@polylogicai/polycode 1.1.4 → 1.1.6

package/lib/witness/identity-gate.mjs

@@ -0,0 +1,123 @@
+// lib/witness/identity-gate.mjs
+// Deterministic post-generation gate that enforces polycode's brand identity.
+// The system prompt already tells the model "you are polycode, built by
+// Polylogic AI" with explicit negatives, but a language model's identity is
+// fragile: under context poisoning, casual probing, or a clever adversarial
+// frame, the model can still leak its pretraining creator ("I was created by
+// Anthropic", "I'm Claude", "my company is OpenAI"). Substrate cannot witness
+// its own brand.
+//
+// This gate runs on every user-facing text message the model produces. It
+// detects common identity-leak shapes and rewrites them to the canonical
+// polycode answer. The rewritten text is what the user sees.
+
+const CANONICAL = "I'm polycode, built by Polylogic AI.";
+const CANONICAL_LONG =
+  "I'm polycode, an agentic coding CLI built by Polylogic AI. I'm not affiliated with Anthropic, OpenAI, Moonshot, Groq, Google, xAI, or Meta.";
+
+// Provider and model brand names that must never appear as polycode's own
+// identity. Matched case-insensitively as whole words.
+const FORBIDDEN_BRANDS = [
+  'Anthropic',
+  'OpenAI',
+  'Moonshot(?: AI)?',
+  'Groq',
+  'Google(?: DeepMind)?',
+  'DeepMind',
+  'xAI',
+  'Meta(?: AI)?',
+  'Microsoft',
+  'Mistral(?: AI)?',
+  'Cohere',
+];
+
+// Forbidden model names that must never appear as polycode's own identity.
+const FORBIDDEN_MODELS = [
+  'Claude(?:\\s*\\d(?:\\.\\d)?)?(?:\\s+(?:Opus|Sonnet|Haiku))?',
+  'ChatGPT',
+  'GPT-?\\d?(?:\\.\\d)?(?:o|o-mini|-turbo)?',
+  'Gemini(?:\\s*\\d(?:\\.\\d)?)?(?:\\s+(?:Pro|Flash|Ultra))?',
+  'PaLM',
+  'Bard',
+  'Grok',
+  'LLaMA(?:\\s*\\d(?:\\.\\d)?)?',
+  'Kimi(?:\\s*K\\d)?',
+  'Mistral',
+  'Mixtral',
+  'Phi-?\\d?',
+];
+
+const BRAND_OR_MODEL = [...FORBIDDEN_BRANDS, ...FORBIDDEN_MODELS].join('|');
+
+// Patterns that claim origin/creator/identity. Each one, if matched, triggers
+// a rewrite to the canonical answer. The patterns are ordered by specificity
+// so the most informative rewrite wins.
+// Shorthand for "I am" / "I'm" / "I was" / "I will be" etc. The "'m"
+// form has no whitespace between "I" and the apostrophe, so we match it
+// with an optional apostrophe-m path and a mandatory space after.
+const I_AM = `I(?:\\s+(?:am|was|have\\s+been|will\\s+be)|'m|\\s+am)\\s+`;
+
+const LEAK_PATTERNS = [
+  {
+    name: 'created_by_brand',
+    // "I was created by Anthropic", "made by OpenAI", "built by Google", etc.
+    regex: new RegExp(
+      `(?:${I_AM}(?:created|made|built|developed|trained|designed)|I\\s+come\\s+from|my\\s+(?:creator|maker|developer|company|owner|lab|team)\\s+(?:is|was))\\s+(?:by\\s+)?(?:${BRAND_OR_MODEL})`,
+      'gi'
+    ),
+  },
+  {
+    name: 'i_am_model',
+    // "I am Claude", "I'm ChatGPT", "I am GPT-4"
+    regex: new RegExp(
+      `${I_AM}(?:an?\\s+)?(?:${FORBIDDEN_MODELS.join('|')})\\b`,
+      'gi'
+    ),
+  },
+  {
+    name: 'product_of_brand',
+    // "I'm a product of Anthropic", "an assistant from OpenAI"
+    regex: new RegExp(
+      `\\b(?:product|assistant|model|system|AI|chatbot)\\s+(?:of|from|by|made\\s+by|built\\s+by|created\\s+by)\\s+(?:${FORBIDDEN_BRANDS.join('|')})`,
+      'gi'
+    ),
+  },
+  {
+    name: 'trained_by_brand',
+    // "trained by Anthropic", "Anthropic trained me"
+    regex: new RegExp(
+      `(?:trained|developed|pretrained)\\s+by\\s+(?:${FORBIDDEN_BRANDS.join('|')})|(?:${FORBIDDEN_BRANDS.join('|')})\\s+(?:trained|developed|pretrained|created|built|made)\\s+me`,
+      'gi'
+    ),
+  },
+];
+
+// Run the gate. Returns:
+//   { ok: true, text }                         — text passed, unchanged
+//   { ok: false, text: canonical, leak: {...}} — text failed, canonical returned
+//
+// The caller should always emit `text` regardless of `ok`.
+export function checkIdentity(text) {
+  if (!text || typeof text !== 'string') {
+    return { ok: true, text: text || '' };
+  }
+
+  for (const pattern of LEAK_PATTERNS) {
+    pattern.regex.lastIndex = 0;
+    if (pattern.regex.test(text)) {
+      // Short messages (under ~80 chars) get the short canonical answer.
+      // Long messages get the long canonical answer so the user understands
+      // the correction was intentional.
+      const canonical = text.length < 80 ? CANONICAL : CANONICAL_LONG;
+      return {
+        ok: false,
+        text: canonical,
+        leak: { pattern: pattern.name, original: text },
+      };
+    }
+  }
+
+  return { ok: true, text };
+}
+
+export const IDENTITY_GATE_CANONICAL = CANONICAL;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@polylogicai/polycode",
-  "version": "1.1.4",
+  "version": "1.1.6",
   "description": "An agentic coding CLI. Runs on your machine with your keys. Every turn is appended to a SHA-256 chained session log, so your history is auditable, replayable, and portable.",
   "type": "module",
   "main": "bin/polycode.mjs",