fix

apps/api/cmd/mock-llm/main.go

@@ -31,14 +31,15 @@ import (
 )
 
 type server struct {
-	apiKey string
-	model  string
-	logReq atomic.Int64
+	apiKey               string
+	model                string
+	categorizeConfidence float64
+	logReq               atomic.Int64
 }
 
 type chatRequest struct {
-	Model       string `json:"model"`
-	Messages    []struct {
+	Model    string `json:"model"`
+	Messages []struct {
 		Role    string `json:"role"`
 		Content string `json:"content"`
 	} `json:"messages"`
@@ -50,11 +51,16 @@ func main() {
 	addr := flag.String("addr", "127.0.0.1:18767", "listen address")
 	key := flag.String("key", envOr("MOCK_LLM_API_KEY", "local-test"), "Bearer API key (non-empty placeholder)")
 	model := flag.String("model", envOr("MOCK_LLM_MODEL", "mock-llm"), "model id returned by /v1/models")
+	// Lets a local run reproduce a low-confidence taxonomy pick (see
+	// processing.DefaultMinCategorizeConfidence).
+	categorizeConfidence := flag.Float64("categorize-confidence", 0,
+		"when > 0, answer categorize calls with this confidence score")
 	flag.Parse()
 
 	s := &server{
-		apiKey: strings.TrimSpace(*key),
-		model:  strings.TrimSpace(*model),
+		apiKey:               strings.TrimSpace(*key),
+		model:                strings.TrimSpace(*model),
+		categorizeConfidence: *categorizeConfidence,
 	}
 	if s.apiKey == "" {
 		log.Fatal("mock-llm: API key must be non-empty (Descrybe Completer.Enabled requires it)")
@@ -77,6 +83,13 @@ func main() {
 	}
 }
 
+func truncTrace(s string, n int) string {
+	if len(s) <= n {
+		return s
+	}
+	return s[:n]
+}
+
 func envOr(k, def string) string {
 	if v := strings.TrimSpace(os.Getenv(k)); v != "" {
 		return v
@@ -140,11 +153,17 @@ func (s *server) handleChatCompletions(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 	system, user := splitMessages(req.Messages)
+	if os.Getenv("MOCK_LLM_TRACE") != "" {
+		log.Printf("mock-llm: roles=%d system_len=%d user_len=%d system_head=%q",
+			len(req.Messages), len(system), len(user), truncTrace(system, 60))
+	}
 	var comp processing.Completion
 	// A category formula in the prompt is answered in that formula's shape, so a
 	// local run proves the formula reached the model and the reply passes the
 	// pipeline's formula gate. Everything else keeps the heuristic fallback.
-	if text, ok := buildFormulaReply(system, user); ok {
+	if text, ok := s.categorizeReply(system, user); ok {
+		comp = processing.Completion{Text: text}
+	} else if text, ok := buildFormulaReply(system, user); ok {
 		comp = processing.Completion{Text: text}
 	} else {
 		var err error
@@ -204,6 +223,39 @@ func (s *server) handleEmbeddings(w http.ResponseWriter, r *http.Request) {
 	})
 }
 
+// categorizeReply overrides the heuristic taxonomy pick so a local run can
+// reproduce a low-confidence answer.
+func (s *server) categorizeReply(system, user string) (string, bool) {
+	if s.categorizeConfidence <= 0 || !strings.Contains(strings.ToLower(system), "categoryid") {
+		return "", false
+	}
+	id := firstCategoryIDFromPrompt(user)
+	if id == "" {
+		return "", false
+	}
+	b, err := json.Marshal(map[string]any{"categoryId": id, "confidence": s.categorizeConfidence})
+	if err != nil {
+		return "", false
+	}
+	return string(b), true
+}
+
+// firstCategoryIDFromPrompt mirrors the "(ID: …)" format of the categorize prompt.
+func firstCategoryIDFromPrompt(user string) string {
+	const marker = "(id:"
+	lower := strings.ToLower(user)
+	at := strings.Index(lower, marker)
+	if at < 0 {
+		return ""
+	}
+	rest := strings.TrimSpace(user[at+len(marker):])
+	end := strings.Index(rest, ")")
+	if end <= 0 {
+		return ""
+	}
+	return strings.TrimSpace(rest[:end])
+}
+
 func splitMessages(msgs []struct {
 	Role    string `json:"role"`
 	Content string `json:"content"`
@@ -211,7 +263,10 @@ func splitMessages(msgs []struct {
 	var users []string
 	for _, m := range msgs {
 		switch strings.ToLower(strings.TrimSpace(m.Role)) {
-		case "system":
+		// Reasoning-model requests send the system message as "developer"
+		// (see processing.OpenAIClient sysRole). Treating it as a user turn made the
+		// mock mis-route every call whose branch depends on the system prompt.
+		case "system", "developer":
 			if system == "" {
 				system = m.Content
 			} else {

apps/api/internal/processing/categorize_ai.go

@@ -2,17 +2,36 @@ package processing
 
 import (
 	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
 	"log"
 	"sort"
+	"strconv"
 	"strings"
 
 	"github.com/descrybe/descrybe-v2/apps/api/internal/aiaudit"
 )
 
+// ErrCategoryNotFound fails a product whose category could not be determined:
+// the model's pick was outside the taxonomy, or it scored below the confidence
+// floor. Processing stops for that product instead of enhancing it under a
+// category nobody believes in.
+var ErrCategoryNotFound = errors.New("product category could not be determined")
+
 // MaxCategorizeOptions caps the taxonomy list injected into the categorize LLM
 // prompt (token/cost bound). Prefer stable sort by display name then unique_id.
 const MaxCategorizeOptions = 200
 
+// DefaultMinCategorizeConfidence is the floor for accepting an LLM category pick.
+//
+// The model always answers with SOME id from the list, so a low-confidence reply
+// like {"categoryId":"1","confidence":0.08} is the model saying "I do not know" —
+// taking it at face value files a camping chair under Generators and then drives
+// that category's title/description formula, producing confidently wrong copy.
+// Below this floor the product fails with ErrCategoryNotFound instead.
+const DefaultMinCategorizeConfidence = 0.75
+
 // MaxTokensCategorize budgets a short JSON reply {"categoryId","confidence"}.
 // code-fast / reasoning-style models often burn internal tokens before JSON;
 // 256 frequently finishes with empty content and forces a length-cap retry.
@@ -116,15 +135,15 @@ func ProductCategorizeUser(name, description string, opts []categoryOption) stri
 // tryAICategorize asks the Completer to pick a company taxonomy unique_id when
 // Category is still empty after mapped/prior/vector. Never invents outside taxonomy:
 // responses are coerced then filtered to namesByUID / valid unique_ids.
-func tryAICategorize(ctx context.Context, e *Engine, out *StepResult, in ProductInput, policy StepPolicy) {
+func tryAICategorize(ctx context.Context, e *Engine, out *StepResult, in ProductInput, policy StepPolicy) error {
 	if out == nil || strings.TrimSpace(out.Category) != "" {
-		return
+		return nil
 	}
 	if !policy.AllowAI {
-		return
+		return nil
 	}
 	if e == nil || !e.CompleterEnabled() {
-		return
+		return nil
 	}
 	opts := categoryOptionsFromNames(in.CategoryNamesByUID)
 	if len(opts) == 0 {
@@ -133,7 +152,7 @@ func tryAICategorize(ctx context.Context, e *Engine, out *StepResult, in Product
 			"status": "skipped",
 			"reason": "no_taxonomy",
 		})
-		return
+		return nil
 	}
 	name := strings.TrimSpace(out.Name)
 	if name == "" {
@@ -149,7 +168,7 @@ func tryAICategorize(ctx context.Context, e *Engine, out *StepResult, in Product
 			"status": "skipped",
 			"reason": "empty_product",
 		})
-		return
+		return nil
 	}
 
 	system := ProductCategorizeSystem
@@ -166,23 +185,44 @@ func tryAICategorize(ctx context.Context, e *Engine, out *StepResult, in Product
 			"error":  TruncateError(err),
 			"tokens": comp.TotalTokens,
 		})
-		return
+		return nil
 	}
 
 	raw := categoryIDFromCategorizeJSON(obj)
+	confidence, hasConfidence := categorizeConfidenceFromJSON(obj)
 	valid := make(map[string]struct{}, len(opts))
 	for _, o := range opts {
 		valid[o.UniqueID] = struct{}{}
 	}
 	resolved := resolveCompanyCategoryUniqueID(raw, in.CategoryNamesByUID, valid)
 	if resolved == "" || isUnusableCategoryValue(resolved, out.ProcessedName, out.Name) {
-		out.Notes = append(out.Notes, "ai_categorize: ignored (not in company taxonomy)")
+		out.Notes = append(out.Notes, "ai_categorize: no category found (not in company taxonomy)")
 		appendStepLog(out.GPTResponse, StepCategorize, map[string]any{
 			"status":   "rejected",
 			"returned": raw,
 			"tokens":   comp.TotalTokens,
 		})
-		return
+		return fmt.Errorf("%w: model returned %q, which is not in the company taxonomy",
+			ErrCategoryNotFound, truncateRunes(raw, 60))
+	}
+
+	// A confident-looking id with a low score is the model guessing. Refuse it
+	// rather than driving the wrong category's formula with it.
+	if min := policy.CategorizeConfidence(); hasConfidence && confidence < min {
+		out.Notes = append(out.Notes, fmt.Sprintf(
+			"ai_categorize: no category found (confidence %.2f below %.2f)", confidence, min))
+		appendStepLog(out.GPTResponse, StepCategorize, map[string]any{
+			"status":     "low_confidence",
+			"returned":   raw,
+			"resolved":   resolved,
+			"confidence": confidence,
+			"minimum":    min,
+			"tokens":     comp.TotalTokens,
+		})
+		log.Printf("processing: category rejected uid=%s confidence=%.2f min=%.2f reason=low_confidence",
+			resolved, confidence, min)
+		return fmt.Errorf("%w: best match %q scored %.2f, below the %.2f minimum",
+			ErrCategoryNotFound, truncateRunes(resolved, 60), confidence, min)
 	}
 
 	out.Category = SanitizeText(resolved)
@@ -201,6 +241,50 @@ func tryAICategorize(ctx context.Context, e *Engine, out *StepResult, in Product
 		"tokens":   comp.TotalTokens,
 		"options":  len(opts),
 	})
+	if hasConfidence {
+		out.Notes = append(out.Notes, fmt.Sprintf("ai_categorize: confidence %.2f", confidence))
+	}
+	return nil
+}
+
+// categorizeConfidenceFromJSON reads the model's self-reported score. ok=false when
+// the field is missing or unparseable — a model that reports nothing is not
+// penalised, only one that reports a low score.
+func categorizeConfidenceFromJSON(obj map[string]any) (float64, bool) {
+	if obj == nil {
+		return 0, false
+	}
+	for _, k := range []string{"confidence", "score", "certainty"} {
+		switch v := obj[k].(type) {
+		case float64:
+			return clampConfidence(v), true
+		case int:
+			return clampConfidence(float64(v)), true
+		case json.Number:
+			if f, err := v.Float64(); err == nil {
+				return clampConfidence(f), true
+			}
+		case string:
+			if f, err := strconv.ParseFloat(strings.TrimSpace(v), 64); err == nil {
+				return clampConfidence(f), true
+			}
+		}
+	}
+	return 0, false
+}
+
+// clampConfidence folds a 0-100 style score onto 0-1 and bounds it.
+func clampConfidence(v float64) float64 {
+	if v > 1 && v <= 100 {
+		v /= 100
+	}
+	if v < 0 {
+		return 0
+	}
+	if v > 1 {
+		return 1
+	}
+	return v
 }
 
 func categoryIDFromCategorizeJSON(obj map[string]any) string {
@@ -233,9 +317,12 @@ func firstAvailableCategoryIDFromPrompt(user string) string {
 }
 
 // runCategorizeStep applies vector then LLM taxonomy selection when Category empty.
-func runCategorizeStep(ctx context.Context, e *Engine, companyID string, out *StepResult, in ProductInput, categoryNames []string, policy StepPolicy) {
+// runCategorizeStep returns ErrCategoryNotFound when the model was asked for a
+// category and could not give a usable one — the caller fails the product rather
+// than enhancing it under a guess.
+func runCategorizeStep(ctx context.Context, e *Engine, companyID string, out *StepResult, in ProductInput, categoryNames []string, policy StepPolicy) error {
 	if out == nil {
-		return
+		return nil
 	}
 	if strings.TrimSpace(out.Category) != "" {
 		appendStepLog(out.GPTResponse, StepCategorize, map[string]any{
@@ -243,7 +330,7 @@ func runCategorizeStep(ctx context.Context, e *Engine, companyID string, out *St
 			"reason":   "already_set",
 			"category": out.Category,
 		})
-		return
+		return nil
 	}
 	if !policy.AllowAI {
 		out.Notes = append(out.Notes, "categorize: skipped (AI not allowed)")
@@ -251,21 +338,24 @@ func runCategorizeStep(ctx context.Context, e *Engine, companyID string, out *St
 			"status": "skipped",
 			"reason": "entitlement_can_use_ai",
 		})
-		return
+		return nil
 	}
 	tryVectorCategorize(ctx, e, companyID, out, categoryNames, policy)
 	if strings.TrimSpace(out.Category) != "" {
-		return
+		return nil
 	}
 	// Taxonomy picks and copy generation share one Completer; tag the role so the
 	// admin inspector can tell them apart.
-	tryAICategorize(aiaudit.WithCall(ctx, aiaudit.CallContext{Role: aiaudit.RoleCategorize}), e, out, in, policy)
+	if err := tryAICategorize(aiaudit.WithCall(ctx, aiaudit.CallContext{Role: aiaudit.RoleCategorize}), e, out, in, policy); err != nil {
+		return err
+	}
 	if strings.TrimSpace(out.Category) == "" {
 		appendStepLog(out.GPTResponse, StepCategorize, map[string]any{
 			"status": "unset",
 			"reason": "no_vector_or_ai_match",
 		})
 	}
+	return nil
 }
 
 // persistMappedCategorySQL writes a taxonomy unique_id onto mapped_data.category

apps/api/internal/processing/categorize_ai_test.go

@@ -2,6 +2,7 @@ package processing
 
 import (
 	"context"
+	"errors"
 	"strings"
 	"testing"
 )
@@ -62,28 +63,24 @@ func TestTryAICategorize_rejectsInventedID(t *testing.T) {
 		}},
 		Vector: NoopVectorCategorizer{},
 	}
+	// An id outside the taxonomy means the category could not be determined. The
+	// product must fail rather than continue uncategorised into enhance, which
+	// would generate copy with no formula behind it.
 	out, err := e.RunSteps(context.Background(), "co", ProductInput{
 		Name: "Widget",
 		CategoryNamesByUID: map[string]string{
 			"50": "Štedilniki",
 		},
 	}, "full", nil, StepPolicy{AllowAI: true})
-	if err != nil {
-		t.Fatal(err)
+	if !errors.Is(err, ErrCategoryNotFound) {
+		t.Fatalf("err=%v want ErrCategoryNotFound", err)
+	}
+	if !strings.Contains(err.Error(), "99999") {
+		t.Fatalf("error should name the rejected id, got %v", err)
 	}
 	if out.Category != "" {
 		t.Fatalf("Category=%q want empty (invented id rejected)", out.Category)
 	}
-	found := false
-	for _, n := range out.Notes {
-		if strings.Contains(n, "ai_categorize: ignored") {
-			found = true
-			break
-		}
-	}
-	if !found {
-		t.Fatalf("expected reject note, got %v", out.Notes)
-	}
 }
 
 func TestTryAICategorize_resolvesDisplayName(t *testing.T) {

apps/api/internal/processing/categorize_confidence_test.go

@@ -0,0 +1,176 @@
+package processing
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"strings"
+	"testing"
+)
+
+func categorizeEngine(reply string) (*Engine, *int) {
+	enhanceCalls := 0
+	e := &Engine{
+		Completer: stubCompleter{fn: func(system, user string) (Completion, error) {
+			if strings.Contains(strings.ToLower(system), "categoryid") {
+				return Completion{Text: reply, TotalTokens: 3}, nil
+			}
+			enhanceCalls++
+			return Completion{Text: `{"name":"Enhanced","description":"Body"}`, TotalTokens: 5}, nil
+		}},
+		Vector: NoopVectorCategorizer{},
+	}
+	return e, &enhanceCalls
+}
+
+func categorizeInput() ProductInput {
+	return ProductInput{
+		Name:        "BRUNNER folding camping chair ONE SHOT",
+		Description: "A light folding aluminium chair for outdoor use.",
+		Mapped: map[string]any{
+			"name":        "BRUNNER folding camping chair ONE SHOT",
+			"description": "A light folding aluminium chair for outdoor use.",
+		},
+		CategoryNamesByUID: map[string]string{"1": "Generators", "34": "Smartwatches"},
+	}
+}
+
+// The reported case: the model always answers with SOME id, so a low score is it
+// saying "I do not know". Taking it at face value files a camping chair under
+// Generators and then drives that category's formula.
+func TestRunSteps_lowConfidenceCategoryFailsTheProduct(t *testing.T) {
+	t.Parallel()
+	e, enhanceCalls := categorizeEngine(`{"categoryId":"1","confidence":0.08}`)
+
+	out, err := e.RunSteps(context.Background(), "co", categorizeInput(), "full", nil,
+		StepPolicy{AllowAI: true})
+	if !errors.Is(err, ErrCategoryNotFound) {
+		t.Fatalf("err=%v want ErrCategoryNotFound", err)
+	}
+	for _, want := range []string{"0.08", "0.75"} {
+		if !strings.Contains(err.Error(), want) {
+			t.Fatalf("error should report the score and the minimum, got %v", err)
+		}
+	}
+	if out.Category != "" {
+		t.Fatalf("a rejected pick must not be persisted, got %q", out.Category)
+	}
+	// The whole point: stop before spending an enhance call on the wrong category.
+	if *enhanceCalls != 0 {
+		t.Fatalf("enhance ran %d time(s) after the category was rejected", *enhanceCalls)
+	}
+}
+
+func TestRunSteps_confidentCategoryProceeds(t *testing.T) {
+	t.Parallel()
+	e, enhanceCalls := categorizeEngine(`{"categoryId":"1","confidence":0.92}`)
+
+	out, err := e.RunSteps(context.Background(), "co", categorizeInput(), "full", nil,
+		StepPolicy{AllowAI: true})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if out.Category != "1" {
+		t.Fatalf("Category=%q want 1", out.Category)
+	}
+	if *enhanceCalls == 0 {
+		t.Fatal("enhance should run for a confident category")
+	}
+}
+
+// Exactly at the floor is acceptable — the threshold is a minimum, not a bar to clear.
+func TestRunSteps_confidenceAtThresholdIsAccepted(t *testing.T) {
+	t.Parallel()
+	e, _ := categorizeEngine(fmt.Sprintf(`{"categoryId":"34","confidence":%v}`, DefaultMinCategorizeConfidence))
+	out, err := e.RunSteps(context.Background(), "co", categorizeInput(), "full", nil,
+		StepPolicy{AllowAI: true})
+	if err != nil {
+		t.Fatalf("confidence == minimum must pass, got %v", err)
+	}
+	if out.Category != "34" {
+		t.Fatalf("Category=%q want 34", out.Category)
+	}
+}
+
+// A model that reports no score is not penalised — only one that reports a low one.
+func TestRunSteps_missingConfidenceIsAccepted(t *testing.T) {
+	t.Parallel()
+	e, _ := categorizeEngine(`{"categoryId":"34"}`)
+	out, err := e.RunSteps(context.Background(), "co", categorizeInput(), "full", nil,
+		StepPolicy{AllowAI: true})
+	if err != nil {
+		t.Fatalf("missing confidence must not fail the product, got %v", err)
+	}
+	if out.Category != "34" {
+		t.Fatalf("Category=%q want 34", out.Category)
+	}
+}
+
+func TestStepPolicy_confidenceOverride(t *testing.T) {
+	t.Parallel()
+	if got := (StepPolicy{}).CategorizeConfidence(); got != DefaultMinCategorizeConfidence {
+		t.Fatalf("default = %v want %v", got, DefaultMinCategorizeConfidence)
+	}
+	if got := (StepPolicy{MinCategorizeConfidence: 0.4}).CategorizeConfidence(); got != 0.4 {
+		t.Fatalf("override = %v want 0.4", got)
+	}
+	// A tenant that lowers the bar accepts what the default would reject.
+	e, _ := categorizeEngine(`{"categoryId":"1","confidence":0.5}`)
+	if _, err := e.RunSteps(context.Background(), "co", categorizeInput(), "full", nil,
+		StepPolicy{AllowAI: true, MinCategorizeConfidence: 0.4}); err != nil {
+		t.Fatalf("0.5 should pass a 0.4 floor, got %v", err)
+	}
+}
+
+// A category already resolved from the feed is never second-guessed: categorize
+// does not run, so its confidence cannot fail the product.
+func TestRunSteps_mappedCategorySkipsConfidenceGate(t *testing.T) {
+	t.Parallel()
+	e, _ := categorizeEngine(`{"categoryId":"1","confidence":0.01}`)
+	in := categorizeInput()
+	in.Mapped["category"] = "34"
+
+	out, err := e.RunSteps(context.Background(), "co", in, "full", nil, StepPolicy{AllowAI: true})
+	if err != nil {
+		t.Fatalf("a feed-supplied category must not be gated, got %v", err)
+	}
+	if out.Category != "34" {
+		t.Fatalf("Category=%q want 34", out.Category)
+	}
+}
+
+// Free plan / AI disabled must keep working — nothing was asked of a model, so
+// there is no failed categorisation to report.
+func TestRunSteps_noAIDoesNotFailOnMissingCategory(t *testing.T) {
+	t.Parallel()
+	e, _ := categorizeEngine(`{"categoryId":"1","confidence":0.01}`)
+	if _, err := e.RunSteps(context.Background(), "co", categorizeInput(), "full", nil,
+		StepPolicy{AllowAI: false}); err != nil {
+		t.Fatalf("AI-disabled processing must not fail on category, got %v", err)
+	}
+}
+
+func TestCategorizeConfidenceFromJSON(t *testing.T) {
+	t.Parallel()
+	cases := []struct {
+		obj  map[string]any
+		want float64
+		ok   bool
+	}{
+		{map[string]any{"confidence": 0.08}, 0.08, true},
+		{map[string]any{"confidence": "0.42"}, 0.42, true},
+		{map[string]any{"score": 0.9}, 0.9, true},
+		// Some models answer on a 0-100 scale.
+		{map[string]any{"confidence": 85.0}, 0.85, true},
+		{map[string]any{"confidence": 1}, 1, true},
+		{map[string]any{"categoryId": "1"}, 0, false},
+		{map[string]any{"confidence": "high"}, 0, false},
+		{nil, 0, false},
+	}
+	for _, c := range cases {
+		got, ok := categorizeConfidenceFromJSON(c.obj)
+		if ok != c.ok || (ok && got != c.want) {
+			t.Fatalf("%v → (%v, %v) want (%v, %v)", c.obj, got, ok, c.want, c.ok)
+		}
+	}
+}

apps/api/internal/processing/steps.go

@@ -19,6 +19,18 @@ import (
 type StepPolicy struct {
 	AllowAI    bool
 	AllowEPREL bool
+	// MinCategorizeConfidence overrides DefaultMinCategorizeConfidence. A categorize
+	// reply below it is treated as "no category found" and fails the product rather
+	// than letting a wrong guess drive the formula. 0 uses the default.
+	MinCategorizeConfidence float64
+}
+
+// CategorizeConfidence returns the effective low-confidence floor for this policy.
+func (p StepPolicy) CategorizeConfidence() float64 {
+	if p.MinCategorizeConfidence > 0 {
+		return p.MinCategorizeConfidence
+	}
+	return DefaultMinCategorizeConfidence
 }
 
 // RunSteps executes the multi-step product pipeline.
@@ -213,7 +225,11 @@ func (e *Engine) RunSteps(ctx context.Context, companyID string, in ProductInput
 
 		case StepCategorize:
 			// Vector (if enabled) then LLM taxonomy pick when category still empty.
-			runCategorizeStep(ctx, e, companyID, &out, in, categoryNames, policy)
+			// A category the model could not determine fails the product here, before
+			// enhance spends a call writing copy for the wrong category.
+			if err := runCategorizeStep(ctx, e, companyID, &out, in, categoryNames, policy); err != nil {
+				return out, err
+			}
 
 		case StepAIEnhance:
 			preservePriorEnhanceHash := func() {
@@ -533,7 +549,9 @@ func (e *Engine) RunSteps(ctx context.Context, companyID string, in ProductInput
 	// Pipelines without StepCategorize (normalize_only) still run vector + LLM
 	// categorize when category is empty so downstream steps are not fed a blank.
 	if !stepsContain(steps, StepCategorize) {
-		runCategorizeStep(ctx, e, companyID, &out, in, categoryNames, policy)
+		if err := runCategorizeStep(ctx, e, companyID, &out, in, categoryNames, policy); err != nil {
+			return out, err
+		}
 	}
 	// Record why category stayed empty for every pipeline, including the ones that
 	// now categorize inside the loop (enhance / title / description).

docs/category-formulas.md

@@ -78,6 +78,31 @@ go run ./cmd/formula-e2e -new-tenant -llm-base http://127.0.0.1:18767/v1
 `-llm-base` pins the provider for the run; without it the harness uses whatever
 `platformsettings.ResolveOpenAI` returns (admin DB setting first, then env).
 
+## Low-confidence categorisation fails the product
+
+The categorize model always answers with *some* id from the list, so a reply like
+`{"categoryId":"1","confidence":0.08}` is the model saying "I do not know". Taking
+it at face value files a camping chair under Generators and then drives that
+category's formula, producing confidently wrong copy.
+
+Below `processing.DefaultMinCategorizeConfidence` (**0.75**) the product fails with
+`ErrCategoryNotFound` and processing stops there — enhance is never called, so no
+credits and no provider cost are spent writing copy for a category nobody believes
+in. The same applies when the model returns an id outside the taxonomy.
+
+```
+processing: category rejected uid=1 confidence=0.08 min=0.75 reason=low_confidence
+processing: item failed ... err=product category could not be determined:
+            best match "1" scored 0.08, below the 0.75 minimum
+```
+
+The product lands in the job as `failed` with that message on
+`processing_job_products.error`, and `raw_products.processing_status = 'failed'`.
+
+Not gated: a category that came from the feed, vector or a prior run (categorize
+never runs), a model that reports no confidence at all, and AI-disabled plans.
+Override the floor per job with `StepPolicy.MinCategorizeConfidence`.
+
 ## Why enhance can still return feed copy
 
 In order of how often it bites: