File size: 21,404 Bytes
c4c41cc
a380f06
c4c41cc
a380f06
 
 
e2d3e54
 
a380f06
 
 
 
 
 
 
 
 
 
 
e2d3e54
 
 
 
 
 
 
 
 
 
a380f06
 
 
 
 
 
 
 
 
c4c41cc
a380f06
 
 
c4c41cc
a380f06
c4c41cc
 
 
 
a380f06
 
e2d3e54
a380f06
 
 
e2d3e54
 
 
a380f06
 
 
 
 
e2d3e54
a380f06
 
 
 
 
 
 
 
 
 
 
 
c4c41cc
 
 
 
a380f06
c4c41cc
a380f06
c4c41cc
 
 
 
 
 
 
 
 
 
a380f06
 
c4c41cc
a380f06
c4c41cc
a380f06
c4c41cc
a380f06
c4c41cc
a380f06
c4c41cc
a380f06
c4c41cc
a380f06
c4c41cc
 
a380f06
 
 
 
 
 
 
 
 
eb515b5
c4c41cc
 
a380f06
e2d3e54
 
 
 
 
 
 
a380f06
 
e2d3e54
 
 
 
 
 
 
 
 
eb515b5
e2d3e54
 
c4c41cc
a380f06
eb515b5
a380f06
c4c41cc
a380f06
eb515b5
c4c41cc
eb515b5
a380f06
e2d3e54
a380f06
eb515b5
 
 
a380f06
e2d3e54
a380f06
 
 
 
 
eb515b5
a380f06
eb515b5
c4c41cc
eb515b5
a380f06
e2d3e54
c4c41cc
a380f06
 
 
 
 
 
c4c41cc
a380f06
 
 
c4c41cc
eb515b5
a380f06
 
 
 
 
 
 
eb515b5
c4c41cc
 
 
eb515b5
 
a380f06
e2d3e54
 
 
 
 
a380f06
eb515b5
 
c4c41cc
 
 
 
 
 
 
eb515b5
a380f06
e2d3e54
 
 
 
eb515b5
 
 
 
c4c41cc
 
 
 
 
a380f06
e2d3e54
eb515b5
a380f06
 
e2d3e54
a380f06
e2d3e54
a380f06
 
 
e2d3e54
a380f06
 
eb515b5
 
 
c4c41cc
eb515b5
 
a380f06
e2d3e54
eb515b5
 
 
e2d3e54
eb515b5
 
 
 
c4c41cc
 
 
 
 
eb515b5
c4c41cc
 
 
 
a380f06
 
 
 
eb515b5
 
 
 
 
 
 
 
 
a380f06
e2d3e54
a380f06
eb515b5
a380f06
eb515b5
a380f06
 
eb515b5
 
 
 
 
 
a380f06
 
e2d3e54
eb515b5
a380f06
e2d3e54
 
eb515b5
 
 
 
 
 
 
a380f06
e2d3e54
 
eb515b5
 
e2d3e54
 
a380f06
eb515b5
 
a380f06
eb515b5
 
 
a380f06
c4c41cc
a380f06
eb515b5
 
a380f06
e2d3e54
c4c41cc
e2d3e54
a380f06
 
 
 
eb515b5
 
 
 
 
 
 
 
a380f06
 
 
 
0f9daa0
a380f06
 
eb515b5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a380f06
 
eb515b5
 
 
 
a380f06
 
 
eb515b5
 
c4c41cc
 
eb515b5
a380f06
eb515b5
a380f06
eb515b5
 
 
 
 
 
a380f06
e2d3e54
 
 
 
 
 
 
 
 
c4c41cc
 
 
 
 
 
 
 
eb515b5
 
 
 
c4c41cc
 
 
 
 
eb515b5
 
c4c41cc
 
 
 
 
 
eb515b5
c4c41cc
 
eb515b5
 
 
c4c41cc
eb515b5
c4c41cc
 
eb515b5
 
 
c4c41cc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eb515b5
 
c4c41cc
eb515b5
 
 
 
 
 
 
c4c41cc
eb515b5
c4c41cc
 
eb515b5
c4c41cc
eb515b5
c4c41cc
eb515b5
 
 
 
 
c4c41cc
 
 
eb515b5
 
 
 
 
 
c4c41cc
eb515b5
c4c41cc
 
eb515b5
 
 
c4c41cc
 
eb515b5
a380f06
eb515b5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a380f06
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
import os, re, json, time, math
from dataclasses import dataclass
from typing import List, Dict, Tuple, Optional

import gradio as gr

# Optional imports for email classifier (loaded lazily).
# Space still runs if these aren't available (pure lexical fallback).
try:
    import torch
    from transformers import AutoTokenizer, AutoModelForSequenceClassification
except Exception:
    torch = None
    AutoTokenizer = None
    AutoModelForSequenceClassification = None

# =========================
# Config (env-overridable)
# =========================
EMAIL_CLASSIFIER_ID   = os.getenv("EMAIL_CLASSIFIER_ID", "your-username/mini-phish")  # <- swap to your HF repo when ready
EMAIL_BACKBONE_ID     = os.getenv("EMAIL_BACKBONE_ID", "microsoft/MiniLM-L6-H384-uncased")
THRESHOLD_TAU         = float(os.getenv("THRESHOLD_TAU", "0.40"))
MAX_SEQ_LEN           = int(os.getenv("MAX_SEQ_LEN", "320"))
SUBJECT_TOKEN_BUDGET  = int(os.getenv("SUBJECT_TOKEN_BUDGET", "64"))
FUSION_EMAIL_W        = float(os.getenv("FUSION_EMAIL_W", "0.6"))
FUSION_URL_W          = float(os.getenv("FUSION_URL_W", "0.4"))
URL_OVERRIDE_HIGH     = float(os.getenv("URL_OVERRIDE_HIGH", "0.85"))
URL_OVERRIDE_KW       = float(os.getenv("URL_OVERRIDE_KW", "0.70"))
ALLOWLIST_SAFE_CAP    = float(os.getenv("ALLOWLIST_SAFE_CAP", "0.15"))

# =========================
# Simple data classes
# =========================
@dataclass
class UrlResult:
    url: str
    risk: float
    reasons: List[str]
    contrib: Dict[str, float]     # per‑reason contribution for transparency

@dataclass
class EmailResult:
    p_email: float                # final probability after boosts
    kw_hits: List[str]
    strong_hits: List[str]        # subset of kw_hits considered strong
    token_counts: Dict[str, int]  # {"subject_tokens":..,"body_tokens":..,"sequence_len":..}
    p_raw: Optional[float]        # raw model probability (before boosts); None in lexical fallback
    path: Optional[str]           # "classifier" | "backbone" | None (lexical)

# =========================
# URL extraction & heuristics (swap with your real URL model when ready)
# =========================
URL_REGEX = r'(?i)\b((?:https?://|www\.)[^\s<>")]+)'

SUSPICIOUS_TLDS = {
    ".xyz", ".top", ".click", ".link", ".ru", ".cn", ".country", ".gq", ".ga", ".ml", ".tk"
}
SHORTENERS = {"bit.ly","t.co","tinyurl.com","goo.gl","ow.ly","is.gd","cutt.ly","tiny.one","lnkd.in"}

def extract_urls(text: str) -> List[str]:
    if not text: return []
    urls = re.findall(URL_REGEX, text)
    uniq, seen = [], set()
    for u in urls:
        u = u.strip().strip(').,;\'"')
        if u and u not in seen:
            uniq.append(u)
            seen.add(u)
    return uniq

def url_host(url: str) -> str:
    host = re.sub(r"^https?://", "", url, flags=re.I).split("/")[0].lower()
    return host

def score_url_heuristic(url: str) -> UrlResult:
    """
    Heuristic scoring with a transparent per‑reason contribution map.
    This keeps the POC explainable and makes the Forensics panel richer.
    """
    host = url_host(url)
    score = 0.0
    reasons = []
    contrib = {}

    def add(amount: float, tag: str):
        nonlocal score
        score += amount
        reasons.append(tag)
        contrib[tag] = round(contrib.get(tag, 0.0) + amount, 3)

    base = 0.05
    add(base, "base")

    if len(url) > 140:
        add(0.15, "very_long_url")
    if "@" in url or "%" in url:
        add(0.20, "special_chars")
    if any(host.endswith(t) for t in SUSPICIOUS_TLDS):
        add(0.35, "suspicious_tld")
    if any(s in host for s in SHORTENERS):
        add(0.50, "shortener")
    if host.count(".") >= 3:
        add(0.20, "deep_subdomain")
    if len(re.findall(r"[A-Z]", url)) > 16:
        add(0.10, "mixed_case")

    score = min(score, 1.0)
    return UrlResult(url=url, risk=score, reasons=reasons, contrib=contrib)

def score_urls(urls: List[str]) -> List[UrlResult]:
    return [score_url_heuristic(u) for u in urls]

# =========================
# Email classifier with fallback
# =========================
_tokenizer = None
_model = None
_model_loaded_from = None  # "classifier", "backbone", or None
_model_load_ms = None
_model_quantized = False

# Strong vs normal cues (lowercase)
STRONG_CUES = [
    "otp", "one-time password", "one time password", "cvv", "pin", "pan",
    "password", "bank details", "netbanking", "debit card", "credit card",
    "lottery", "jackpot", "prize", "reward", "winner", "you have won",
    "send otp", "share otp", "confirm otp", "verify otp",
    "account restricted", "reactivate account", "unlock your account"
]

NORMAL_CUES = [
    "verify your account", "update your password", "immediately",
    "within 24 hours", "suspended", "unusual activity", "confirm",
    "login", "click", "invoice", "payment", "security alert",
    "urgent", "limited time"
]

LEXICAL_CUES = sorted(set(STRONG_CUES + NORMAL_CUES))

def load_email_model() -> Tuple[object, object, str]:
    """Try to load EMAIL_CLASSIFIER_ID; on failure, fall back to backbone with small head.
       Apply dynamic int8 quantization for CPU if available."""
    global _tokenizer, _model, _model_loaded_from, _model_load_ms, _model_quantized
    if _tokenizer is not None and _model is not None:
        return _tokenizer, _model, _model_loaded_from

    start = time.perf_counter()
    if AutoTokenizer is None or AutoModelForSequenceClassification is None or torch is None:
        _model_loaded_from = None
        _model_load_ms = round((time.perf_counter() - start) * 1000, 2)
        return None, None, _model_loaded_from  # environment without torch/transformers

    # Preferred classifier
    try:
        _tokenizer = AutoTokenizer.from_pretrained(EMAIL_CLASSIFIER_ID)
        _model = AutoModelForSequenceClassification.from_pretrained(EMAIL_CLASSIFIER_ID)
        _model_loaded_from = "classifier"
    except Exception:
        # Fallback: backbone + fresh 2-class head
        try:
            _tokenizer = AutoTokenizer.from_pretrained(EMAIL_BACKBONE_ID)
            _model = AutoModelForSequenceClassification.from_pretrained(
                EMAIL_BACKBONE_ID, num_labels=2, problem_type="single_label_classification"
            )
            _model_loaded_from = "backbone"
        except Exception:
            _tokenizer, _model, _model_loaded_from = None, None, None
            _model_load_ms = round((time.perf_counter() - start) * 1000, 2)
            return None, None, _model_loaded_from

    # Dynamic quantization (CPU)
    _model_quantized = False
    try:
        _model.eval()
        _model.to("cpu")
        if hasattr(torch, "quantization"):
            from torch.quantization import quantize_dynamic
            _model = quantize_dynamic(_model, {torch.nn.Linear}, dtype=torch.qint8)  # type: ignore
            _model_quantized = True
    except Exception:
        pass

    _model_load_ms = round((time.perf_counter() - start) * 1000, 2)
    return _tokenizer, _model, _model_loaded_from

def _truncate_for_budget(tokens_subject: List[int], tokens_body: List[int], max_len: int, subj_budget: int):
    subj = tokens_subject[:subj_budget]
    remain = max(0, max_len - len(subj))
    body = tokens_body[:remain]
    return subj + body

def score_email(subject: str, body: str) -> Tuple[EmailResult, Dict]:
    """Return EmailResult + debug dict with probability, hits, boosts, timings, token counts, and model info."""
    dbg = {"path": None, "p_raw": None, "boost_from_strong": 0.0, "boost_from_normal": 0.0,
           "timing_ms": {}, "token_counts": {}, "model_info": {}}

    t0 = time.perf_counter()
    text = (subject or "") + "\n" + (body or "")
    low = text.lower()

    strong_hits = [c for c in STRONG_CUES if c in low]
    normal_hits = [c for c in NORMAL_CUES if c in low]
    all_hits = sorted(set(strong_hits + normal_hits))

    tok, mdl, path = load_email_model()
    dbg["path"] = path
    dbg["model_info"] = {
        "loaded_from": path,
        "classifier_id": EMAIL_CLASSIFIER_ID,
        "backbone_id": EMAIL_BACKBONE_ID,
        "quantized": _model_quantized,
        "model_load_ms": _model_load_ms
    }

    if tok is None or mdl is None:
        # Pure lexical fallback (no model available):
        base = 0.10
        p_email = base + 0.18 * len(strong_hits) + 0.07 * len(normal_hits)
        p_email = float(max(0.01, min(0.99, p_email)))
        dbg["p_raw"] = None
        dbg["boost_from_strong"] = 0.18 * len(strong_hits)
        dbg["boost_from_normal"] = 0.07 * len(normal_hits)
        dbg["timing_ms"]["email_infer"] = round((time.perf_counter() - t0) * 1000, 2)
        dbg["token_counts"] = {"subject_tokens": 0, "body_tokens": 0, "sequence_len": 0}
        return EmailResult(
            p_email=p_email, kw_hits=all_hits, strong_hits=strong_hits,
            token_counts=dbg["token_counts"], p_raw=None, path=path
        ), dbg

    # Model path (MiniLM or your classifier)
    enc_t0 = time.perf_counter()
    encoded_subj = tok.encode(subject or "", add_special_tokens=False)
    encoded_body = tok.encode(body or "", add_special_tokens=False)
    input_ids = _truncate_for_budget(encoded_subj, encoded_body, MAX_SEQ_LEN - 2, SUBJECT_TOKEN_BUDGET)
    input_ids = [tok.cls_token_id] + input_ids + [tok.sep_token_id]
    attn_mask = [1] * len(input_ids)

    ids = torch.tensor([input_ids], dtype=torch.long)
    mask = torch.tensor([attn_mask], dtype=torch.long)

    with torch.no_grad():
        out = mdl(input_ids=ids, attention_mask=mask)
    if hasattr(out, "logits"):
        logits = out.logits[0].detach().cpu().numpy().tolist()
        exps = [math.exp(x) for x in logits]
        p_raw = float(exps[1] / (exps[0] + exps[1]))  # assume label 1 = phishing
    else:
        p_raw = 0.5

    # Nudge with cues: stronger boost for strong hits
    boost_s = 0.10 * len(strong_hits)
    boost_n = 0.03 * len(normal_hits)
    p_email = float(max(0.01, min(0.99, p_raw + boost_s + boost_n)))

    dbg["p_raw"] = round(p_raw, 3)
    dbg["boost_from_strong"] = round(boost_s, 3)
    dbg["boost_from_normal"] = round(boost_n, 3)
    dbg["timing_ms"]["email_infer"] = round((time.perf_counter() - enc_t0) * 1000, 2)
    dbg["token_counts"] = {
        "subject_tokens": len(encoded_subj),
        "body_tokens": len(encoded_body),
        "sequence_len": len(input_ids)
    }

    return EmailResult(
        p_email=p_email, kw_hits=all_hits, strong_hits=strong_hits,
        token_counts=dbg["token_counts"], p_raw=p_raw, path=path
    ), dbg

# =========================
# Fusion
# =========================
def fuse(email_res: EmailResult, url_results: List[UrlResult], allowlist_domains: List[str]) -> Tuple[Dict, Dict]:
    """Return fused decision dict + debug dict explaining the math & overrides."""
    fdbg = {
        "weights": {"email": FUSION_EMAIL_W, "url": FUSION_URL_W},
        "threshold_tau": THRESHOLD_TAU,
        "overrides": {"url_high": URL_OVERRIDE_HIGH, "url_kw": URL_OVERRIDE_KW, "allowlist_safe_cap": ALLOWLIST_SAFE_CAP},
        "applied_overrides": [],
    }

    r_url_max = max([u.risk for u in url_results], default=0.0)
    no_urls = (len(url_results) == 0)

    # Allowlist check
    allowlist_hit = False
    matched_allow = None
    for u in url_results:
        h = url_host(u.url)
        for d in [d.strip().lower() for d in allowlist_domains if d.strip()]:
            if h.endswith(d):
                allowlist_hit = True
                matched_allow = d
                break
        if allowlist_hit:
            break

    # Base fusion
    r_before = FUSION_EMAIL_W * email_res.p_email + FUSION_URL_W * r_url_max

    # URL-driven overrides
    kw_flag = 1 if email_res.kw_hits else 0
    r_after = r_before
    if r_url_max >= URL_OVERRIDE_HIGH:
        r_after = max(r_after, 0.90)
        fdbg["applied_overrides"].append("URL_OVERRIDE_HIGH")
    elif kw_flag and r_url_max >= URL_OVERRIDE_KW:
        r_after = max(r_after, 0.90)
        fdbg["applied_overrides"].append("URL_OVERRIDE_KW")

    # Email-only strong-cue override
    if no_urls and len(email_res.strong_hits) > 0:
        r_after = max(r_after, 0.85)
        fdbg["applied_overrides"].append("EMAIL_ONLY_STRONG_CUES")

    # Allowlist cap
    if allowlist_hit:
        r_after = min(r_after, ALLOWLIST_SAFE_CAP)
        fdbg["applied_overrides"].append(f"ALLOWLIST({matched_allow})")

    verdict = "UNSAFE" if r_after >= THRESHOLD_TAU else "SAFE"

    fused = {
        "P_email": round(email_res.p_email, 3),
        "P_email_raw": round(email_res.p_raw, 3) if email_res.p_raw is not None else None,
        "R_url_max": round(r_url_max, 3),
        "R_total": round(r_after, 3),
        "R_total_before_overrides": round(r_before, 3),
        "kw_hits": email_res.kw_hits,
        "strong_hits": email_res.strong_hits,
        "token_counts": email_res.token_counts,
        "no_urls": no_urls,
        "allowlist_hit": allowlist_hit,
        "verdict": verdict
    }

    fdbg.update({
        "components": {"P_email": fused["P_email"], "R_url_max": fused["R_url_max"]},
        "no_urls": no_urls,
        "allowlist_hit": allowlist_hit,
        "matched_allow": matched_allow
    })
    return fused, fdbg

# =========================
# Gradio UI
# =========================
with gr.Blocks(title="PhishingMail-Lab") as demo:
    gr.Markdown("# 🧪 PhishingMail‑Lab\n**POC** — Hybrid (email + URL) with explainable forensics.")

    with gr.Row():
        with gr.Column(scale=3):
            subject = gr.Textbox(label="Subject", placeholder="Subject: Important account update")
            body = gr.Textbox(label="Email Body (paste text or HTML)", lines=12, placeholder="Paste the email content here...")
            with gr.Row():
                allowlist = gr.Textbox(label="Allowlist domains (comma-separated)", placeholder="microsoft.com, amazon.in")
                tau = gr.Slider(0, 1, value=THRESHOLD_TAU, step=0.01, label="Decision Threshold τ")
            analyze_btn = gr.Button("Analyze", variant="primary")

            verdict = gr.Label(label="Verdict")
            # Banner under verdict
            context_banner = gr.Markdown(visible=False)
            fusion_json = gr.JSON(label="Fusion & Flags")
            url_table = gr.Dataframe(headers=["URL","Risk","Reasons"], label="Per‑URL risk (heuristics demo)", interactive=False)

        # Forensics column
        with gr.Column(scale=2):
            gr.Markdown("### 🔎 Forensics")
            forensics_json = gr.JSON(label="Forensics (structured log)")
            forensics_md = gr.Markdown(label="Forensics (human‑readable)")

    def run(subject_text, body_text, allowlist_text, tau_val):
        # Timers for forensics
        t_all = time.perf_counter()

        # Update threshold
        global THRESHOLD_TAU
        THRESHOLD_TAU = float(tau_val)

        # URL pipeline
        t0 = time.perf_counter()
        raw_text = (subject_text or "") + "\n" + (body_text or "")
        urls = list(dict.fromkeys(extract_urls(raw_text)))  # uniq & ordered
        t1 = time.perf_counter()
        url_results = score_urls(urls)
        t2 = time.perf_counter()

        # Email pipeline
        email_res, email_dbg = score_email(subject_text or "", body_text or "")

        # Fusion
        allow_domains = [d.strip().lower() for d in (allowlist_text or "").split(",") if d.strip()]
        fused, fuse_dbg = fuse(email_res, url_results, allow_domains)

        # Build banner text/visibility
        banners = []
        if fused.get("no_urls"):
            banners.append("⚠️ **No URLs found** — decision based **only on email body**.")
        if fused.get("allowlist_hit"):
            banners.append("🛈 **Allowlist active** — risk **capped** for trusted domain.")
        banner_text = "<br>".join(banners) if banners else ""
        banner_visible = bool(banners)

        # Forensics JSON (deeper detail)
        per_url = [{
            "url": u.url,
            "risk": round(u.risk,3),
            "reasons": u.reasons,
            "contrib": u.contrib
        } for u in url_results]

        fx = {
            "config": {
                "weights": {"email": FUSION_EMAIL_W, "url": FUSION_URL_W},
                "threshold_tau": THRESHOLD_TAU,
                "overrides": {
                    "url_high": URL_OVERRIDE_HIGH,
                    "url_kw": URL_OVERRIDE_KW,
                    "allowlist_safe_cap": ALLOWLIST_SAFE_CAP
                },
                "model_ids": {"classifier": EMAIL_CLASSIFIER_ID, "backbone": EMAIL_BACKBONE_ID}
            },
            "input_summary": {
                "chars_subject": len(subject_text or ""),
                "chars_body": len(body_text or ""),
                "num_urls": len(urls),
                "allowlist_domains": allow_domains
            },
            "email": {
                "path": email_dbg["path"] or "lexical-fallback",
                "p_email_final": fused["P_email"],
                "p_email_raw": email_dbg["p_raw"],
                "boost_from_strong": email_dbg["boost_from_strong"],
                "boost_from_normal": email_dbg["boost_from_normal"],
                "token_counts": email_dbg["token_counts"],
                "kw_hits": email_res.kw_hits,
                "strong_hits": email_res.strong_hits,
                "model_info": email_dbg["model_info"]
            },
            "urls": per_url,
            "fusion": {
                "equation": f"R_total = {FUSION_EMAIL_W} * P_email + {FUSION_URL_W} * R_url_max",
                "values": {
                    "P_email": fused["P_email"],
                    "R_url_max": fused["R_url_max"],
                    "R_total_before_overrides": fused["R_total_before_overrides"],
                    "R_total_final": fused["R_total"],
                    "overrides_applied": fuse_dbg["applied_overrides"]
                },
                "decision": {
                    "threshold_tau": THRESHOLD_TAU,
                    "verdict": fused["verdict"]
                },
                "flags": {
                    "no_urls": fused["no_urls"],
                    "allowlist_hit": fused["allowlist_hit"]
                }
            },
            "timings_ms": {
                "model_load": email_dbg["model_info"]["model_load_ms"],
                "url_extract": round((t1 - t0) * 1000, 2),
                "url_score": round((t2 - t1) * 1000, 2),
                "email_infer": email_dbg["timing_ms"].get("email_infer"),
                "total": round((time.perf_counter() - t_all) * 1000, 2)
            }
        }

        # Forensics Markdown (human‑readable, denser detail)
        lines = []
        lines.append(f"**Verdict:** `{fused['verdict']}`  |  **R_total:** `{fused['R_total']}`  (before: `{fused['R_total_before_overrides']}`)  |  **τ:** `{THRESHOLD_TAU}`")
        lines.append(f"**Fusion:**  R = {FUSION_EMAIL_W}×P_email + {FUSION_URL_W}×R_url_max  →  {FUSION_EMAIL_W}×{fused['P_email']} + {FUSION_URL_W}×{fused['R_url_max']}")
        if fuse_dbg["applied_overrides"]:
            lines.append(f"**Overrides:** {', '.join(fuse_dbg['applied_overrides'])}")
        else:
            lines.append("**Overrides:** (none)")
        if fused["no_urls"]:
            lines.append("• No URLs found → email‑only decision path.")
        if fused["allowlist_hit"]:
            lines.append("• Allowlist matched → risk capped.")
        lines.append("")
        lines.append(f"**Email path:** `{email_dbg['path'] or 'lexical-fallback'}`  |  p_raw={email_dbg['p_raw']}  |  +strong={email_dbg['boost_from_strong']}  |  +normal={email_dbg['boost_from_normal']}")
        tc = email_dbg["token_counts"]
        lines.append(f"• Tokens: subject={tc.get('subject_tokens',0)}, body={tc.get('body_tokens',0)}, sequence_len={tc.get('sequence_len',0)} (max={MAX_SEQ_LEN})  |  subject_budget={SUBJECT_TOKEN_BUDGET}")
        if email_res.strong_hits:
            lines.append(f"• Strong cues: {', '.join(email_res.strong_hits)}")
        if email_res.kw_hits:
            lines.append(f"• All cues: {', '.join(email_res.kw_hits)}")
        lines.append("")
        if per_url:
            lines.append("**URLs & contributions:**")
            for u in per_url:
                contrib_str = ", ".join([f"{k}:{v}" for k,v in u["contrib"].items()])
                lines.append(f"• {u['url']}  → risk={u['risk']}  |  reasons=({', '.join(u['reasons']) or 'none'})  |  contrib=({contrib_str or 'n/a'})")
        else:
            lines.append("**URLs:** (none)")
        lines.append("")
        lines.append(f"**Model info:** loaded_from={email_dbg['model_info']['loaded_from']}, quantized={email_dbg['model_info']['quantized']}, load_ms={email_dbg['model_info']['model_load_ms']}")
        lines.append("")
        lines.append("**Timings (ms):** " + json.dumps(fx["timings_ms"]))

        forensic_markdown = "\n".join(lines)

        rows = [[u.url, round(u.risk,3), ", ".join(u.reasons)] for u in url_results]
        return (
            fused["verdict"],
            gr.update(value=banner_text, visible=banner_visible),
            fused,
            rows,
            fx,
            forensic_markdown
        )

    analyze_btn.click(
        run,
        [subject, body, allowlist, tau],
        [verdict, context_banner, fusion_json, url_table, forensics_json, forensics_md]
    )

if __name__ == "__main__":
    demo.launch()