayed {
+		ordinals[index] = fmt.Sprintf("%d", groupIndex+1)
+	}
+	if len(groupIndexes) == 1 {
+		return "Duplicate group " + ordinals[0]
+	}
+	label := "Duplicate groups " + strings.Join(ordinals, ", ")
+	if omitted := len(groupIndexes) - len(displayed); omitted > 0 {
+		label += fmt.Sprintf(", and %d additional affected group(s)", omitted)
+	}
+	return label
+}
+
 // PlanGroup identifies one validated group using original candidate positions.
 type PlanGroup struct {
 	memberPositions   []int

internal/framework/semanticreconcile/proposal_test.go

@@ -2,9 +2,11 @@ package semanticreconcile
 
 import (
 	"reflect"
+	"strings"
 	"testing"
 
 	"gitea.maximumdirect.net/eric/notarius/internal/core/source"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/contracts"
 )
 
 func TestAssessProducesAStableOriginalPositionPlan(t *testing.T) {
@@ -50,6 +52,72 @@ func TestIssueDetailsPreservesIssueOrder(t *testing.T) {
 	}
 }
 
+func TestCorrectionGuidanceCoversEveryIssueCategoryWithoutExposingInternalLabels(t *testing.T) {
+	if len(issueCorrectionProse) != len(allIssueCategories) {
+		t.Fatalf("correction prose entries = %d, categories = %d", len(issueCorrectionProse), len(allIssueCategories))
+	}
+	issues := make([]Issue, len(allIssueCategories))
+	seen := make(map[IssueCategory]struct{}, len(allIssueCategories))
+	for index, category := range allIssueCategories {
+		if _, duplicate := seen[category]; duplicate {
+			t.Fatalf("duplicate authoritative issue category %q", category)
+		}
+		seen[category] = struct{}{}
+		if strings.TrimSpace(issueCorrectionProse[category]) == "" {
+			t.Fatalf("issue category %q has no model-facing prose", category)
+		}
+		issues[index] = Issue{GroupIndex: index, Category: category}
+	}
+
+	details, err := CorrectionDetails(issues)
+	if err != nil {
+		t.Fatalf("CorrectionDetails() error = %v", err)
+	}
+	guidance, err := CorrectionGuidance(issues)
+	if err != nil {
+		t.Fatalf("CorrectionGuidance() error = %v", err)
+	}
+	if len(details) != len(allIssueCategories) || !strings.Contains(guidance, "Duplicate group 1") || !strings.Contains(guidance, "complete corrected JSON response") || len(guidance) > contracts.MaxNormalizeRetryCorrectionGuidanceBytes {
+		t.Fatalf("correction details = %#v guidance = %q", details, guidance)
+	}
+	for _, category := range allIssueCategories {
+		if strings.Contains(guidance, string(category)) {
+			t.Fatalf("model guidance exposed internal category %q: %q", category, guidance)
+		}
+	}
+}
+
+func TestCorrectionDetailsDeduplicatesAndBoundsAffectedGroupLists(t *testing.T) {
+	issues := make([]Issue, 0, 257)
+	for group := 0; group < 256; group++ {
+		issues = append(issues, Issue{GroupIndex: group, Category: IssueMemberUnknown})
+	}
+	issues = append(issues, Issue{GroupIndex: 0, Category: IssueMemberUnknown})
+
+	details, err := CorrectionDetails(issues)
+	if err != nil {
+		t.Fatalf("CorrectionDetails() error = %v", err)
+	}
+	if len(details) != 1 || !strings.Contains(details[0], "additional affected group") || strings.Count(details[0], "Duplicate groups") != 1 {
+		t.Fatalf("CorrectionDetails() = %#v, want one bounded grouped detail", details)
+	}
+	guidance, err := CorrectionGuidance(issues)
+	if err != nil || len(guidance) > contracts.MaxNormalizeRetryCorrectionGuidanceBytes {
+		t.Fatalf("CorrectionGuidance() = %q, %v", guidance, err)
+	}
+}
+
+func TestCorrectionDetailsRejectsUnknownOrInvalidIssues(t *testing.T) {
+	for _, issues := range [][]Issue{
+		{{GroupIndex: 0, Category: "future_unmapped_category"}},
+		{{GroupIndex: -1, Category: IssueMemberUnknown}},
+	} {
+		if details, err := CorrectionDetails(issues); err == nil || details != nil {
+			t.Fatalf("CorrectionDetails(%#v) = %#v, %v; want fail-closed error", issues, details, err)
+		}
+	}
+}
+
 func TestAssessRejectsEveryUnsafeLocalGroupShape(t *testing.T) {
 	preparation := proposalPreparation(t)
 	tests := []struct {

internal/modules/dnd/normalize/itemregistry/normalizer.go

@@ -23,7 +23,7 @@ const (
 	Key                 = "dnd/item-registry"
 	PromptID            = "dnd.item_registry.normalize"
 	PromptVersion       = "v1"
-	normalizationPolicy = "dnd.item_registry.normalize.v3"
+	normalizationPolicy = "dnd.item_registry.normalize.v4"
 	NormalizationPolicy = normalizationPolicy
 
 	ReasonCodeItemFieldsNormalized                = "item_fields_normalized"
@@ -129,7 +129,7 @@ func (n *Normalizer) Normalize(ctx context.Context, req contracts.TypedNormalize
 	case semanticreconcile.SkippedInsufficientCandidates:
 		return normalizationResult(deterministic, findings, nil, nil)
 	case semanticreconcile.SkippedLimitExceeded:
-		return fallbackResult(deterministic, findings, nil, semanticFallbackFinding(-1))
+		return fallbackResult(deterministic, findings, nil, semanticFallbackFinding(-1, "the candidate set exceeded the configured semantic-reconciliation limits"))
 	case semanticreconcile.RetryableInvalidStructuredOutput:
 		return n.invalidStructuredResult(deterministic, findings)
 	case semanticreconcile.Complete, semanticreconcile.RetryableDiscardedProposalGroups:
@@ -150,22 +150,38 @@ func (n *Normalizer) Normalize(ctx context.Context, req contracts.TypedNormalize
 }
 
 func (n *Normalizer) invalidStructuredResult(value dnd.ItemRegistry, findings []diagnostics.Finding) (contracts.TypedNormalizeResult[dnd.ItemRegistry], error) {
-	return retryResultWithFallback(value, findings, nil, nil, ReasonCodeItemSemanticRetryProposalInvalid, "semantic proposal requires retry: invalid structured output", semanticFallbackFinding(-1))
+	return retryResultWithFallback(value, findings, nil, nil, ReasonCodeItemSemanticRetryProposalInvalid, "", "semantic proposal requires retry: invalid structured output", semanticFallbackFinding(-1, "the final model response did not satisfy the structured-output contract"))
 }
 
 func retryResult(value dnd.ItemRegistry, findings, advisoryFindings []diagnostics.Finding, reconciliation semanticreconcile.Result, rejectedGroups int) (contracts.TypedNormalizeResult[dnd.ItemRegistry], error) {
-	details := semanticreconcile.IssueDetails(reconciliation.Issues())
+	issues := reconciliation.Issues()
+	operatorDetails := semanticreconcile.IssueDetails(issues)
+	correctionDetails, err := semanticreconcile.CorrectionDetails(issues)
+	if err != nil {
+		return contracts.TypedNormalizeResult[dnd.ItemRegistry]{}, normalizerErrorf("build semantic correction details: %w", err)
+	}
+	var additionalGuidance []string
 	if rejectedGroups > 0 {
-		details = append(details, "currency may only be consolidated with aliases of one denomination")
+		const currencyGuidance = "Consolidate currency only with aliases of the same denomination; keep different denominations separate and do not combine currency with non-currency items."
+		operatorDetails = append(operatorDetails, "currency may only be consolidated with aliases of one denomination")
+		correctionDetails = append(correctionDetails, currencyGuidance)
+		additionalGuidance = append(additionalGuidance, currencyGuidance)
+	}
+	correctionGuidance, err := semanticreconcile.CorrectionGuidance(issues, additionalGuidance...)
+	if err != nil {
+		return contracts.TypedNormalizeResult[dnd.ItemRegistry]{}, normalizerErrorf("build semantic correction guidance: %w", err)
 	}
 	discardedGroups := reconciliation.DiscardedGroupCount() + rejectedGroups
-	return retryResultWithFallback(value, findings, advisoryFindings, reconciliation.ModelCandidate(), ReasonCodeItemSemanticRetryProposalInvalid, diagnostics.Aggregate("semantic proposal requires retry", details), semanticFallbackFinding(discardedGroups))
+	return retryResultWithFallback(value, findings, advisoryFindings, reconciliation.ModelCandidate(), ReasonCodeItemSemanticRetryProposalInvalid, correctionGuidance, diagnostics.Aggregate("semantic proposal requires retry", operatorDetails), semanticFallbackFinding(discardedGroups, correctionDetails...))
 }
 
-func semanticFallbackFinding(discarded int) diagnostics.Finding {
-	message := "semantic proposal could not be applied"
+func semanticFallbackFinding(discarded int, details ...string) diagnostics.Finding {
+	message := "semantic reconciliation could not be completed; all affected candidates were preserved separately"
 	if discarded >= 0 {
-		message = fmt.Sprintf("%d proposal group(s) omitted after semantic proposal retry exhaustion", discarded)
+		message = fmt.Sprintf("%d proposal group(s) omitted from the final semantic proposal; all affected candidates were preserved separately", discarded)
+	}
+	if len(details) > 0 {
+		message = diagnostics.Aggregate(message, details)
 	}
 	return diagnostics.Finding{Scope: "items", ReasonCode: ReasonCodeItemSemanticReconciliationExhausted, Message: message}
 }
@@ -196,7 +212,7 @@ func fallbackResult(value dnd.ItemRegistry, findings, advisoryFindings []diagnos
 	return result, nil
 }
 
-func retryResultWithFallback(value dnd.ItemRegistry, findings, advisoryFindings []diagnostics.Finding, candidate *contracts.ModelCandidate, reasonCode, message string, fallback diagnostics.Finding) (contracts.TypedNormalizeResult[dnd.ItemRegistry], error) {
+func retryResultWithFallback(value dnd.ItemRegistry, findings, advisoryFindings []diagnostics.Finding, candidate *contracts.ModelCandidate, reasonCode, correctionGuidance, message string, fallback diagnostics.Finding) (contracts.TypedNormalizeResult[dnd.ItemRegistry], error) {
 	result, err := normalizationResult(value, findings, advisoryFindings, candidate)
 	if err != nil {
 		return contracts.TypedNormalizeResult[dnd.ItemRegistry]{}, err
@@ -205,7 +221,7 @@ func retryResultWithFallback(value dnd.ItemRegistry, findings, advisoryFindings
 	if err != nil {
 		return contracts.TypedNormalizeResult[dnd.ItemRegistry]{}, normalizerErrorf("collect fallback diagnostic: %w", err)
 	}
-	result.Retry = &contracts.NormalizeRetry{ReasonCode: reasonCode, Message: message, FallbackDiagnostics: fallbackGroups}
+	result.Retry = &contracts.NormalizeRetry{ReasonCode: reasonCode, Message: message, CorrectionGuidance: correctionGuidance, FallbackDiagnostics: fallbackGroups}
 	return result, nil
 }
 

internal/modules/dnd/normalize/itemregistry/normalizer_test.go

@@ -289,7 +289,10 @@ func TestNormalizeAppliesIndependentGroupAndCountsAllOmissions(t *testing.T) {
 	if !hasDiagnostic(result.Diagnostics, ReasonCodeDuplicateItemCollapsed, contracts.DiagnosticDispositionObservation) || !hasDiagnostic(result.Diagnostics, ReasonCodeItemSemanticProposalInvalid, contracts.DiagnosticDispositionAdvisory) {
 		t.Fatalf("diagnostics = %#v, want accepted and guarded-group diagnostics", result.Diagnostics)
 	}
-	if len(result.Retry.FallbackDiagnostics) != 1 || !strings.Contains(result.Retry.FallbackDiagnostics[0].Samples[0].Message, "2 proposal group(s)") {
+	if !strings.Contains(result.Retry.CorrectionGuidance, "aliases of the same denomination") || !strings.Contains(result.Retry.CorrectionGuidance, "do not combine currency with non-currency items") || !strings.Contains(result.Retry.CorrectionGuidance, "supplied candidate list") || strings.Contains(result.Retry.CorrectionGuidance, "member_unknown") || strings.Contains(result.Retry.CorrectionGuidance, ReasonCodeItemSemanticRetryProposalInvalid) {
+		t.Fatalf("correction guidance = %q, want shared and item-specific prose only", result.Retry.CorrectionGuidance)
+	}
+	if len(result.Retry.FallbackDiagnostics) != 1 || !strings.Contains(result.Retry.FallbackDiagnostics[0].Samples[0].Message, "2 proposal group(s)") || !strings.Contains(result.Retry.FallbackDiagnostics[0].Samples[0].Message, "preserved separately") || !strings.Contains(result.Retry.FallbackDiagnostics[0].Samples[0].Message, "same denomination") {
 		t.Fatalf("retry = %#v, want one guarded and one malformed group counted", result.Retry)
 	}
 }
@@ -321,6 +324,9 @@ func TestNormalizeRetryFallbackErrorsAndIdempotence(t *testing.T) {
 	if err != nil || invalid.Retry == nil || invalid.Retry.ReasonCode != ReasonCodeItemSemanticRetryProposalInvalid {
 		t.Fatalf("invalid result = %#v, %v", invalid, err)
 	}
+	if invalid.Retry.CorrectionGuidance != "" || !strings.Contains(invalid.Retry.FallbackDiagnostics[0].Samples[0].Message, "structured-output contract") {
+		t.Fatalf("invalid structured retry = %#v, want feedback-free retry with final-cause fallback", invalid.Retry)
+	}
 	_, err = newNormalizer(t, &recordingNormalizerClient{err: errors.New("provider unavailable")}).Normalize(context.Background(), normalizeRequestWithSource(input, doc))
 	if err == nil || !strings.Contains(err.Error(), "provider unavailable") {
 		t.Fatalf("provider error = %v", err)

internal/modules/dnd/normalize/locationregistry/normalizer.go

@@ -24,7 +24,7 @@ const (
 	Key                 = "dnd/location-registry"
 	PromptID            = "dnd.location_registry.normalize"
 	PromptVersion       = "v1"
-	normalizationPolicy = "dnd.location_registry.normalize.v3"
+	normalizationPolicy = "dnd.location_registry.normalize.v4"
 	NormalizationPolicy = normalizationPolicy
 
 	ReasonCodeLocationFieldsNormalized                = "location_fields_normalized"
@@ -129,7 +129,7 @@ func (n *Normalizer) Normalize(ctx context.Context, req contracts.TypedNormalize
 	case semanticreconcile.SkippedInsufficientCandidates:
 		return normalizationResult(deterministic, findings, nil)
 	case semanticreconcile.SkippedLimitExceeded:
-		return fallbackResult(deterministic, findings, semanticFallbackFinding(-1))
+		return fallbackResult(deterministic, findings, semanticFallbackFinding(-1, "the candidate set exceeded the configured semantic-reconciliation limits"))
 	case semanticreconcile.RetryableInvalidStructuredOutput:
 		return n.invalidStructuredResult(deterministic, findings)
 	case semanticreconcile.Complete, semanticreconcile.RetryableDiscardedProposalGroups:
@@ -149,17 +149,29 @@ func (n *Normalizer) Normalize(ctx context.Context, req contracts.TypedNormalize
 }
 
 func (n *Normalizer) invalidStructuredResult(value dnd.LocationRegistry, findings []diagnostics.Finding) (contracts.TypedNormalizeResult[dnd.LocationRegistry], error) {
-	return retryResultWithFallback(value, findings, nil, "semantic proposal requires retry: invalid structured output", semanticFallbackFinding(-1))
+	return retryResultWithFallback(value, findings, nil, "", "semantic proposal requires retry: invalid structured output", semanticFallbackFinding(-1, "the final model response did not satisfy the structured-output contract"))
 }
 
 func retryResult(value dnd.LocationRegistry, findings []diagnostics.Finding, reconciliation semanticreconcile.Result) (contracts.TypedNormalizeResult[dnd.LocationRegistry], error) {
-	return retryResultWithFallback(value, findings, reconciliation.ModelCandidate(), diagnostics.Aggregate("semantic proposal requires retry", semanticreconcile.IssueDetails(reconciliation.Issues())), semanticFallbackFinding(reconciliation.DiscardedGroupCount()))
+	issues := reconciliation.Issues()
+	correctionDetails, err := semanticreconcile.CorrectionDetails(issues)
+	if err != nil {
+		return contracts.TypedNormalizeResult[dnd.LocationRegistry]{}, normalizerErrorf("build semantic correction details: %w", err)
+	}
+	correctionGuidance, err := semanticreconcile.CorrectionGuidance(issues)
+	if err != nil {
+		return contracts.TypedNormalizeResult[dnd.LocationRegistry]{}, normalizerErrorf("build semantic correction guidance: %w", err)
+	}
+	return retryResultWithFallback(value, findings, reconciliation.ModelCandidate(), correctionGuidance, diagnostics.Aggregate("semantic proposal requires retry", semanticreconcile.IssueDetails(issues)), semanticFallbackFinding(reconciliation.DiscardedGroupCount(), correctionDetails...))
 }
 
-func semanticFallbackFinding(discarded int) diagnostics.Finding {
-	message := "semantic proposal could not be applied"
+func semanticFallbackFinding(discarded int, details ...string) diagnostics.Finding {
+	message := "semantic reconciliation could not be completed; all affected candidates were preserved separately"
 	if discarded >= 0 {
-		message = fmt.Sprintf("%d proposal group(s) omitted after semantic proposal retry exhaustion", discarded)
+		message = fmt.Sprintf("%d proposal group(s) omitted from the final semantic proposal; all affected candidates were preserved separately", discarded)
+	}
+	if len(details) > 0 {
+		message = diagnostics.Aggregate(message, details)
 	}
 	return diagnostics.Finding{Scope: "locations", ReasonCode: ReasonCodeLocationSemanticReconciliationExhausted, Message: message}
 }
@@ -185,7 +197,7 @@ func fallbackResult(value dnd.LocationRegistry, findings []diagnostics.Finding,
 	return result, nil
 }
 
-func retryResultWithFallback(value dnd.LocationRegistry, findings []diagnostics.Finding, candidate *contracts.ModelCandidate, message string, fallback diagnostics.Finding) (contracts.TypedNormalizeResult[dnd.LocationRegistry], error) {
+func retryResultWithFallback(value dnd.LocationRegistry, findings []diagnostics.Finding, candidate *contracts.ModelCandidate, correctionGuidance, message string, fallback diagnostics.Finding) (contracts.TypedNormalizeResult[dnd.LocationRegistry], error) {
 	result, err := normalizationResult(value, findings, candidate)
 	if err != nil {
 		return contracts.TypedNormalizeResult[dnd.LocationRegistry]{}, err
@@ -194,7 +206,7 @@ func retryResultWithFallback(value dnd.LocationRegistry, findings []diagnostics.
 	if err != nil {
 		return contracts.TypedNormalizeResult[dnd.LocationRegistry]{}, normalizerErrorf("collect fallback diagnostic: %w", err)
 	}
-	result.Retry = &contracts.NormalizeRetry{ReasonCode: ReasonCodeLocationSemanticProposalInvalid, Message: message, FallbackDiagnostics: fallbackGroups}
+	result.Retry = &contracts.NormalizeRetry{ReasonCode: ReasonCodeLocationSemanticProposalInvalid, Message: message, CorrectionGuidance: correctionGuidance, FallbackDiagnostics: fallbackGroups}
 	return result, nil
 }
 

internal/modules/dnd/normalize/locationregistry/normalizer_test.go

@@ -128,7 +128,10 @@ func TestNormalizeRejectsUnsafeAndOverlappingGroupsWithoutLosingCandidates(t *te
 	if err != nil || result.Retry == nil || len(result.Value.Locations) != 3 || !strings.Contains(result.Retry.Message, "overlapping_member") {
 		t.Fatalf("Normalize() = %#v, %v; want safe retry fallback", result, err)
 	}
-	if len(result.Retry.FallbackDiagnostics) != 1 || !strings.Contains(result.Retry.FallbackDiagnostics[0].Samples[0].Message, "2 proposal group") {
+	if !strings.Contains(result.Retry.CorrectionGuidance, "Duplicate groups 1, 2") || !strings.Contains(result.Retry.CorrectionGuidance, "at most one duplicate group") || strings.Contains(result.Retry.CorrectionGuidance, "overlapping_member") {
+		t.Fatalf("correction guidance = %q, want model-facing overlap explanation", result.Retry.CorrectionGuidance)
+	}
+	if len(result.Retry.FallbackDiagnostics) != 1 || !strings.Contains(result.Retry.FallbackDiagnostics[0].Samples[0].Message, "2 proposal group") || !strings.Contains(result.Retry.FallbackDiagnostics[0].Samples[0].Message, "preserved separately") {
 		t.Fatalf("fallback diagnostics = %#v", result.Retry.FallbackDiagnostics)
 	}
 }
@@ -188,6 +191,9 @@ func TestNormalizeHandlesRetryFallbackAndErrors(t *testing.T) {
 	if err != nil || invalid.Retry == nil || invalid.Retry.ReasonCode != ReasonCodeLocationSemanticProposalInvalid {
 		t.Fatalf("invalid result = %#v, %v", invalid, err)
 	}
+	if invalid.Retry.CorrectionGuidance != "" || !strings.Contains(invalid.Retry.FallbackDiagnostics[0].Samples[0].Message, "structured-output contract") || !strings.Contains(invalid.Retry.FallbackDiagnostics[0].Samples[0].Message, "preserved separately") {
+		t.Fatalf("invalid structured retry = %#v, want feedback-free retry and informative fallback", invalid.Retry)
+	}
 	_, err = newNormalizer(t, &recordingLocationNormalizerClient{err: errors.New("provider unavailable")}).Normalize(context.Background(), normalizeRequestWithSource(input, doc))
 	if err == nil || !strings.Contains(err.Error(), "provider unavailable") {
 		t.Fatalf("provider error = %v", err)

internal/modules/dnd/normalize/npcregistry/normalizer.go

@@ -23,7 +23,7 @@ const (
 	Key                 = "dnd/npc-registry"
 	PromptID            = "dnd.npc_registry.normalize"
 	PromptVersion       = "v1"
-	normalizationPolicy = "dnd.npc_registry.normalize.v5"
+	normalizationPolicy = "dnd.npc_registry.normalize.v6"
 	NormalizationPolicy = normalizationPolicy
 
 	ReasonCodeNPCFieldsNormalized                = "npc_fields_normalized"
@@ -128,7 +128,7 @@ func (n *Normalizer) Normalize(ctx context.Context, req contracts.TypedNormalize
 	case semanticreconcile.SkippedInsufficientCandidates:
 		return normalizationResult(deterministic, findings, nil)
 	case semanticreconcile.SkippedLimitExceeded:
-		return fallbackResult(deterministic, findings, semanticFallbackFinding(-1))
+		return fallbackResult(deterministic, findings, semanticFallbackFinding(-1, "the candidate set exceeded the configured semantic-reconciliation limits"))
 	case semanticreconcile.RetryableInvalidStructuredOutput:
 		return n.invalidStructuredResult(deterministic, findings)
 	case semanticreconcile.Complete, semanticreconcile.RetryableDiscardedProposalGroups:
@@ -148,17 +148,29 @@ func (n *Normalizer) Normalize(ctx context.Context, req contracts.TypedNormalize
 }
 
 func (n *Normalizer) invalidStructuredResult(value dnd.NPCRegistry, findings []diagnostics.Finding) (contracts.TypedNormalizeResult[dnd.NPCRegistry], error) {
-	return retryResultWithFallback(value, findings, nil, "semantic proposal requires retry: invalid structured output", semanticFallbackFinding(-1))
+	return retryResultWithFallback(value, findings, nil, "", "semantic proposal requires retry: invalid structured output", semanticFallbackFinding(-1, "the final model response did not satisfy the structured-output contract"))
 }
 
 func retryResult(value dnd.NPCRegistry, findings []diagnostics.Finding, reconciliation semanticreconcile.Result) (contracts.TypedNormalizeResult[dnd.NPCRegistry], error) {
-	return retryResultWithFallback(value, findings, reconciliation.ModelCandidate(), diagnostics.Aggregate("semantic proposal requires retry", semanticreconcile.IssueDetails(reconciliation.Issues())), semanticFallbackFinding(reconciliation.DiscardedGroupCount()))
+	issues := reconciliation.Issues()
+	correctionDetails, err := semanticreconcile.CorrectionDetails(issues)
+	if err != nil {
+		return contracts.TypedNormalizeResult[dnd.NPCRegistry]{}, normalizerErrorf("build semantic correction details: %w", err)
+	}
+	correctionGuidance, err := semanticreconcile.CorrectionGuidance(issues)
+	if err != nil {
+		return contracts.TypedNormalizeResult[dnd.NPCRegistry]{}, normalizerErrorf("build semantic correction guidance: %w", err)
+	}
+	return retryResultWithFallback(value, findings, reconciliation.ModelCandidate(), correctionGuidance, diagnostics.Aggregate("semantic proposal requires retry", semanticreconcile.IssueDetails(issues)), semanticFallbackFinding(reconciliation.DiscardedGroupCount(), correctionDetails...))
 }
 
-func semanticFallbackFinding(discardedGroups int) diagnostics.Finding {
-	message := "semantic proposal could not be applied"
+func semanticFallbackFinding(discardedGroups int, details ...string) diagnostics.Finding {
+	message := "semantic reconciliation could not be completed; all affected candidates were preserved separately"
 	if discardedGroups >= 0 {
-		message = fmt.Sprintf("%d proposal group(s) omitted after semantic proposal retry exhaustion", discardedGroups)
+		message = fmt.Sprintf("%d proposal group(s) omitted from the final semantic proposal; all affected candidates were preserved separately", discardedGroups)
+	}
+	if len(details) > 0 {
+		message = diagnostics.Aggregate(message, details)
 	}
 	return diagnostics.Finding{Scope: "npcs", ReasonCode: ReasonCodeNPCSemanticReconciliationExhausted, Message: message}
 }
@@ -184,7 +196,7 @@ func fallbackResult(value dnd.NPCRegistry, findings []diagnostics.Finding, fallb
 	return result, nil
 }
 
-func retryResultWithFallback(value dnd.NPCRegistry, findings []diagnostics.Finding, candidate *contracts.ModelCandidate, message string, fallback diagnostics.Finding) (contracts.TypedNormalizeResult[dnd.NPCRegistry], error) {
+func retryResultWithFallback(value dnd.NPCRegistry, findings []diagnostics.Finding, candidate *contracts.ModelCandidate, correctionGuidance, message string, fallback diagnostics.Finding) (contracts.TypedNormalizeResult[dnd.NPCRegistry], error) {
 	result, err := normalizationResult(value, findings, candidate)
 	if err != nil {
 		return contracts.TypedNormalizeResult[dnd.NPCRegistry]{}, err
@@ -193,7 +205,7 @@ func retryResultWithFallback(value dnd.NPCRegistry, findings []diagnostics.Findi
 	if err != nil {
 		return contracts.TypedNormalizeResult[dnd.NPCRegistry]{}, normalizerErrorf("collect fallback diagnostic: %w", err)
 	}
-	result.Retry = &contracts.NormalizeRetry{ReasonCode: ReasonCodeNPCSemanticProposalInvalid, Message: message, FallbackDiagnostics: fallbackGroups}
+	result.Retry = &contracts.NormalizeRetry{ReasonCode: ReasonCodeNPCSemanticProposalInvalid, Message: message, CorrectionGuidance: correctionGuidance, FallbackDiagnostics: fallbackGroups}
 	return result, nil
 }
 

internal/modules/dnd/normalize/npcregistry/semantic_normalizer_test.go

@@ -115,10 +115,13 @@ func TestNormalizeUnsafeProposalReturnsSafeRetryFallback(t *testing.T) {
 	if result.Retry.ReasonCode != ReasonCodeNPCSemanticProposalInvalid || !strings.Contains(result.Retry.Message, "group 1: member_unknown") {
 		t.Fatalf("retry = %#v, want bounded invalid-proposal diagnostics", result.Retry)
 	}
+	if !strings.Contains(result.Retry.CorrectionGuidance, "Duplicate group 2") || !strings.Contains(result.Retry.CorrectionGuidance, "supplied candidate list") || strings.Contains(result.Retry.CorrectionGuidance, "member_unknown") || strings.Contains(result.Retry.CorrectionGuidance, ReasonCodeNPCSemanticProposalInvalid) {
+		t.Fatalf("correction guidance = %q, want contextual prose without internal labels", result.Retry.CorrectionGuidance)
+	}
 	if len(result.Value.NPCs) != 2 || result.Value.NPCs[0].Name != "Mira Thorn" || result.Value.NPCs[1].Name != "Captain Vale" {
 		t.Fatalf("fallback NPCs = %#v, want independently safe group applied", result.Value.NPCs)
 	}
-	if len(result.Retry.FallbackDiagnostics) != 1 || result.Retry.FallbackDiagnostics[0].ReasonCode != ReasonCodeNPCSemanticReconciliationExhausted || !strings.Contains(result.Retry.FallbackDiagnostics[0].Samples[0].Message, "1 proposal group") {
+	if len(result.Retry.FallbackDiagnostics) != 1 || result.Retry.FallbackDiagnostics[0].ReasonCode != ReasonCodeNPCSemanticReconciliationExhausted || !strings.Contains(result.Retry.FallbackDiagnostics[0].Samples[0].Message, "1 proposal group") || !strings.Contains(result.Retry.FallbackDiagnostics[0].Samples[0].Message, "preserved separately") || !strings.Contains(result.Retry.FallbackDiagnostics[0].Samples[0].Message, "supplied candidate list") {
 		t.Fatalf("fallback diagnostics = %#v, want exact omitted-group warning", result.Retry.FallbackDiagnostics)
 	}
 }
@@ -134,6 +137,9 @@ func TestNormalizeInvalidStructuredOutputAndOperationalErrorsRemainDistinct(t *t
 	if err != nil || result.Retry == nil || result.Retry.ReasonCode != ReasonCodeNPCSemanticProposalInvalid || len(result.Value.NPCs) != 2 {
 		t.Fatalf("invalid structured result = %#v, %v; want deterministic retry fallback", result, err)
 	}
+	if result.Retry.CorrectionGuidance != "" || !strings.Contains(result.Retry.FallbackDiagnostics[0].Samples[0].Message, "structured-output contract") || !strings.Contains(result.Retry.FallbackDiagnostics[0].Samples[0].Message, "preserved separately") {
+		t.Fatalf("invalid structured retry = %#v, want feedback-free retry and informative fallback", result.Retry)
+	}
 	operational := errors.New("provider unavailable")
 	if _, err := newNormalizer(t, &recordingNPCNormalizerClient{err: operational}).Normalize(context.Background(), normalizeRequestWithSource(input, doc)); !errors.Is(err, operational) || errors.Is(err, contracts.ErrInvalidStructuredOutput) {
 		t.Fatalf("operational completion error = %v, want ordinary error", err)