Ship hardening wave: off-loop CSV inference, model_ids cap/dedupe, real detector Hub ids, comment fixes

Dockerfile

@@ -50,11 +50,13 @@ names = ['cardiffnlp/twitter-roberta-base-sentiment-latest', \
 [(AutoTokenizer.from_pretrained(n), AutoModelForSequenceClassification.from_pretrained(n)) for n in names]"
 
 # Detectors are baked differently — into local model directories, not the HF
-# cache. The registry gives detectors display names ("desklib-ai-text-detector-
-# v1.01"), not valid Hub IDs, so resolve_model_source() can only reach them via
-# their local_path; the Hub-name fallback that works for sentiment would 404
-# here. So snapshot_download each real detector repo into the exact path
-# resolve_model_source() computes at runtime (models/ sibling of the app dir).
+# cache. Each registry detector sets a local_path, and resolve_model_source()
+# checks that on-disk dir FIRST (falling back to the Hub name only if it is
+# absent), so runtime resolves detectors from local weights. snapshot_download
+# each real detector repo into the exact path resolve_model_source() computes at
+# runtime (models/ sibling of the app dir). The registry names are the real Hub
+# repos too (the same ids used as the snapshot_download source), so the Hub-name
+# fallback would resolve as well — exactly like the sentiment models above.
 #
 # All THREE detectors are baked, not lazy-loaded: the AI Detector tab's default
 # action runs /api/ai-detect/compare with no model_ids, which scores every

backend/app/main.py

@@ -71,7 +71,9 @@ if os.getenv("PUBLIC_DEPLOY") == "1":
     # this protects.
     limiter = Limiter(key_func=client_ip, application_limits=["30/minute"])
     app.state.limiter = limiter
-    app.add_exception_handler(RateLimitExceeded, _rate_limit_exceeded_handler)
+    # No app.add_exception_handler(RateLimitExceeded, ...): the middleware below
+    # catches RateLimitExceeded and returns the 429 itself, so the exception
+    # never propagates to an app-level handler — registering one would be dead code.
 
     # NOT slowapi's SlowAPIMiddleware: it resolves the handler by scanning
     # app.routes for an `endpoint` attribute, but FastAPI 0.139 wraps included

backend/app/model_registry.py

@@ -76,7 +76,7 @@ MODEL_REGISTRY: dict[str, ModelConfig] = {
     # AI text detector models
     "desklib-ai-detector": ModelConfig(
         id="desklib-ai-detector",
-        name="desklib-ai-text-detector-v1.01",
+        name="desklib/ai-text-detector-v1.01",
         task=ModelTask.AI_TEXT_DETECTION,
         labels=("human", "ai"),
         domain="general AI-written text detection",
@@ -91,7 +91,7 @@ MODEL_REGISTRY: dict[str, ModelConfig] = {
     ),
     "fakespot-ai-detector": ModelConfig(
         id="fakespot-ai-detector",
-        name="fakespot-roberta-base-ai-text-detection-v1",
+        name="fakespot-ai/roberta-base-ai-text-detection-v1",
         task=ModelTask.AI_TEXT_DETECTION,
         labels=("human", "ai"),
         domain="AI-generated review/text detection",
@@ -100,7 +100,7 @@ MODEL_REGISTRY: dict[str, ModelConfig] = {
     ),
     "oxidane-ai-detector": ModelConfig(
         id="oxidane-ai-detector",
-        name="oxidane-tmr-ai-text-detector",
+        name="Oxidane/tmr-ai-text-detector",
         task=ModelTask.AI_TEXT_DETECTION,
         labels=("human", "ai"),
         domain="general AI text detection",

backend/app/routes.py

@@ -1,6 +1,7 @@
 """API routes. Endpoints stay thin: validation lives in schemas (Pydantic),
 ML logic lives in SentimentModel — routes just wire the two together."""
 
+import asyncio
 import csv
 import io
 import os
@@ -198,7 +199,11 @@ async def analyze_csv(
     JSON, so each failure mode gets its own explicit 400 — the caller should
     always learn WHY their file was rejected."""
     reject_non_default_model_id(model_id)
-    raw = await file.read()
+    # Bounded read: pull at most MAX_CSV_BYTES + 1 bytes off the wire, never the
+    # whole upload. The +1 byte is all the size check below needs to fire, so the
+    # guard can't be defeated by a body larger than RAM — an attacker streaming a
+    # multi-GB file never gets more than ~5MB buffered here.
+    raw = await file.read(MAX_CSV_BYTES + 1)
     if len(raw) > MAX_CSV_BYTES:
         # 413 Payload Too Large, not 400: the upload isn't malformed — it's
         # simply too big, which is exactly what 413 means. Checked before any
@@ -236,7 +241,9 @@ async def analyze_csv(
     if not texts:
         raise HTTPException(status_code=400, detail="No non-empty rows found")
 
-    results = model.predict(texts)
+    # Inference is CPU-bound and blocking; run it off the event loop so
+    # concurrent requests stay responsive.
+    results = await asyncio.to_thread(model.predict, texts)
     items = [{"text": t, **r} for t, r in zip(texts, results)]
     return {"results": items, "aggregates": aggregate(results)}
 
@@ -268,13 +275,9 @@ async def compare(req: CompareRequest, request: Request):
     """
     model_ids = req.model_ids or DEFAULT_COMPARE_MODELS
 
-    # Public-deployment allowlist first: a disabled model is rejected before
-    # get_or_load_model can ever spend memory/bandwidth on its weights.
-    for model_id in model_ids:
-        reject_disabled_model(model_id)
-
-    # Validate every id BEFORE loading any weights — fail fast at the boundary
-    # so a bad id never costs a multi-hundred-MB load first.
+    # Validate identity BEFORE the allowlist: an unknown id must return a
+    # truthful 400, not masquerade as a 403 "disabled". Every id is checked
+    # BEFORE loading any weights, so a bad id never costs a multi-hundred-MB load.
     configs = []
     for model_id in model_ids:
         try:
@@ -292,12 +295,21 @@ async def compare(req: CompareRequest, request: Request):
             )
         configs.append((model_id, cfg))
 
+    # Public-deployment allowlist next: a known-but-disabled model is rejected
+    # before get_or_load_model can ever spend memory/bandwidth on its weights.
+    for model_id, _ in configs:
+        reject_disabled_model(model_id)
+
     results = []
     for model_id, cfg in configs:
         model = await get_or_load_model(request.app, model_id)
-        # perf_counter is monotonic — the correct clock for measuring a duration.
+        # Inference is CPU-bound and blocking; run it off the event loop (like
+        # the lazy load() in get_or_load_model) so concurrent requests stay
+        # responsive. perf_counter is monotonic — the right clock for a duration
+        # — and times the predict call itself (the thread hand-off is sub-ms
+        # noise; the expensive one-time load already happened before this loop).
         start = time.perf_counter()
-        prediction = model.predict([req.text])[0]
+        prediction = (await asyncio.to_thread(model.predict, [req.text]))[0]
         latency_ms = (time.perf_counter() - start) * 1000
         scores = prediction["scores"]
         results.append(
@@ -321,13 +333,12 @@ async def _score_detectors(request: Request, text: str, model_ids: list[str]) ->
     """Validate, load, and score a set of AI text detectors on one input.
 
     Shared by /api/ai-detect and /api/ai-detect/compare so both apply the same
-    guards in the same order: allowlist (403) → known id (400) → detector-only
-    (400), ALL before any weights load. Sentiment ids are refused here — mixing
-    the two task families is the one thing this endpoint exists to prevent.
+    guards in the same order: known id (400) → detector-only (400) → allowlist
+    (403), ALL before any weights load. Identity is checked before the allowlist
+    so an unknown id returns a truthful 400 instead of masquerading as a 403
+    "disabled". Sentiment ids are refused here — mixing the two task families is
+    the one thing this endpoint exists to prevent.
     """
-    for model_id in model_ids:
-        reject_disabled_model(model_id)
-
     configs = []
     for model_id in model_ids:
         try:
@@ -346,11 +357,19 @@ async def _score_detectors(request: Request, text: str, model_ids: list[str]) ->
             )
         configs.append((model_id, cfg))
 
+    # Public-deployment allowlist last: a known-but-disabled detector is rejected
+    # before get_or_load_model can ever spend memory/bandwidth on its weights.
+    for model_id, _ in configs:
+        reject_disabled_model(model_id)
+
     rows = []
     for model_id, cfg in configs:
         model = await get_or_load_model(request.app, model_id)
+        # Inference is CPU-bound and blocking; run it off the event loop so
+        # concurrent requests stay responsive. Latency times the predict call
+        # (thread hand-off is sub-ms noise; the one-time load already happened).
         start = time.perf_counter()
-        prediction = model.predict([text])[0]
+        prediction = (await asyncio.to_thread(model.predict, [text]))[0]
         latency_ms = (time.perf_counter() - start) * 1000
         scores = prediction["scores"]
         rows.append(

backend/app/schemas.py

@@ -7,11 +7,17 @@ assume clean input, which keeps it simple.
 
 from pydantic import BaseModel, Field, field_validator
 
+from app.model_registry import MODEL_REGISTRY
+
 # Character cap is a cheap pre-tokenization guard. The tokenizer still
 # truncates to 512 tokens, but rejecting huge payloads early protects the
 # server from pathological requests.
 MAX_CHARS = 2000
 MAX_BATCH = 500
+# A multi-model request can name at most every registry model once. Capping at
+# the registry size (and de-duplicating below) stops one anonymous request from
+# queueing thousands of inferences by padding or repeating model_ids.
+MAX_MODEL_IDS = len(MODEL_REGISTRY)
 
 
 class AnalyzeRequest(BaseModel):
@@ -88,9 +94,26 @@ class DynamicAnalyzeResponse(BaseModel):
     scores: dict[str, float]
 
 
-class CompareRequest(AnalyzeRequest):
+class _MultiModelRequest(AnalyzeRequest):
+    """Shared contract for endpoints that fan one input across several named
+    models. model_ids is capped at the registry size and de-duplicated
+    (order-preserving) at the boundary, so one anonymous request can never
+    queue thousands of inferences by repeating or padding the list."""
+
+    model_ids: list[str] | None = Field(default=None, max_length=MAX_MODEL_IDS)
+
+    @field_validator("model_ids")
+    @classmethod
+    def dedupe_model_ids(cls, v: list[str] | None) -> list[str] | None:
+        if v is None:
+            return v
+        # dict.fromkeys keeps the first occurrence of each id and preserves order.
+        return list(dict.fromkeys(v))
+
+
+class CompareRequest(_MultiModelRequest):
     # Omit model_ids to compare the default lineup (see routes.DEFAULT_COMPARE_MODELS).
-    model_ids: list[str] | None = None
+    pass
 
 
 class CompareItem(DynamicAnalyzeResponse):
@@ -114,10 +137,10 @@ class CompareResponse(BaseModel):
 # the softmax detectors (fakespot/oxidane) all funnel into the same shape.
 
 
-class AiDetectRequest(AnalyzeRequest):
+class AiDetectRequest(_MultiModelRequest):
     # Omit model_ids: /api/ai-detect uses the default detector; /api/ai-detect/compare
     # uses every detector (see routes).
-    model_ids: list[str] | None = None
+    pass
 
 
 class AiDetectItem(DynamicAnalyzeResponse):

backend/tests/test_ai_detect.py

@@ -58,7 +58,7 @@ def test_ai_detect_honors_explicit_detector(client_with_detectors):
 
 def test_ai_detect_row_carries_registry_metadata(client_with_detectors):
     result = client_with_detectors.post("/api/ai-detect", json={"text": "x"}).json()["result"]
-    assert result["name"] == "desklib-ai-text-detector-v1.01"
+    assert result["name"] == "desklib/ai-text-detector-v1.01"
     assert result["domain"]
     assert result["note"]
     assert isinstance(result["latency_ms"], (int, float))
@@ -134,6 +134,17 @@ def test_compare_rejects_sentiment_model(client_with_detectors):
     assert "sentiment" in resp.json()["detail"].lower()
 
 
+def test_compare_dedupes_duplicate_detector_ids(client_with_detectors):
+    # Same dedupe guard as /api/compare: a repeated detector id collapses to one
+    # row, so one request can't queue the same detector many times over.
+    resp = client_with_detectors.post(
+        "/api/ai-detect/compare",
+        json={"text": "x", "model_ids": ["desklib-ai-detector", "desklib-ai-detector"]},
+    )
+    assert resp.status_code == 200
+    assert [r["model_id"] for r in resp.json()["results"]] == ["desklib-ai-detector"]
+
+
 # --- cross-task rejection: detectors are refused by /api/compare ----------
 
 

backend/tests/test_compare.py

@@ -91,6 +91,40 @@ def test_compare_allows_models_on_the_allowlist(monkeypatch, client_with_compare
     assert len(resp.json()["results"]) == 2
 
 
+def test_compare_dedupes_duplicate_model_ids(client_with_compare_models):
+    # A repeated id must collapse to ONE inference/row — one anonymous request
+    # can't queue the same model dozens of times by padding the list.
+    resp = client_with_compare_models.post(
+        "/api/compare",
+        json={"text": "good", "model_ids": ["twitter-roberta", "twitter-roberta"]},
+    )
+    assert resp.status_code == 200
+    assert [r["model_id"] for r in resp.json()["results"]] == ["twitter-roberta"]
+
+
+def test_compare_rejects_over_cap_model_ids(client_with_compare_models):
+    # More ids than the registry holds is rejected at the boundary (422) before
+    # any weights load — a hard cap on how many inferences one request can queue.
+    from app.schemas import MAX_MODEL_IDS
+
+    too_many = [f"model-{i}" for i in range(MAX_MODEL_IDS + 1)]
+    resp = client_with_compare_models.post(
+        "/api/compare", json={"text": "good", "model_ids": too_many}
+    )
+    assert resp.status_code == 422
+
+
+def test_compare_unknown_id_is_400_even_under_allowlist(monkeypatch, client_with_compare_models):
+    # Guard order: an unknown id returns a truthful 400, NOT a 403 masquerading
+    # as "disabled", even when the ENABLED_MODELS allowlist is active.
+    monkeypatch.setenv("ENABLED_MODELS", "twitter-roberta")
+    resp = client_with_compare_models.post(
+        "/api/compare", json={"text": "good", "model_ids": ["not-a-real-model"]}
+    )
+    assert resp.status_code == 400
+    assert "not-a-real-model" in resp.json()["detail"]
+
+
 def test_lifespan_seeds_default_model_into_cache(monkeypatch):
     """The startup-loaded default must be seeded into model_cache so /api/compare
     reuses that one copy instead of loading a second ~500MB set of weights."""

backend/tests/test_csv.py

@@ -70,3 +70,13 @@ def test_csv_over_max_bytes_is_413(client_with_model):
     resp = _upload(client_with_model, csv_bytes)
     assert resp.status_code == 413
     assert "5MB" in resp.json()["detail"]
+
+
+def test_csv_far_over_max_bytes_still_413(client_with_model):
+    # A body many times the limit still 413s: the endpoint uses a BOUNDED read
+    # (file.read(MAX_CSV_BYTES + 1)), so the guard fires on the first MAX+1 bytes
+    # and can't be defeated by a body larger than RAM. Behavior is all we can
+    # assert here; the bounded read is what keeps that behavior cheap.
+    csv_bytes = b"text\n" + b"x" * (MAX_CSV_BYTES * 3)
+    resp = _upload(client_with_model, csv_bytes)
+    assert resp.status_code == 413

evals/ai_detection_report.md

@@ -71,7 +71,7 @@ Accuracy and macro F1 against the seed labels, plus single-text latency (p50 = t
 | fakespot-ai-detector | 0.567 | 0.562 | 8.5 | 32.9 |
 | oxidane-ai-detector | 0.600 | 0.524 | 6.5 | 8.5 |
 
-### `desklib-ai-detector` — desklib-ai-text-detector-v1.01
+### `desklib-ai-detector` — desklib/ai-text-detector-v1.01
 
 Per-class precision / recall / F1:
 
@@ -107,7 +107,7 @@ Misclassified (15 of 30) — P(ai) shows how confidently it was wrong:
 | There are many factors to consider, and ultimately the right choice depends on your own needs. | ambiguous | human | ai | 1.00 |
 | Overall, it was a memorable trip that offered both challenges and a few lessons I won't forget. | ambiguous | human | ai | 1.00 |
 
-### `fakespot-ai-detector` — fakespot-roberta-base-ai-text-detection-v1
+### `fakespot-ai-detector` — fakespot-ai/roberta-base-ai-text-detection-v1
 
 Per-class precision / recall / F1:
 
@@ -141,7 +141,7 @@ Misclassified (13 of 30) — P(ai) shows how confidently it was wrong:
 | A thorough strategy such as this secures the best results and encourages steady expansion in every area. | ai_paraphrased | ai | human | 0.19 |
 | my two cents plus the draft: The proposed framework offers a robust and scalable foundation for future development. | mixed_human_ai | ai | human | 0.14 |
 
-### `oxidane-ai-detector` — oxidane-tmr-ai-text-detector
+### `oxidane-ai-detector` — Oxidane/tmr-ai-text-detector
 
 Per-class precision / recall / F1:
 

frontend/README.md

@@ -1,32 +1,17 @@
-# React + TypeScript + Vite
+# SentimentScope — frontend
 
-This template provides a minimal setup to get React working in Vite with HMR and some Oxlint rules.
+React + TypeScript + Vite SPA for SentimentScope. See the [root README](../README.md)
+for the full project overview, architecture, and deployment topologies.
 
-Currently, two official plugins are available:
+## Dev commands
 
-- [@vitejs/plugin-react](https://github.com/vitejs/vite-plugin-react/blob/main/packages/plugin-react) uses [Oxc](https://oxc.rs)
-- [@vitejs/plugin-react-swc](https://github.com/vitejs/vite-plugin-react/blob/main/packages/plugin-react-swc) uses [SWC](https://swc.rs/)
-
-## React Compiler
-
-The React Compiler is not enabled on this template because of its impact on dev & build performances. To add it, see [this documentation](https://react.dev/learn/react-compiler/installation).
-
-## Expanding the Oxlint configuration
-
-If you are developing a production application, we recommend enabling type-aware lint rules by installing `oxlint-tsgolint` and editing `.oxlintrc.json`:
-
-```json
-{
-  "$schema": "./node_modules/oxlint/configuration_schema.json",
-  "plugins": ["react", "typescript", "oxc"],
-  "options": {
-    "typeAware": true
-  },
-  "rules": {
-    "react/rules-of-hooks": "error",
-    "react/only-export-components": ["warn", { "allowConstantExport": true }]
-  }
-}
+```bash
+npm install       # install dependencies
+npm run dev       # Vite dev server (proxies /api -> backend on :8000)
+npm test -- --run # run the vitest suite once
+npm run lint      # oxlint
+npm run build     # type-check (tsc -b) + production build
 ```
 
-See the [Oxlint rules documentation](https://oxc.rs/docs/guide/usage/linter/rules) for the full list of rules and categories.
+The UI only ever calls relative `/api/...` paths; the Vite dev proxy, nginx
+(compose), and FastAPI static mount (Spaces) each wire those to the backend.

frontend/nginx.conf

@@ -6,6 +6,9 @@ server {
     # container. Mirrors the Vite dev proxy — the frontend build is identical
     # in both environments.
     location /api/ {
+        # Match the backend's 5MB CSV cap so oversized uploads are rejected at
+        # the edge, before they ever reach uvicorn.
+        client_max_body_size 5m;
         proxy_pass http://backend:8000;
     }
 

frontend/src/App.test.tsx

@@ -22,6 +22,7 @@ beforeEach(() => {
       {
         id: "twitter-roberta",
         name: "cardiffnlp/twitter-roberta-base-sentiment-latest",
+        task: "sentiment",
         labels: ["negative", "neutral", "positive"],
         domain: "social / short English text",
         note: "",

frontend/src/api.ts

@@ -52,6 +52,7 @@ export interface ModelInfo {
 export interface ModelSummary {
   id: string;
   name: string;
+  task: string;
   labels: string[];
   domain: string;
   note: string;
@@ -83,11 +84,6 @@ export interface AiDetectItem {
   note: string;
 }
 
-export interface AiDetectResponse {
-  result: AiDetectItem;
-  warning: string;
-}
-
 export interface AiDetectCompareResponse {
   results: AiDetectItem[];
   disagreement: boolean;
@@ -133,8 +129,5 @@ export const getModels = (task?: "sentiment" | "ai_text_detection") =>
 export const compareModels = (text: string, model_ids?: string[]) =>
   request<{ results: CompareItem[] }>("/api/compare", postJson({ text, model_ids }));
 
-export const detectAiText = (text: string) =>
-  request<AiDetectResponse>("/api/ai-detect", postJson({ text }));
-
 export const compareAiDetectors = (text: string, model_ids?: string[]) =>
   request<AiDetectCompareResponse>("/api/ai-detect/compare", postJson({ text, model_ids }));

frontend/src/components/AiTextDetector.test.tsx

@@ -16,7 +16,7 @@ const WARNING =
 
 const item = (over: Partial<AiDetectItem> = {}): AiDetectItem => ({
   model_id: "desklib-ai-detector",
-  name: "desklib-ai-text-detector-v1.01",
+  name: "desklib/ai-text-detector-v1.01",
   domain: "general AI-written text detection",
   label: "ai",
   scores: { human: 0.1, ai: 0.9 },
@@ -57,7 +57,7 @@ it("renders a readout per detector with its verdict, scores, latency, and note",
     results: [
       item({
         model_id: "desklib-ai-detector",
-        name: "desklib-ai-text-detector-v1.01",
+        name: "desklib/ai-text-detector-v1.01",
         label: "ai",
         scores: { human: 0.02, ai: 0.98 },
         confidence: 0.98,
@@ -66,7 +66,7 @@ it("renders a readout per detector with its verdict, scores, latency, and note",
       }),
       item({
         model_id: "fakespot-ai-detector",
-        name: "fakespot-roberta-base-ai-text-detection-v1",
+        name: "fakespot-ai/roberta-base-ai-text-detection-v1",
         label: "human",
         scores: { human: 0.7, ai: 0.3 },
         confidence: 0.7,
@@ -85,7 +85,7 @@ it("renders a readout per detector with its verdict, scores, latency, and note",
   const desklib = await screen.findByRole("group", { name: "desklib-ai-detector" });
   const fakespot = screen.getByRole("group", { name: "fakespot-ai-detector" });
 
-  expect(within(desklib).getByText("desklib-ai-text-detector-v1.01")).toBeInTheDocument();
+  expect(within(desklib).getByText("desklib/ai-text-detector-v1.01")).toBeInTheDocument();
   expect(within(desklib).getByText("general AI-written text detection")).toBeInTheDocument();
   expect(within(desklib).getByText("2.0%")).toBeInTheDocument(); // P(human)
   expect(within(desklib).getAllByText("98.0%").length).toBeGreaterThan(0); // P(ai) + confidence

frontend/src/components/CompareModels.test.tsx

@@ -15,6 +15,7 @@ const registry = {
     {
       id: "twitter-roberta",
       name: "cardiffnlp/twitter-roberta-base-sentiment-latest",
+      task: "sentiment",
       labels: ["negative", "neutral", "positive"],
       domain: "social / short English text",
       note: "",
@@ -24,6 +25,7 @@ const registry = {
     {
       id: "distilbert-sst2",
       name: "distilbert-base-uncased-finetuned-sst-2-english",
+      task: "sentiment",
       labels: ["negative", "positive"],
       domain: "general binary sentiment",
       note: "",
@@ -33,6 +35,7 @@ const registry = {
     {
       id: "finbert",
       name: "ProsusAI/finbert",
+      task: "sentiment",
       labels: ["negative", "neutral", "positive"],
       domain: "financial text",
       note: "Useful for finance/news sentences; misleading outside that domain.",

frontend/src/components/CompareModels.tsx

@@ -74,6 +74,7 @@ function CompareForm({ registryPromise }: { registryPromise: Promise<ModelSummar
         <textarea
           className="w-full rounded-lg border border-slate-300 bg-white p-3 focus:border-slate-500 focus:outline-none focus-visible:ring-2 focus-visible:ring-slate-300"
           rows={3}
+          maxLength={2000}
           aria-label="Text to compare"
           value={text}
           onChange={(e) => setText(e.target.value)}