AI classifier correction notes + misread review (AI tab)

Add a global, temporal ai_notes list appended to the classifier prompt
(seeded once from no-PII defaults, documented in README), managed inline
on a new AI tab with a read-only view of the assembled prompt. Every
AI-run upload records the browser-round-tripped suggestion blob + model;
misreads are derived (final field != AI guess) and reviewed one by one
(image + per-field guess-vs-entered + notes-since), attributing which
note fixed each or closing unresolved. Update SPEC (new section 10),
DESIGN item 15, README, and changelog (0.0.3).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

CHANGELOG.md

@@ -4,6 +4,22 @@ All notable changes to this project are documented here. Versions are git tags;
 release tags `X.Y.Z` are built and deployed automatically (pre-release tags such
 as `0.0.0a1` are built and staged only).
 
+## [0.0.3] - 2026-06-20
+
+### Added
+- AI classifier correction notes + misread review (new "AI" tab).
+  - A global list of free-text correction notes is appended to the classifier
+    prompt; managed inline (add/edit/delete). The notes table is temporal (edits
+    soft-delete + insert), seeded once from a no-PII default list (documented in the
+    README), and read live per classification.
+  - A read-only view of the exact assembled prompt.
+  - Every AI-run upload records the suggestion blob + model (round-tripped from the
+    browser). Misreads — where a final field differs from the AI's guess — are
+    derived and listed for review; reviewing shows the image and per-field
+    guess-vs-entered, and lets the user attribute which note(s) fixed it (or close
+    it unresolved).
+  - See SPEC.md §10 and DESIGN.md item 15.
+
 ## [0.0.2] - 2026-06-20
 
 ### Added

DESIGN.md

@@ -456,4 +456,98 @@ Out of scope (for now):
     accumulate cruft and a cleanup surface (akin to item 6 for people) will be
     wanted eventually, but not in this item.
   - Adding/removing tags on an already-saved receipt (no edit/detail page yet, same
-    limitation as attachments in item 10).
+    limitation as attachments in item 10).
+
+15. AI classifier notes + failure review (implemented — see SPEC.md §10)
+
+A closed loop for improving receipt classification over time: the user accumulates
+corrective instructions ("notes") that are appended to the classifier prompt, and
+reviews past misreads one by one to author and attribute those fixes. Big change;
+spans storage, the upload path, the classifier prompt, and a new top-level tab.
+
+Motivation: classification will misread some receipts (a vendor's odd date format,
+a statement that lists the patient under "Guarantor", etc.). Rather than hardcode
+ever-more rules, let the two users add their own corrections as they hit failures,
+and give them a place to study failures and decide what fixed them.
+
+A. Notes — what they are
+  - A single, GLOBAL, free-text list of correction lines, appended to the system
+    prompt as a "Corrections/Notes" appendix. Global because at classify time the
+    app does not yet know the category/vendor, so scoped notes couldn't be selected.
+  - Authored English, separate from the prompt's DERIVED parts (people + name
+    variants, category names/examples, today's date), which stay computed in code.
+    The static prompt scaffolding (the rules/warnings) stays in code too — notes
+    only AUGMENT it; this item does not externalize the whole prompt.
+  - Curated, not append-forever: each active note costs tokens on every scan and too
+    many dilute the instructions, so delete/edit matter as much as add. Realistically
+    a handful.
+
+B. Notes — storage (temporal)
+  - Table ai_notes(id, text, created_at, deleted_at). Active set = deleted_at IS
+    NULL, ordered by created_at; that set is what gets appended to the prompt.
+  - EDIT = soft-delete the old row + insert a new one (never update in place), so the
+    full history is preserved. The notes active at any time T are
+    created_at <= T AND (deleted_at IS NULL OR deleted_at > T) — which is what lets a
+    failure be matched against the notes that existed when it happened (section E).
+  - Read per /classify call (live; no restart needed). Rides along in /export/db and
+    the daily backup like everything else.
+  - SEEDED ONCE from a hardcoded default list in code (the same defaults are
+    published in the README for humans), inserted only when the table is completely
+    empty — so a deliberately-deleted default does not resurrect on restart. Never
+    read from a file/config at runtime; no export/mirror file. Defaults carry no PII;
+    user-authored notes live only in the DB (private, not in git).
+
+C. Capturing classifications (the plumbing)
+  - The /classify result is produced server-side but only reaches the browser (to
+    pre-fill the form); by POST /upload the server no longer holds it. So the browser
+    PERSISTS the suggestion blob + model and sends it back in a hidden field on
+    submit. (The API key is and remains server-side — only the suggestions round-trip.)
+  - Store one row per AI-run upload:
+    classifications(receipt_id PK/FK, model, response_json, created_at,
+                    reviewed BOOL, reviewed_at, resolution).
+    Skip-AI or classification-disabled uploads create NO row.
+  - response_json is the /classify blob (suggested amount/date/category_id/person_id,
+    the raw_* text the model read, cost, model). Treat it as DIAGNOSTIC data, not
+    ground truth — it is client-supplied and could be tampered with; nothing
+    security-relevant depends on it. Mild PII (raw read name) but DB-only.
+
+D. Failures are DERIVED, not stored
+  - A "failure"/miss = at least one of the 4 AI-suggested fields (amount, date,
+    category, who) differs from the receipt's final stored value, where AI returning
+    null/empty and the user filling it COUNTS as an override (the AI missed it).
+  - Receipts are immutable (no edit) and the blob is immutable, so the comparison
+    inputs never change — derived overrides can never drift, so we store no
+    redundant per-field booleans. Per-field detail (which of the 4, AI-guess vs.
+    corrected) is computed on demand from the blob vs. the receipt when a failure is
+    displayed. Volume is tiny, so deriving each time is cheap. (If SQL-level
+    filtering/stats ever matter, a single computed had_override flag could be added
+    purely as an index — not the four booleans.)
+  - Review state (reviewed, reviewed_at, resolution) and the fix links below are the
+    only non-derivable things, so they ARE stored.
+
+E. Review — the loop
+  - New top-level "AI" tab with three sections (one tab unless it grows):
+      1. Notes — list active notes; add / edit / delete (each edit = soft-delete +
+         insert per section B).
+      2. Prompt — READ-ONLY view of the live assembled system prompt (static
+         scaffolding + injected people/categories/today + current active notes), so
+         the user sees exactly what is sent. View-only for now; not editable.
+      3. Review — unreviewed failures, one at a time.
+  - Reviewing one failure shows: the receipt IMAGE, the MODEL used, which fields were
+    overridden (AI guess vs. corrected, derived from the blob), and the NOTES ADDED
+    SINCE this failure's created_at (from the ai_notes temporal history). The user
+    marks which of those notes were the fix → recorded in miss_fixes(receipt_id,
+    note_id); if none fixed it, close as resolution='unresolved'. Either way set
+    reviewed=1, reviewed_at. ("Switched to a stronger model" as a resolution reason
+    may be added later, since model is recorded.)
+
+F. Out of scope / future
+  - Per-category or vendor-scoped notes (global only, by design above).
+  - Automatic re-classification/recompute of past failures. The data supports it
+    (model + active-notes-at-time-T + the stored image), but review stays MANUAL —
+    the user eyeballs the image to understand the miss.
+  - Editing the static prompt scaffolding (view-only here) and externalizing the
+    whole prompt.
+  - Accuracy-rate stats across all classifications.
+  - No prompting the user to write a note at upload time — failures are recorded
+    silently and dealt with later in the Review tab (likely on a real computer).

README.md

@@ -0,0 +1,35 @@
+# HSA Receipt Tracker
+
+A small, mobile-first web app for two household users to capture and archive
+HSA-eligible receipts (photo or PDF) for future reimbursement and tax
+substantiation, with optional AI auto-fill of the amount/date/category/patient.
+
+- **What it does (current behavior):** [SPEC.md](SPEC.md) — the source of truth.
+- **Why it's built this way (history & rationale):** [DESIGN.md](DESIGN.md).
+- **Version log:** [CHANGELOG.md](CHANGELOG.md).
+
+## Running
+
+It's a single static Go binary (`CGO_ENABLED=0`, pure-Go SQLite). Configure via
+environment (see [.env.example](.env.example)); `./scripts/build.sh` builds it and
+`./scripts/run.sh` runs it locally. Deployment notes: [deploy/INSTALL.md](deploy/INSTALL.md).
+
+## AI classifier correction notes
+
+When an API key is configured, each upload is read by the model to pre-fill the
+form. You can steer it with **correction notes** — free-text rules appended to the
+classifier prompt — managed under the **AI** tab. When the model misreads a
+receipt, that upload is recorded; the AI tab lets you review misreads one by one
+and attribute which note fixed each.
+
+Notes live only in the database (private, never committed, included in `/export/db`
+backups). On first run the table is seeded once with these **default notes** (no
+PII), which you can edit or delete:
+
+1. Amounts that use a comma as the decimal separator (e.g. "12,50") mean 12.50, not 1250.
+2. When both a service/visit date and a separate statement, print, or due date appear, use the service date.
+3. "Patient Pay", "You Paid", "Amount Due", and "Patient Responsibility" are the amount actually paid — prefer them over subtotals or insurance-covered amounts.
+
+These defaults are defined in code (`internal/storage/ai_notes.go`); this list is
+the human-readable copy. They are only seeded when the notes table is empty, so a
+deleted default does not come back on restart.

SPEC.md

@@ -79,7 +79,7 @@ A missing/invalid catalog is a fatal startup error.
 ## 4. Data model (SQLite)
 
 - **categories** — `id`, `label` (unique). Seeded; renamable via Manage.
-- **people** — `id`, `label` (unique). Seeded from catalog; renamable via Manage.
+- **people** — `id`, `label` (unique). Seeded from catalog; renamable via Manage. Format is "<FIRST NAME> <LAST NAME>"
 - **receipts** — `id` (UUID), `uploaded_by`, `uploaded_at`, `receipt_date`,
   `amount_cents` (integer — money is never a float), `category_id` (FK, required),
   `person_id` (FK, nullable), `file_path`, `image_data` (BLOB), `file_size_bytes`,
@@ -89,6 +89,12 @@ A missing/invalid catalog is a fatal startup error.
   their own.
 - **tags** — `id`, `label` (unique, **case-insensitive** via `COLLATE NOCASE`).
 - **receipt_tags** — (`receipt_id`, `tag_id`) many-to-many, composite primary key.
+- **ai_notes** — `id`, `text`, `created_at`, `deleted_at` (nullable). Temporal:
+  edits soft-delete + insert, so the active set at any past time is recoverable.
+- **classifications** — `receipt_id` (PK/FK), `model`, `response_json` (the AI
+  suggestion blob), `created_at`, `reviewed`, `reviewed_at`, `resolution`.
+- **miss_fixes** — (`receipt_id`, `note_id`) linking a reviewed misread to the
+  note(s) credited with fixing it.
 
 `PRAGMA foreign_keys=ON`, `journal_mode=WAL`, `busy_timeout=5000` are set on open.
 
@@ -138,9 +144,13 @@ state — the user reviews and submits normally.
   computed locally from the response's token usage and a hardcoded per-model rate
   table. An unknown model id shows token usage but omits the ¢ figure (never a
   guess).
+- The active **correction notes** (§10) are appended to the classifier prompt. The
+  suggestion the model returns is round-tripped through the form and **recorded**
+  against the saved receipt, so misreads can be reviewed later (§10). Skip-AI and
+  disabled-classification uploads record nothing.
 
 ### 5.4 Duplicate warning
-Whenever date and amount are both known, the form calls `GET /duplicates` and warns
+Whenever date and amount are both known (edit: those are mandatory fields aren't they?), the form calls `GET /duplicates` and warns
 if a non-deleted receipt already has the **same `receipt_date` and `amount_cents`**.
 The warning lists each match (amount, date, category, who, uploader, upload time,
 filename) and **disables submit** until the user ticks "Add it anyway". No match →
@@ -235,11 +245,40 @@ any attachments, and the receipt's tag chips.
 - On startup, stray partial-name people (e.g. a leftover "Jude") are merged into the
   unambiguous canonical person ("Jude Tremblay"), reassigning their receipts first,
   so no receipt loses its who. Seeding is idempotent.
-- Tags are **not** managed here (see §13 gaps).
+- Tags are **not** managed here (see §15 gaps).
 
 ---
 
-## 10. Export
+## 10. AI classifier notes & misread review
+
+`GET /ai` — a tab for improving classification over time.
+
+- **Correction notes:** a single global list of free-text rules, appended to the
+  classifier prompt (§5.3). Managed inline — add (`POST /ai/notes`), edit (`POST
+  /ai/notes/edit`), delete (`POST /ai/notes/delete`). The table is **temporal**: an
+  edit soft-deletes the old row and inserts a new one, so the notes active when any
+  past receipt was classified are recoverable. Seeded once from a hardcoded default
+  list (documented in the README; no PII) only when the table is empty, so a deleted
+  default does not return. Read live on every classify call (no restart).
+- **Read-only prompt view:** renders the exact system prompt that would be sent today
+  (static scaffolding + injected people/categories/today + active notes), when
+  classification is enabled.
+- **Misread review:** every AI-run upload stores the suggestion blob + model. A
+  **miss** is *derived* (never stored) — any of the four AI fields (amount, date,
+  category, who) differing from the receipt's final value, where AI-null-then-filled
+  counts as a miss. The review queue lists unreviewed misses; `GET /ai/review/{id}`
+  shows the receipt image, the AI-guess-vs-entered per field, and the notes added
+  *since* that classification. The user ticks which notes fixed it (`POST
+  /ai/review/{id}`) → resolution `fixed` and `miss_fixes` links; ticking none closes
+  it `unresolved`. Either way the row is marked reviewed.
+
+Notes live only in the DB (private, not in git; included in `/export/db` and
+backups). The stored suggestion blob is client-supplied diagnostic data, not ground
+truth.
+
+---
+
+## 11. Export
 
 - `GET /export/db` — downloads a consistent point-in-time copy of the SQLite
   database produced via `VACUUM INTO` (the live file is never locked/served
@@ -252,7 +291,7 @@ any attachments, and the receipt's tag chips.
 
 ---
 
-## 11. Scheduled backups
+## 12. Scheduled backups
 
 A background goroutine writes **metadata-only** snapshots (blobs stripped from both
 receipts and attachments) into `BACKUP_DIR`:
@@ -267,7 +306,7 @@ receipts and attachments) into `BACKUP_DIR`:
 
 ---
 
-## 12. Durability & crash recovery
+## 13. Durability & crash recovery
 
 - WAL mode makes the DB self-healing: an interrupted write is rolled forward if
   committed, discarded otherwise, bringing the DB up at its last committed state.
@@ -277,7 +316,7 @@ receipts and attachments) into `BACKUP_DIR`:
 
 ---
 
-## 13. Deployment & operations
+## 14. Deployment & operations
 
 - The app builds as a fully static binary (`CGO_ENABLED=0`; the SQLite driver is
   pure Go), runnable on any `linux/<arch>`.
@@ -289,7 +328,7 @@ receipts and attachments) into `BACKUP_DIR`:
 
 ---
 
-## 14. Known gaps / not yet implemented
+## 15. Known gaps / not yet implemented
 
 These are intentionally absent today (the schema or storage layer may support some;
 the UI/endpoint does not):
@@ -307,3 +346,7 @@ the UI/endpoint does not):
 - **Archive/zip export** (`/export/archive`): not implemented; only `/export/db`.
 - **HEIC** images are not decodable in pure Go; browser camera uploads arrive as
   JPEG in practice.
+
+Related future work now that AI notes + review exist (§10): per-category/vendor-
+scoped notes, automatic re-classification of past misses, editing the static prompt
+scaffolding, and accuracy-rate stats. See [DESIGN.md](DESIGN.md) item 15.

internal/classify/classify.go

@@ -66,6 +66,12 @@ func New(apiKey, model string, persons []config.Person, categories []config.Cate
 	}
 }
 
+// PromptPreview renders the exact system prompt that would be sent today with the
+// given notes, for the read-only view on the AI tab.
+func (c *Classifier) PromptPreview(today time.Time, notes []string) string {
+	return BuildSystemPrompt(c.Persons, c.Categories, today.Format("2006-01-02"), notes)
+}
+
 const toolName = "record_receipt"
 
 // --- request / response shapes ---
@@ -139,8 +145,8 @@ type toolInput struct {
 // Classify sends the receipt to the model and returns its normalized reading.
 // today is used so the date heuristics ("closest to today, never future") are
 // testable; pass time.Now().
-func (c *Classifier) Classify(ctx context.Context, today time.Time, image []byte, mimeType string) (Suggestion, error) {
-	system := BuildSystemPrompt(c.Persons, c.Categories, today.Format("2006-01-02"))
+func (c *Classifier) Classify(ctx context.Context, today time.Time, image []byte, mimeType string, notes []string) (Suggestion, error) {
+	system := BuildSystemPrompt(c.Persons, c.Categories, today.Format("2006-01-02"), notes)
 
 	imgBlock := apiBlock{
 		Type:   "image",

internal/classify/classify_test.go

@@ -67,7 +67,7 @@ func TestClassifyHappyPath(t *testing.T) {
 		"raw_amount": "$42.50",
 	})
 
-	got, err := c.Classify(context.Background(), time.Now(), []byte("img"), "image/jpeg")
+	got, err := c.Classify(context.Background(), time.Now(), []byte("img"), "image/jpeg", nil)
 	if err != nil {
 		t.Fatalf("Classify: %v", err)
 	}
@@ -99,7 +99,7 @@ func TestClassifyNullsAndFallback(t *testing.T) {
 		"raw_amount": "",
 	})
 
-	got, err := c.Classify(context.Background(), time.Now(), []byte("img"), "image/jpeg")
+	got, err := c.Classify(context.Background(), time.Now(), []byte("img"), "image/jpeg", nil)
 	if err != nil {
 		t.Fatalf("Classify: %v", err)
 	}
@@ -122,7 +122,7 @@ func TestClassifyRejectsNonCanonicalPerson(t *testing.T) {
 		"person":   "Dr. Emily Smith", // a provider, not a configured patient
 		"category": "Medical",
 	})
-	got, err := c.Classify(context.Background(), time.Now(), []byte("img"), "image/jpeg")
+	got, err := c.Classify(context.Background(), time.Now(), []byte("img"), "image/jpeg", nil)
 	if err != nil {
 		t.Fatalf("Classify: %v", err)
 	}
@@ -148,7 +148,7 @@ func TestClassifyIntegration(t *testing.T) {
 	if err != nil {
 		t.Skipf("set HSA_CLASSIFY_IMG to a receipt image: %v", err)
 	}
-	got, err := c.Classify(context.Background(), time.Now(), img, "image/jpeg")
+	got, err := c.Classify(context.Background(), time.Now(), img, "image/jpeg", nil)
 	if err != nil {
 		t.Fatalf("Classify: %v", err)
 	}

internal/classify/prompt.go

@@ -122,7 +122,9 @@ func uniqueCompound(p config.Person, all []config.Person) bool {
 // catalog and today's date (YYYY-MM-DD). The catalog's canonical labels are the only
 // allowed outputs; variants and examples are presented as illustrations, never as
 // data to extract.
-func BuildSystemPrompt(persons []config.Person, categories []config.Category, today string) string {
+// notes are user-authored corrections appended verbatim as a final section; pass nil
+// for none.
+func BuildSystemPrompt(persons []config.Person, categories []config.Category, today string, notes []string) string {
 	var b strings.Builder
 
 	b.WriteString("You extract data from a US health-care receipt for HSA reimbursement. ")
@@ -178,6 +180,18 @@ func BuildSystemPrompt(persons []config.Person, categories []config.Category, to
 	b.WriteString("RAW — also pass back the literal text you read for the name, date, and amount (raw_name, raw_date, raw_amount), exactly as printed, for auditing. Use \"\" if nothing was found.\n")
 	b.WriteString("If the image holds more than one receipt, read the primary (largest/topmost) one.\n")
 
+	// User corrections learned from past misreads (see the AI tab). Appended last so
+	// they take precedence over the general guidance above.
+	if len(notes) > 0 {
+		b.WriteString("\nADDITIONAL CORRECTIONS — learned from past mistakes on these specific receipts; apply them:\n")
+		for _, n := range notes {
+			n = strings.TrimSpace(n)
+			if n != "" {
+				fmt.Fprintf(&b, "  - %s\n", n)
+			}
+		}
+	}
+
 	return b.String()
 }
 

internal/classify/prompt_test.go

@@ -69,19 +69,19 @@ func TestBuildSystemPrompt(t *testing.T) {
 		{Name: "Pharmacy", Examples: []string{"CVS", "Rx"}},
 		{Name: "Other"},
 	}
-	p := BuildSystemPrompt(persons, categories, "2026-06-17")
+	p := BuildSystemPrompt(persons, categories, "2026-06-17", []string{"Treat handwritten totals as authoritative."})
 
 	for _, want := range []string{
 		"Jean-Michel Tremblay", // canonical label
 		"Lynna Nguyen",
 		"Pharmacy",
-		"CVS",                // example injected
-		"2026-06-17",         // today
-		"AMBIGUOUS",          // the ambiguity rule
-		"IGNORE everyone",    // ignore-other-people rule
-		"null",               // unknown handling
-		"raw_name",           // raw echo
-		"CLOSEST to today",   // date tie-breaker
+		"CVS",              // example injected
+		"2026-06-17",       // today
+		"AMBIGUOUS",        // the ambiguity rule
+		"IGNORE everyone",  // ignore-other-people rule
+		"null",             // unknown handling
+		"raw_name",         // raw echo
+		"CLOSEST to today", // date tie-breaker
 	} {
 		if !strings.Contains(p, want) {
 			t.Errorf("system prompt missing %q", want)

internal/storage/ai_notes.go

@@ -0,0 +1,130 @@
+package storage
+
+import (
+	"database/sql"
+	"fmt"
+	"strings"
+	"time"
+)
+
+// Note is one classifier correction note. Notes are global free-text lines appended
+// to the classifier prompt. The table is temporal: an edit soft-deletes the old row
+// and inserts a new one, so the set of notes active at any past time is recoverable.
+type Note struct {
+	ID        int64
+	Text      string
+	CreatedAt time.Time
+}
+
+// defaultNotes seed the ai_notes table on first run (when it is completely empty).
+// They carry no PII and are also documented in the README; users curate from here.
+var defaultNotes = []string{
+	`Amounts that use a comma as the decimal separator (e.g. "12,50") mean 12.50, not 1250.`,
+	`When both a service/visit date and a separate statement, print, or due date appear, use the service date.`,
+	`"Patient Pay", "You Paid", "Amount Due", and "Patient Responsibility" are the amount actually paid — prefer them over subtotals or insurance-covered amounts.`,
+}
+
+// seedNotesIfEmpty inserts the default notes only when the table has no rows at all,
+// so a deliberately-deleted default does not resurrect on the next restart.
+func seedNotesIfEmpty(db *sql.DB) error {
+	var n int
+	if err := db.QueryRow(`SELECT COUNT(*) FROM ai_notes`).Scan(&n); err != nil {
+		return fmt.Errorf("count ai_notes: %w", err)
+	}
+	if n > 0 {
+		return nil
+	}
+	now := time.Now().UTC().Format(rfc3339)
+	for _, t := range defaultNotes {
+		if _, err := db.Exec(`INSERT INTO ai_notes(text, created_at) VALUES (?, ?)`, t, now); err != nil {
+			return fmt.Errorf("seed ai_notes: %w", err)
+		}
+	}
+	return nil
+}
+
+// ListActiveNotes returns the live notes (not soft-deleted), oldest first — the set
+// appended to the classifier prompt.
+func (s *Store) ListActiveNotes() ([]Note, error) {
+	return s.queryNotes(`SELECT id, text, created_at FROM ai_notes
+	                     WHERE deleted_at IS NULL ORDER BY created_at, id`)
+}
+
+// NotesCreatedAfter returns the live notes created strictly after t — the
+// "rules added since this failure" shown when reviewing a miss.
+func (s *Store) NotesCreatedAfter(t time.Time) ([]Note, error) {
+	return s.queryNotes(`SELECT id, text, created_at FROM ai_notes
+	                     WHERE deleted_at IS NULL AND created_at > ? ORDER BY created_at, id`,
+		t.UTC().Format(rfc3339))
+}
+
+func (s *Store) queryNotes(q string, args ...any) ([]Note, error) {
+	rows, err := s.db.Query(q, args...)
+	if err != nil {
+		return nil, fmt.Errorf("query notes: %w", err)
+	}
+	defer rows.Close()
+	var out []Note
+	for rows.Next() {
+		var n Note
+		var created string
+		if err := rows.Scan(&n.ID, &n.Text, &created); err != nil {
+			return nil, fmt.Errorf("scan note: %w", err)
+		}
+		n.CreatedAt, _ = time.Parse(rfc3339, created)
+		out = append(out, n)
+	}
+	return out, rows.Err()
+}
+
+// AddNote inserts a new active note and returns its id.
+func (s *Store) AddNote(text string) (int64, error) {
+	text = strings.TrimSpace(text)
+	if text == "" {
+		return 0, fmt.Errorf("note cannot be empty")
+	}
+	res, err := s.db.Exec(`INSERT INTO ai_notes(text, created_at) VALUES (?, ?)`,
+		text, time.Now().UTC().Format(rfc3339))
+	if err != nil {
+		return 0, fmt.Errorf("add note: %w", err)
+	}
+	return res.LastInsertId()
+}
+
+// DeleteNote soft-deletes a note (preserving history).
+func (s *Store) DeleteNote(id int64) error {
+	_, err := s.db.Exec(`UPDATE ai_notes SET deleted_at = ? WHERE id = ? AND deleted_at IS NULL`,
+		time.Now().UTC().Format(rfc3339), id)
+	if err != nil {
+		return fmt.Errorf("delete note: %w", err)
+	}
+	return nil
+}
+
+// EditNote changes a note's text by soft-deleting the old row and inserting a new
+// one, so the temporal history (what was active when) is preserved. Returns the new id.
+func (s *Store) EditNote(id int64, text string) (int64, error) {
+	text = strings.TrimSpace(text)
+	if text == "" {
+		return 0, fmt.Errorf("note cannot be empty")
+	}
+	tx, err := s.db.Begin()
+	if err != nil {
+		return 0, fmt.Errorf("begin edit note: %w", err)
+	}
+	defer tx.Rollback()
+
+	now := time.Now().UTC().Format(rfc3339)
+	if _, err := tx.Exec(`UPDATE ai_notes SET deleted_at = ? WHERE id = ? AND deleted_at IS NULL`, now, id); err != nil {
+		return 0, fmt.Errorf("retire old note: %w", err)
+	}
+	res, err := tx.Exec(`INSERT INTO ai_notes(text, created_at) VALUES (?, ?)`, text, now)
+	if err != nil {
+		return 0, fmt.Errorf("insert edited note: %w", err)
+	}
+	newID, err := res.LastInsertId()
+	if err != nil {
+		return 0, err
+	}
+	return newID, tx.Commit()
+}

internal/storage/ai_notes_test.go

@@ -0,0 +1,106 @@
+package storage
+
+import (
+	"path/filepath"
+	"testing"
+	"time"
+)
+
+func notesTestStore(t *testing.T) *Store {
+	t.Helper()
+	s, err := Open(filepath.Join(t.TempDir(), "notes.db"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { s.Close() })
+	return s
+}
+
+func TestNotes_SeededOnFirstOpen(t *testing.T) {
+	s := notesTestStore(t)
+	notes, err := s.ListActiveNotes()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(notes) != len(defaultNotes) {
+		t.Fatalf("seeded %d notes, want %d", len(notes), len(defaultNotes))
+	}
+}
+
+func TestNotes_DeletedDefaultDoesNotResurrect(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "n.db")
+
+	s, err := Open(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	all, _ := s.ListActiveNotes()
+	if err := s.DeleteNote(all[0].ID); err != nil {
+		t.Fatal(err)
+	}
+	s.Close()
+
+	// Reopen: seeding must NOT run again (table is non-empty thanks to history).
+	s2, err := Open(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer s2.Close()
+	notes, _ := s2.ListActiveNotes()
+	if len(notes) != len(defaultNotes)-1 {
+		t.Errorf("after delete+reopen got %d active notes, want %d", len(notes), len(defaultNotes)-1)
+	}
+}
+
+func TestNotes_EditKeepsHistory(t *testing.T) {
+	s := notesTestStore(t)
+	id, err := s.AddNote("original")
+	if err != nil {
+		t.Fatal(err)
+	}
+	newID, err := s.EditNote(id, "revised")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if newID == id {
+		t.Error("edit should create a new row id")
+	}
+	active, _ := s.ListActiveNotes()
+	var texts []string
+	for _, n := range active {
+		texts = append(texts, n.Text)
+	}
+	if contains(texts, "original") {
+		t.Error("old text should no longer be active")
+	}
+	if !contains(texts, "revised") {
+		t.Error("revised text should be active")
+	}
+}
+
+func TestNotes_CreatedAfter(t *testing.T) {
+	s := notesTestStore(t)
+	cutoff := time.Now().UTC()
+	// created_at has second resolution (RFC3339); make sure the new note sorts after.
+	time.Sleep(1100 * time.Millisecond)
+	if _, err := s.AddNote("late note"); err != nil {
+		t.Fatal(err)
+	}
+	after, err := s.NotesCreatedAfter(cutoff)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(after) != 1 || after[0].Text != "late note" {
+		t.Fatalf("NotesCreatedAfter = %v, want just the late note", after)
+	}
+}
+
+func contains(ss []string, want string) bool {
+	for _, s := range ss {
+		if s == want {
+			return true
+		}
+	}
+	return false
+}

internal/storage/classifications.go

@@ -0,0 +1,123 @@
+package storage
+
+import (
+	"database/sql"
+	"fmt"
+	"time"
+)
+
+// ClassificationRow is a stored AI classification of a receipt, joined with the
+// receipt's final field values so the caller can derive which fields the user
+// overrode (a "miss"). The response blob is opaque diagnostic data here — the web
+// layer parses it; storage never interprets it.
+type ClassificationRow struct {
+	ReceiptID    string
+	Model        string
+	ResponseJSON string
+	CreatedAt    time.Time
+	Reviewed     bool
+	ReviewedAt   *time.Time
+	Resolution   string
+
+	// Final receipt values, for deriving overrides and display.
+	FinalAmountCents int64
+	FinalReceiptDate time.Time
+	FinalCategoryID  int64
+	FinalPersonID    *int64
+}
+
+// InsertClassification records the AI suggestion (its response blob + model) for a
+// receipt. One row per AI-run upload; skip-AI/disabled uploads insert nothing.
+func (s *Store) InsertClassification(receiptID, model, responseJSON string, createdAt time.Time) error {
+	_, err := s.db.Exec(
+		`INSERT INTO classifications(receipt_id, model, response_json, created_at)
+		 VALUES (?, ?, ?, ?)`,
+		receiptID, model, responseJSON, createdAt.UTC().Format(rfc3339))
+	if err != nil {
+		return fmt.Errorf("insert classification: %w", err)
+	}
+	return nil
+}
+
+const classificationSelect = `
+SELECT c.receipt_id, c.model, c.response_json, c.created_at, c.reviewed, c.reviewed_at, c.resolution,
+       r.amount_cents, r.receipt_date, r.category_id, r.person_id
+FROM classifications c
+JOIN receipts r ON r.id = c.receipt_id`
+
+// ListUnreviewedClassifications returns not-yet-reviewed classifications of live
+// receipts, newest first. Whether each is an actual "miss" is derived by the caller
+// from the blob vs. the final fields.
+func (s *Store) ListUnreviewedClassifications() ([]ClassificationRow, error) {
+	rows, err := s.db.Query(classificationSelect +
+		` WHERE c.reviewed = 0 AND r.deleted_at IS NULL ORDER BY c.created_at DESC`)
+	if err != nil {
+		return nil, fmt.Errorf("list classifications: %w", err)
+	}
+	defer rows.Close()
+	var out []ClassificationRow
+	for rows.Next() {
+		c, err := scanClassification(rows)
+		if err != nil {
+			return nil, err
+		}
+		out = append(out, c)
+	}
+	return out, rows.Err()
+}
+
+// GetClassification returns the classification for one receipt.
+func (s *Store) GetClassification(receiptID string) (ClassificationRow, error) {
+	return scanClassification(s.db.QueryRow(classificationSelect+` WHERE c.receipt_id = ?`, receiptID))
+}
+
+func scanClassification(sc rowScanner) (ClassificationRow, error) {
+	var c ClassificationRow
+	var createdAt, receiptDate string
+	var reviewed int64
+	var reviewedAt, resolution sql.NullString
+	var personID sql.NullInt64
+	if err := sc.Scan(&c.ReceiptID, &c.Model, &c.ResponseJSON, &createdAt, &reviewed, &reviewedAt, &resolution,
+		&c.FinalAmountCents, &receiptDate, &c.FinalCategoryID, &personID); err != nil {
+		return ClassificationRow{}, fmt.Errorf("scan classification: %w", err)
+	}
+	c.CreatedAt, _ = time.Parse(rfc3339, createdAt)
+	c.FinalReceiptDate, _ = time.Parse(rfc3339, receiptDate)
+	c.Reviewed = reviewed != 0
+	if reviewedAt.Valid {
+		if t, err := time.Parse(rfc3339, reviewedAt.String); err == nil {
+			c.ReviewedAt = &t
+		}
+	}
+	if resolution.Valid {
+		c.Resolution = resolution.String
+	}
+	if personID.Valid {
+		c.FinalPersonID = &personID.Int64
+	}
+	return c, nil
+}
+
+// MarkClassificationReviewed flags a classification reviewed with the given
+// resolution and links the notes the user credited with fixing it.
+func (s *Store) MarkClassificationReviewed(receiptID, resolution string, fixNoteIDs []int64) error {
+	tx, err := s.db.Begin()
+	if err != nil {
+		return fmt.Errorf("begin review: %w", err)
+	}
+	defer tx.Rollback()
+
+	if _, err := tx.Exec(
+		`UPDATE classifications SET reviewed = 1, reviewed_at = ?, resolution = ? WHERE receipt_id = ?`,
+		time.Now().UTC().Format(rfc3339), resolution, receiptID); err != nil {
+		return fmt.Errorf("mark reviewed: %w", err)
+	}
+	for _, nid := range fixNoteIDs {
+		if _, err := tx.Exec(
+			`INSERT OR IGNORE INTO miss_fixes(receipt_id, note_id) VALUES (?, ?)`,
+			receiptID, nid); err != nil {
+			return fmt.Errorf("link fix note: %w", err)
+		}
+	}
+	return tx.Commit()
+}

internal/storage/classifications_test.go

@@ -0,0 +1,69 @@
+package storage
+
+import (
+	"path/filepath"
+	"testing"
+	"time"
+)
+
+func classifyTestStore(t *testing.T) *Store {
+	t.Helper()
+	s, err := Open(filepath.Join(t.TempDir(), "c.db"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { s.Close() })
+	return s
+}
+
+func TestClassifications_InsertListReview(t *testing.T) {
+	s := classifyTestStore(t)
+	insertReceipt(t, s, "r1")
+
+	if err := s.InsertClassification("r1", "haiku", `{"model":"haiku"}`, time.Now().UTC()); err != nil {
+		t.Fatal(err)
+	}
+
+	rows, err := s.ListUnreviewedClassifications()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(rows) != 1 || rows[0].ReceiptID != "r1" || rows[0].Model != "haiku" {
+		t.Fatalf("ListUnreviewedClassifications = %+v", rows)
+	}
+	if rows[0].FinalAmountCents != 100 { // insertReceipt sets 100 cents
+		t.Errorf("joined final amount = %d, want 100", rows[0].FinalAmountCents)
+	}
+
+	// Review it, crediting a note as the fix.
+	noteID, _ := s.AddNote("the fix")
+	if err := s.MarkClassificationReviewed("r1", "fixed", []int64{noteID}); err != nil {
+		t.Fatal(err)
+	}
+	rows, _ = s.ListUnreviewedClassifications()
+	if len(rows) != 0 {
+		t.Errorf("reviewed row still in unreviewed queue: %+v", rows)
+	}
+	got, err := s.GetClassification("r1")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !got.Reviewed || got.Resolution != "fixed" || got.ReviewedAt == nil {
+		t.Errorf("review state not persisted: %+v", got)
+	}
+}
+
+func TestClassifications_DeletedReceiptDropsFromQueue(t *testing.T) {
+	s := classifyTestStore(t)
+	insertReceipt(t, s, "r1")
+	if err := s.InsertClassification("r1", "m", `{}`, time.Now().UTC()); err != nil {
+		t.Fatal(err)
+	}
+	if err := s.SoftDelete("r1"); err != nil {
+		t.Fatal(err)
+	}
+	rows, _ := s.ListUnreviewedClassifications()
+	if len(rows) != 0 {
+		t.Errorf("soft-deleted receipt should not appear in review queue: %+v", rows)
+	}
+}

internal/storage/storage.go

@@ -75,6 +75,27 @@ CREATE TABLE IF NOT EXISTS receipt_tags (
     PRIMARY KEY (receipt_id, tag_id)
 );
 CREATE INDEX IF NOT EXISTS idx_receipt_tags_tag ON receipt_tags(tag_id);
+CREATE TABLE IF NOT EXISTS ai_notes (
+    id         INTEGER PRIMARY KEY AUTOINCREMENT,
+    text       TEXT NOT NULL,
+    created_at TEXT NOT NULL,
+    deleted_at TEXT
+);
+CREATE TABLE IF NOT EXISTS classifications (
+    receipt_id    TEXT PRIMARY KEY REFERENCES receipts(id),
+    model         TEXT NOT NULL,
+    response_json TEXT NOT NULL,
+    created_at    TEXT NOT NULL,
+    reviewed      INTEGER NOT NULL DEFAULT 0,
+    reviewed_at   TEXT,
+    resolution    TEXT
+);
+CREATE INDEX IF NOT EXISTS idx_classifications_reviewed ON classifications(reviewed);
+CREATE TABLE IF NOT EXISTS miss_fixes (
+    receipt_id TEXT    NOT NULL REFERENCES receipts(id),
+    note_id    INTEGER NOT NULL REFERENCES ai_notes(id),
+    PRIMARY KEY (receipt_id, note_id)
+);
 `
 
 // Open opens (creating if needed) the SQLite database at path, applies the schema,
@@ -107,6 +128,10 @@ func Open(path string) (*Store, error) {
 			return nil, fmt.Errorf("seed categories: %w", err)
 		}
 	}
+	if err := seedNotesIfEmpty(db); err != nil {
+		db.Close()
+		return nil, err
+	}
 	return &Store{db: db}, nil
 }
 

internal/web/ai.go

@@ -0,0 +1,273 @@
+package web
+
+import (
+	"encoding/json"
+	"net/http"
+	"net/url"
+	"strconv"
+	"strings"
+	"time"
+
+	"maisym.com/hsa/internal/receipt"
+	"maisym.com/hsa/internal/storage"
+)
+
+// --- AI tab: classifier correction notes + failure review ---
+
+type aiView struct {
+	Notes      []storage.Note
+	Misses     []missSummary // unreviewed failures
+	Prompt     string        // live assembled system prompt, or "" when unavailable
+	PromptNote string        // why the prompt is unavailable (classification off)
+	Error      string
+}
+
+// missSummary is one unreviewed failure in the review queue.
+type missSummary struct {
+	ReceiptID string
+	When      string
+	Amount    string
+	Model     string
+	Fields    string // comma-joined names of the overridden fields
+}
+
+func (s *Server) handleAI(w http.ResponseWriter, r *http.Request) {
+	notes, err := s.store.ListActiveNotes()
+	if err != nil {
+		s.serverError(w, "list notes", err)
+		return
+	}
+	rows, err := s.store.ListUnreviewedClassifications()
+	if err != nil {
+		s.serverError(w, "list classifications", err)
+		return
+	}
+	cats, people, err := s.lookups()
+	if err != nil {
+		s.serverError(w, "load lookups", err)
+		return
+	}
+
+	var misses []missSummary
+	for _, c := range rows {
+		ovs := deriveOverrides(c, cats, people)
+		var names []string
+		for _, o := range ovs {
+			if o.Overridden {
+				names = append(names, o.Field)
+			}
+		}
+		if len(names) == 0 {
+			continue // not a miss — the AI got every field right
+		}
+		misses = append(misses, missSummary{
+			ReceiptID: c.ReceiptID,
+			When:      c.CreatedAt.Local().Format(dateLayout),
+			Amount:    dollars(c.FinalAmountCents),
+			Model:     c.Model,
+			Fields:    strings.Join(names, ", "),
+		})
+	}
+
+	view := aiView{Notes: notes, Misses: misses, Error: r.URL.Query().Get("error")}
+	if s.classifier != nil {
+		var texts []string
+		for _, n := range notes {
+			texts = append(texts, n.Text)
+		}
+		view.Prompt = s.classifier.PromptPreview(time.Now(), texts)
+	} else {
+		view.PromptNote = "Classification is disabled (no API key configured), so the live prompt is unavailable."
+	}
+
+	w.Header().Set("Content-Type", "text/html; charset=utf-8")
+	if err := aiPage.ExecuteTemplate(w, "base", view); err != nil {
+		s.serverError(w, "render ai", err)
+	}
+}
+
+func (s *Server) handleAddNote(w http.ResponseWriter, r *http.Request) {
+	if _, err := s.store.AddNote(strings.TrimSpace(r.FormValue("text"))); err != nil {
+		redirectAI(w, r, "Could not add the note (it may be empty).")
+		return
+	}
+	redirectAI(w, r, "")
+}
+
+func (s *Server) handleEditNote(w http.ResponseWriter, r *http.Request) {
+	id, err := strconv.ParseInt(strings.TrimSpace(r.FormValue("id")), 10, 64)
+	if err != nil {
+		redirectAI(w, r, "Invalid note id.")
+		return
+	}
+	if _, err := s.store.EditNote(id, strings.TrimSpace(r.FormValue("text"))); err != nil {
+		redirectAI(w, r, "Could not save the note (it may be empty).")
+		return
+	}
+	redirectAI(w, r, "")
+}
+
+func (s *Server) handleDeleteNote(w http.ResponseWriter, r *http.Request) {
+	id, err := strconv.ParseInt(strings.TrimSpace(r.FormValue("id")), 10, 64)
+	if err != nil {
+		redirectAI(w, r, "Invalid note id.")
+		return
+	}
+	if err := s.store.DeleteNote(id); err != nil {
+		redirectAI(w, r, "Could not delete the note.")
+		return
+	}
+	redirectAI(w, r, "")
+}
+
+// reviewView is the single-failure review page.
+type reviewView struct {
+	ReceiptID  string
+	When       string
+	Model      string
+	Fields     []fieldOverride
+	SinceNotes []storage.Note // notes added after this failure (candidate fixes)
+}
+
+func (s *Server) handleReview(w http.ResponseWriter, r *http.Request) {
+	id := r.PathValue("id")
+	c, err := s.store.GetClassification(id)
+	if err != nil {
+		http.Error(w, "not found", http.StatusNotFound)
+		return
+	}
+	cats, people, err := s.lookups()
+	if err != nil {
+		s.serverError(w, "load lookups", err)
+		return
+	}
+	since, err := s.store.NotesCreatedAfter(c.CreatedAt)
+	if err != nil {
+		s.serverError(w, "notes since", err)
+		return
+	}
+	w.Header().Set("Content-Type", "text/html; charset=utf-8")
+	if err := reviewPage.ExecuteTemplate(w, "base", reviewView{
+		ReceiptID:  c.ReceiptID,
+		When:       c.CreatedAt.Local().Format("2006-01-02 15:04"),
+		Model:      c.Model,
+		Fields:     deriveOverrides(c, cats, people),
+		SinceNotes: since,
+	}); err != nil {
+		s.serverError(w, "render review", err)
+	}
+}
+
+func (s *Server) handleReviewSubmit(w http.ResponseWriter, r *http.Request) {
+	id := r.PathValue("id")
+	if err := r.ParseForm(); err != nil {
+		http.Error(w, "bad form", http.StatusBadRequest)
+		return
+	}
+	var fixIDs []int64
+	for _, v := range r.Form["fix"] {
+		if nid, err := strconv.ParseInt(v, 10, 64); err == nil {
+			fixIDs = append(fixIDs, nid)
+		}
+	}
+	resolution := "unresolved"
+	if len(fixIDs) > 0 {
+		resolution = "fixed"
+	}
+	if err := s.store.MarkClassificationReviewed(id, resolution, fixIDs); err != nil {
+		s.serverError(w, "mark reviewed", err)
+		return
+	}
+	http.Redirect(w, r, "/ai", http.StatusSeeOther)
+}
+
+func redirectAI(w http.ResponseWriter, r *http.Request, errMsg string) {
+	target := "/ai"
+	if errMsg != "" {
+		target += "?error=" + url.QueryEscape(errMsg)
+	}
+	http.Redirect(w, r, target, http.StatusSeeOther)
+}
+
+// fieldOverride is one of the four AI-suggested fields, with the suggestion, the
+// final value, and whether the user overrode it.
+type fieldOverride struct {
+	Field      string
+	Suggested  string
+	Final      string
+	Overridden bool
+}
+
+// deriveOverrides compares the stored AI suggestion blob against the receipt's final
+// values to determine which of the four fields the user changed. A field the AI left
+// null/empty that the user then filled counts as an override (the AI missed it).
+func deriveOverrides(c storage.ClassificationRow, cats, people []storage.Lookup) []fieldOverride {
+	var cr classifyResponse
+	_ = json.Unmarshal([]byte(c.ResponseJSON), &cr)
+
+	finalAmount := dollars(c.FinalAmountCents)
+	finalDate := c.FinalReceiptDate.Format(dateLayout)
+	finalCat := labelFor(c.FinalCategoryID, cats)
+	finalWho := whoLabel(c.FinalPersonID, people)
+
+	return []fieldOverride{
+		{Field: "amount", Suggested: strPtr(cr.Amount), Final: finalAmount,
+			Overridden: amountOverridden(cr.Amount, c.FinalAmountCents)},
+		{Field: "date", Suggested: strPtr(cr.Date), Final: finalDate,
+			Overridden: dateOverridden(cr.Date, c.FinalReceiptDate)},
+		{Field: "category", Suggested: lookupOr(cr.CategoryID, cats, cr.Category), Final: finalCat,
+			Overridden: int64PtrOverridden(cr.CategoryID, &c.FinalCategoryID)},
+		{Field: "who", Suggested: lookupOr(cr.PersonID, people, cr.Person), Final: finalWho,
+			Overridden: int64PtrOverridden(cr.PersonID, c.FinalPersonID)},
+	}
+}
+
+func amountOverridden(suggested *string, finalCents int64) bool {
+	if suggested == nil {
+		return true
+	}
+	cents, err := receipt.ParseAmountCents(*suggested)
+	if err != nil {
+		return true
+	}
+	return cents != finalCents
+}
+
+func dateOverridden(suggested *string, final time.Time) bool {
+	if suggested == nil {
+		return true
+	}
+	d, err := time.Parse(dateLayout, strings.TrimSpace(*suggested))
+	if err != nil {
+		return true
+	}
+	return d.Year() != final.Year() || d.Month() != final.Month() || d.Day() != final.Day()
+}
+
+func int64PtrOverridden(suggested, final *int64) bool {
+	if suggested == nil || final == nil {
+		return suggested != final // both nil → not overridden; one nil → overridden
+	}
+	return *suggested != *final
+}
+
+func strPtr(p *string) string {
+	if p == nil || strings.TrimSpace(*p) == "" {
+		return "—"
+	}
+	return *p
+}
+
+// lookupOr renders a suggested lookup id as its label, falling back to a label the
+// blob already carried, then to "—".
+func lookupOr(id *int64, set []storage.Lookup, fallback string) string {
+	if id != nil {
+		if l := labelFor(*id, set); l != "" {
+			return l
+		}
+	}
+	if strings.TrimSpace(fallback) != "" {
+		return fallback
+	}
+	return "—"
+}

internal/web/ai_test.go

@@ -0,0 +1,122 @@
+package web
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"strconv"
+	"strings"
+	"testing"
+)
+
+// uploadWithClassifyBlob posts a receipt plus a round-tripped AI suggestion blob.
+func uploadWithClassifyBlob(t *testing.T, s *Server, amount, classifyJSON string) {
+	t.Helper()
+	body, ct := multipartUpload(t, map[string]string{
+		"amount":        amount,
+		"receipt_date":  "2026-06-01",
+		"category_id":   aCategoryID(t, s),
+		"classify_json": classifyJSON,
+	}, "receipt", "r.png", fakePNG())
+	req := httptest.NewRequest(http.MethodPost, "/upload", body)
+	req.Header.Set("Content-Type", ct)
+	req.AddCookie(authCookie(t, s))
+	rec := httptest.NewRecorder()
+	s.Routes().ServeHTTP(rec, req)
+	if rec.Code != http.StatusOK {
+		t.Fatalf("upload status=%d body=%s", rec.Code, rec.Body.String())
+	}
+}
+
+func TestAI_MissRecordedAndReviewable(t *testing.T) {
+	s := testServerWithStore(t)
+
+	// AI suggested $99.99; user submitted $12.34 → an amount miss.
+	blob := `{"amount":"99.99","date":"2026-06-01","category_id":null,"person_id":null,"model":"haiku-test"}`
+	uploadWithClassifyBlob(t, s, "12.34", blob)
+
+	// Stored classification exists.
+	rows, err := s.store.ListUnreviewedClassifications()
+	if err != nil || len(rows) != 1 {
+		t.Fatalf("ListUnreviewedClassifications: err=%v len=%d", err, len(rows))
+	}
+	id := rows[0].ReceiptID
+
+	// /ai lists it as a miss touching "amount".
+	page := get(t, s, "/ai").Body.String()
+	if !strings.Contains(page, "haiku-test") || !strings.Contains(page, "amount") {
+		t.Errorf("/ai missing the amount miss: %s", page)
+	}
+
+	// Review page shows the AI guess vs the entered value.
+	rv := get(t, s, "/ai/review/"+id).Body.String()
+	if !strings.Contains(rv, "99.99") || !strings.Contains(rv, "12.34") {
+		t.Errorf("review page missing guess/final: %s", rv)
+	}
+
+	// Submitting the review (no fix ticked) closes it as unresolved and clears the queue.
+	req := httptest.NewRequest(http.MethodPost, "/ai/review/"+id, strings.NewReader(url.Values{}.Encode()))
+	req.Header.Set("Content-Type", "application/x-www-form-urlencoded")
+	req.AddCookie(authCookie(t, s))
+	rec := httptest.NewRecorder()
+	s.Routes().ServeHTTP(rec, req)
+	if rec.Code != http.StatusSeeOther {
+		t.Fatalf("review submit status=%d", rec.Code)
+	}
+	left, _ := s.store.ListUnreviewedClassifications()
+	if len(left) != 0 {
+		t.Errorf("queue not cleared after review: %d left", len(left))
+	}
+	if c, _ := s.store.GetClassification(id); c.Resolution != "unresolved" {
+		t.Errorf("resolution = %q, want unresolved", c.Resolution)
+	}
+}
+
+func TestAI_NoMissWhenSuggestionMatches(t *testing.T) {
+	s := testServerWithStore(t)
+	cat := aCategoryID(t, s)
+	// AI nailed every field the user submitted → not a miss.
+	blob := `{"amount":"12.34","date":"2026-06-01","category_id":` + cat + `,"person_id":null,"model":"m"}`
+	uploadWithClassifyBlob(t, s, "12.34", blob)
+
+	page := get(t, s, "/ai").Body.String()
+	if !strings.Contains(page, "No misreads to review") {
+		t.Errorf("expected no-misreads message, got: %s", page)
+	}
+}
+
+func TestAI_NotesCRUD(t *testing.T) {
+	s := testServerWithStore(t)
+
+	post := func(path string, form url.Values) {
+		req := httptest.NewRequest(http.MethodPost, path, strings.NewReader(form.Encode()))
+		req.Header.Set("Content-Type", "application/x-www-form-urlencoded")
+		req.AddCookie(authCookie(t, s))
+		rec := httptest.NewRecorder()
+		s.Routes().ServeHTTP(rec, req)
+		if rec.Code != http.StatusSeeOther {
+			t.Fatalf("%s status=%d", path, rec.Code)
+		}
+	}
+
+	post("/ai/notes", url.Values{"text": {"European dates are DD/MM"}})
+	notes, _ := s.store.ListActiveNotes()
+	var added int64
+	found := false
+	for _, n := range notes {
+		if n.Text == "European dates are DD/MM" {
+			added, found = n.ID, true
+		}
+	}
+	if !found {
+		t.Fatal("note not added")
+	}
+
+	post("/ai/notes/delete", url.Values{"id": {strconv.FormatInt(added, 10)}})
+	notes, _ = s.store.ListActiveNotes()
+	for _, n := range notes {
+		if n.ID == added {
+			t.Error("note not deleted")
+		}
+	}
+}

internal/web/scan.go

@@ -27,7 +27,7 @@ type classifyResponse struct {
 	RawDate   string `json:"raw_date"`
 	RawAmount string `json:"raw_amount"`
 
-	Model     string   `json:"model"`              // model that ran, for display
+	Model     string   `json:"model"`                // model that ran, for display
 	CostCents *float64 `json:"cost_cents,omitempty"` // nil when the model's price is unknown
 }
 
@@ -64,7 +64,15 @@ func (s *Server) handleClassify(w http.ResponseWriter, r *http.Request) {
 	}
 	data = normalizeOrientation(data, mimeType) // upright image reads better for the AI
 
-	sug, err := s.classifier.Classify(r.Context(), time.Now(), data, mimeType)
+	// Append the user's correction notes to the prompt (best-effort).
+	var notes []string
+	if ns, err := s.store.ListActiveNotes(); err == nil {
+		for _, n := range ns {
+			notes = append(notes, n.Text)
+		}
+	}
+
+	sug, err := s.classifier.Classify(r.Context(), time.Now(), data, mimeType, notes)
 	if err != nil {
 		s.serverError(w, "classify receipt", err)
 		return

internal/web/static/style.css

@@ -202,3 +202,17 @@ form.row {
 form.row input { flex: 1; padding: .4rem; }
 form.row button { padding: .4rem .7rem; }
 .add input { border: 1px solid #137333; }
+form.row button.danger { background: #fce8e6; border: 1px solid #d93025; color: #a50e0e; }
+
+/* AI tab: read-only prompt + miss-review table. */
+pre.prompt {
+  white-space: pre-wrap;
+  word-break: break-word;
+  background: #f1f3f4;
+  padding: .6rem;
+  border-radius: .4rem;
+  font-size: .78rem;
+  max-height: 28rem;
+  overflow-y: auto;
+}
+table.tally tr.miss th, table.tally tr.miss td { background: #fef7e0; }

internal/web/templates.go

@@ -24,4 +24,6 @@ var (
 	exportPage  = parsePage("export.html")
 	tallyPage   = parsePage("tally.html")
 	recentPage  = parsePage("recent.html")
+	aiPage      = parsePage("ai.html")
+	reviewPage  = parsePage("ai_review.html")
 )

internal/web/templates/ai.html

@@ -0,0 +1,49 @@
+{{define "title"}}AI{{end}}
+{{define "content"}}
+<header>
+  <span><a href="/">← Add receipt</a></span>
+  <span><a href="/recent">Recent</a> · <a href="/tally">Tally</a> · <a href="/manage">Manage</a></span>
+</header>
+<h1>AI classifier</h1>
+{{if .Error}}<div class="err">{{.Error}}</div>{{end}}
+
+<h2>Correction notes</h2>
+<p><small>Free-text rules appended to the classifier prompt to fix recurring misreads. Editing a note keeps the old version in history.</small></p>
+{{range .Notes}}
+<form class="row" method="post" action="/ai/notes/edit">
+  <input type="hidden" name="id" value="{{.ID}}">
+  <input type="text" name="text" value="{{.Text}}">
+  <button type="submit">Save</button>
+  <button type="submit" formaction="/ai/notes/delete" class="danger">✕</button>
+</form>
+{{end}}
+<form class="row add" method="post" action="/ai/notes">
+  <input type="text" name="text" placeholder="New correction note">
+  <button type="submit">Add</button>
+</form>
+
+<h2>Review misreads {{if .Misses}}({{len .Misses}}){{end}}</h2>
+{{if not .Misses}}
+<p><small>No misreads to review — the AI matched what you submitted (or nothing has been classified yet).</small></p>
+{{else}}
+<ul class="recent">
+  {{range .Misses}}
+  <li>
+    <a class="amt" href="/ai/review/{{.ReceiptID}}">${{.Amount}}</a>
+    <span class="meta">{{.When}} · changed: {{.Fields}}</span>
+    <span class="when">{{.Model}}</span>
+  </li>
+  {{end}}
+</ul>
+{{end}}
+
+<h2>Current prompt</h2>
+{{if .Prompt}}
+<details>
+  <summary>View the exact system prompt sent to the model</summary>
+  <pre class="prompt">{{.Prompt}}</pre>
+</details>
+{{else}}
+<p><small>{{.PromptNote}}</small></p>
+{{end}}
+{{end}}

internal/web/templates/ai_review.html

@@ -0,0 +1,38 @@
+{{define "title"}}Review misread{{end}}
+{{define "content"}}
+<header>
+  <span><a href="/ai">← AI</a></span>
+  <span>classified by {{.Model}} · {{.When}}</span>
+</header>
+<h1>Review misread</h1>
+
+<p><a href="/receipt/{{.ReceiptID}}/file" target="_blank">Open the receipt image ↗</a></p>
+
+<table class="tally">
+  <thead><tr><th>Field</th><th>AI guessed</th><th>You entered</th></tr></thead>
+  <tbody>
+    {{range .Fields}}
+    <tr{{if .Overridden}} class="miss"{{end}}>
+      <th>{{.Field}}</th>
+      <td>{{.Suggested}}</td>
+      <td>{{.Final}}{{if .Overridden}} ✎{{end}}</td>
+    </tr>
+    {{end}}
+  </tbody>
+</table>
+<p><small>Rows marked ✎ are where you overrode the AI.</small></p>
+
+<form method="post" action="/ai/review/{{.ReceiptID}}">
+  <h2>What fixed it?</h2>
+  {{if .SinceNotes}}
+  <p><small>Notes added since this receipt was classified — tick any that address this misread:</small></p>
+  {{range .SinceNotes}}
+  <label class="check"><input type="checkbox" name="fix" value="{{.ID}}"> {{.Text}}</label>
+  {{end}}
+  {{else}}
+  <p><small>No notes were added after this receipt was classified. Add a correction note on the <a href="/ai">AI page</a> first if one would help, then come back — or close this as unresolved.</small></p>
+  {{end}}
+  <button type="submit" class="primary">Done reviewing</button>
+  <p><small>Ticking one or more notes marks this as fixed; ticking none closes it as unresolved.</small></p>
+</form>
+{{end}}

internal/web/templates/recent.html

@@ -2,7 +2,7 @@
 {{define "content"}}
 <header>
   <span><a href="/">← Add receipt</a></span>
-  <span><a href="/tally">Tally</a> · <a href="/manage">Manage</a> · <a href="/export">Export</a></span>
+  <span><a href="/tally">Tally</a> · <a href="/ai">AI</a> · <a href="/manage">Manage</a> · <a href="/export">Export</a></span>
 </header>
 <h1>{{.Title}}</h1>
 <p class="tabs">

internal/web/templates/upload.html

@@ -2,7 +2,7 @@
 {{define "content"}}
 <header>
   <span>{{.Subject}}</span>
-  <span><a href="/tally">Tally</a> · <a href="/recent">Recent</a> · <a href="/manage">Manage</a> · <a href="/export">Export</a> · <a href="/logout">Log out</a></span>
+  <span><a href="/tally">Tally</a> · <a href="/recent">Recent</a> · <a href="/ai">AI</a> · <a href="/manage">Manage</a> · <a href="/export">Export</a> · <a href="/logout">Log out</a></span>
 </header>
 <h1>Add receipt</h1>
 {{range .Errors}}<div class="err">{{.}}</div>{{end}}
@@ -58,6 +58,7 @@
     <ul id="dup-list"></ul>
     <label class="check"><input type="checkbox" id="dup-ack"> Add it anyway</label>
   </div>
+  <input type="hidden" name="classify_json" id="classify_json">
   <button type="submit" class="primary" id="submit-btn">Submit</button>
 </form>
 
@@ -189,11 +190,16 @@
 
   skip.addEventListener("change", function () {
     note.style.opacity = skip.checked ? "0.5" : "";
-    if (skip.checked) note.textContent = "AI auto-fill skipped — enter the fields by hand.";
-    else note.innerHTML = baseNote;
+    if (skip.checked) {
+      note.textContent = "AI auto-fill skipped — enter the fields by hand.";
+      setValue("classify_json", ""); // no AI suggestion to record
+    } else {
+      note.innerHTML = baseNote;
+    }
   });
 
   fileInput.addEventListener("change", function () {
+    setValue("classify_json", ""); // a new file invalidates any prior suggestion
     var file = fileInput.files[0];
     if (!file || skip.checked) return;
     note.style.opacity = "";
@@ -212,6 +218,7 @@
         if (d.date) setValue("receipt_date", d.date);
         if (d.category_id != null) setValue("category_id", String(d.category_id));
         if (d.person_id != null) setValue("person_id", String(d.person_id));
+        setValue("classify_json", JSON.stringify(d)); // round-trip for misread review
         var cost = (d.cost_cents != null) ? " · " + d.cost_cents.toFixed(2) + "¢" : "";
         note.textContent = "Pre-filled by " + (d.model || model) + cost + " — review and submit.";
         // Pre-filled date+amount may match an existing receipt — re-check.

internal/web/upload.go

@@ -2,8 +2,10 @@ package web
 
 import (
 	"crypto/rand"
+	"encoding/json"
 	"fmt"
 	"io"
+	"log"
 	"net/http"
 	"os"
 	"path/filepath"
@@ -254,6 +256,16 @@ func (s *Server) handleUpload(w http.ResponseWriter, r *http.Request) {
 		}
 	}
 
+	// Persist the AI suggestion the browser round-tripped, so misreads can be
+	// reviewed later (best-effort, diagnostic — never block the upload on it).
+	if blob := strings.TrimSpace(r.FormValue("classify_json")); blob != "" {
+		var cr classifyResponse
+		_ = json.Unmarshal([]byte(blob), &cr) // model only; blob stored verbatim
+		if err := s.store.InsertClassification(id, cr.Model, blob, time.Now().UTC()); err != nil {
+			log.Printf("warning: store classification for %s: %v", id, err)
+		}
+	}
+
 	// Save each attachment now that the parent receipt id exists, keyed to the
 	// receipt's date+amount so the files sort alongside it.
 	var attachmentNames []string

internal/web/web.go

@@ -75,6 +75,12 @@ func (s *Server) Routes() http.Handler {
 	mux.Handle("GET /recent/receipts", s.requireAuth(http.HandlerFunc(s.handleRecentReceipts)))
 	mux.Handle("GET /receipt/{id}/file", s.requireAuth(http.HandlerFunc(s.handleReceiptFile)))
 	mux.Handle("GET /attachment/{id}/file", s.requireAuth(http.HandlerFunc(s.handleAttachmentFile)))
+	mux.Handle("GET /ai", s.requireAuth(http.HandlerFunc(s.handleAI)))
+	mux.Handle("POST /ai/notes", s.requireAuth(http.HandlerFunc(s.handleAddNote)))
+	mux.Handle("POST /ai/notes/edit", s.requireAuth(http.HandlerFunc(s.handleEditNote)))
+	mux.Handle("POST /ai/notes/delete", s.requireAuth(http.HandlerFunc(s.handleDeleteNote)))
+	mux.Handle("GET /ai/review/{id}", s.requireAuth(http.HandlerFunc(s.handleReview)))
+	mux.Handle("POST /ai/review/{id}", s.requireAuth(http.HandlerFunc(s.handleReviewSubmit)))
 	mux.Handle("GET /manage", s.requireAuth(http.HandlerFunc(s.handleManage)))
 	mux.Handle("POST /manage/categories", s.requireAuth(http.HandlerFunc(s.handleAddCategory)))
 	mux.Handle("POST /manage/categories/rename", s.requireAuth(http.HandlerFunc(s.handleRenameCategory)))