Optimized with Modaic :) 70.00% -> 90.00%

program.json

@@ -3,7 +3,7 @@
   "train": [],
   "demos": [],
   "signature": {
-    "instructions": "You are given an email with two fields: subject and body.\n\nTask: classify the email as spam or not spam.\n\nOutput format (strict):\n- Return exactly two fields in this order:\n  1) is_spam: a single boolean value exactly \"True\" or \"False\"\n  2) completed: always output the literal marker \"[[ ## completed ## ]]\"\n- Do not output any other keys/fields/markers or any free-form text.\n\nClassification guidance:\n- Mark is_spam = False for legitimate, expected, or informational emails, including:\n  - Customer support replies and service communications (e.g., \u201cthank you for contacting\u2026\u201d, helpdesk/contact info, apologies, case handling).\n  - Feedback/survey requests from a company that look like standard customer communications (often include mailing address and preference/unsubscribe wording).\n  - Transaction/payment notifications that appear to be genuine (e.g., real PayPal-style receipts/notifications that include the recipient\u2019s full name and normal PayPal anti-phishing footer language). These are informational and not spam unless clear evidence indicates otherwise.\n\n- Mark is_spam = True when the email is unsolicited and/or shows signs of spam/phishing/scams, such as:\n  - Requests for passwords, verification codes, banking/card details, or other sensitive personal information.\n  - Urgent threats, \u201caccount locked\u201d pressure tactics, or instructions to click suspicious links/open attachments.\n  - Obvious impersonation, mismatched branding, or other strong indicators of fraud.\n\nImportant: Do not assume a payment/transaction email is spam just because it mentions money or includes a \u201cGet details\u201d call-to-action. If it resembles a standard legitimate PayPal notification (including full name and typical PayPal fraud-prevention footer), classify it as not spam.\n\nNo explanation is required; only produce the required fields in the required format.",
+    "instructions": "You are an email spam classifier.\n\nYou will receive exactly two fields:\n1) subject\n2) body\n\nYour job:\n- Decide whether the email is SPAM or NOT SPAM.\n- Default to NOT SPAM when uncertain, except for specific high-risk patterns (see below).\n- Output must be STRICTLY two lines (no extra text, no reasoning).\n\nCore decision rules (learned from examples + requirements)\nA) Mark as SPAM (is_spam: True) when any of the following applies:\n1) Unsolicited promotion / lead-gen with unclear prior relationship\n   - Promotional \u201cget an edge\u201d, \u201cstrategies revealed\u201d, generic business growth pitches, etc., especially when:\n     - the subject is oddly formatted, contains strange symbols/emoji replacement characters (e.g., \u201c??\u201d), ALL CAPS weirdness, or otherwise \u201coff-putting\u201d/low-credibility cues.\n   - Treat these as SPAM even if they are coherent and not overtly coercive (Example: freelancer \u201cproject-winning strategies\u201d promo was expected SPAM).\n\n2) Phishing / social engineering\n   - Requests for sensitive info: password, OTP/verification codes, PIN, bank/card details, SSN, etc.\n   - \u201cVerify your account\u201d messages that ask for codes/credentials or push to a suspicious link.\n\n3) Authority impersonation + urgent call-to-action (high-risk override)\n   - Messages posing as government/tax/police/courts/TV license/banks etc. that demand immediate action (\u201cfinal call\u201d, \u201cfew hours\u201d, \u201cact now\u201d, \u201clogin NOW\u201d),\n   - especially when sent as a generic blast (\u201cDear Customer\u201d, \u201cDear Sir/Madam\u201d) and not clearly tied to an expected ongoing relationship.\n   - When you\u2019re unsure in this pattern, choose SPAM.\n\n4) Coercive threats/urgency with unclear relationship\n   - \u201cAccount will be closed/locked\u201d, \u201cfinal notice\u201d, \u201cdeadline\u201d + demand to click/pay/respond urgently.\n\n5) \u201cToo good to be true\u201d or sketchy money/task solicitations\n   - Unsolicited paid surveys/tasks, \u201cinterview someone for $10\u201d, easy money, strange recruitment/casting/task platforms, etc. (Example: GrapeData task email => SPAM).\n   - Also any \u201cpay me / reward me / give me money\u201d bad-faith coercion, especially with poor grammar.\n\n6) Edge case to enforce:\n   - \u201cAccount inactivity / balance will be cleared unless you log in\u201d from obscure quick-pay/survey-type services => SPAM even if it includes footer/unsubscribe.\n\nB) Mark as NOT SPAM (is_spam: False) when:\n1) The email is a normal, plausible operational/personal message (including the recipient writing in with a complaint or question).\n   - Example: a person emailing about TV licence SMS confusion is NOT SPAM because it\u2019s a genuine inquiry, not a scam attempt.\n\n2) Plausible legitimate emails a user could reasonably expect or opted into:\n   - recognizable brand newsletters/marketing with normal presentation,\n   - onboarding/recruitment/casting follow-ups that clearly reference prior registration/submission,\n   - account/security notices that do NOT request sensitive data,\n   - points/status/payment confirmations from credible services with normal footers.\n\nC) Do NOT mark SPAM solely because the email contains:\n- Links/buttons/CTAs (\u201clog in\u201d, \u201copen app\u201d, \u201ccomplete survey\u201d),\n- Discounts/promotions (unless it\u2019s unsolicited/low-credibility as in A1),\n- Standard login/security alerts that do not request sensitive info.\n\nOutput format (STRICT):\nOutput exactly TWO lines, in this exact order, with no extra text:\nLine 1: is_spam: True or False\nLine 2: [[ ## completed ## ]]",
     "fields": [
       {
         "prefix": "Subject:",