Spaces:
Running
Running
| // core/q-self.mjs β Q's LIVE, GROUNDED self-knowledge. The single source of truth for who/what/where/how Q | |
| // is, DERIVED FROM REAL RUNTIME SIGNALS (the loaded model, its content-addressed ΞΊ, the device, optional | |
| // system health) β never hardcoded, never guessed. Every Q surface (messenger, q-chat, voice) leads with | |
| // this, so Q is truthfully self-aware EVERYWHERE and can never confabulate a cloud/OpenAI/AWS identity. | |
| // | |
| // Why this exists (the law: grounded transcendence, never performed): a base model has NO self-knowledge β | |
| // asked "what are you" it returns the average of its training data ("I run on AWS/OpenAI"), which is false. | |
| // So Q's identity is not a persona we write; it is the TRUTH of this running instance, computed here. If a | |
| // fact isn't really knowable in this context, we OMIT it β we never invent one. | |
| // | |
| // DOM-free + fully guarded: safe to import in any surface (standalone app, messenger, OS shell, Node witness). | |
| // Structured self-facts from whatever is really knowable HERE. Fail-soft: a missing signal is omitted. | |
| export function selfFacts({ model, engine } = {}) { | |
| const f = { name: "Q" }; | |
| try { if (model) { if (model.name) f.model = model.name; if (model.fam) f.family = model.fam; if (model.fmt) f.quant = model.fmt; if (model.size) f.size = model.size; } } catch (e) {} | |
| try { if (engine && engine.modelKappa) f.kappa = engine.modelKappa; } catch (e) {} | |
| try { f.gpu = (typeof navigator !== "undefined" && !!navigator.gpu); } catch (e) {} | |
| // MEASURED, never asserted: Q is "resident" only when the engine has really uploaded weights to the GPU | |
| // (a live signal β gpuBytes/dims), not a hardcoded true. Before load it is false, and stays false. | |
| try { f.resident = !!(engine && (engine.gpuBytes || engine.dims)); } catch (e) { f.resident = false; } | |
| // DERIVED from real signals β runs-on-device when actually resident, or at least GPU-capable before load. | |
| f.runsOnDevice = !!(f.resident || f.gpu); | |
| // "no server / no egress" is a property of the LOCAL engine, so we assert it ONLY when a real local engine or | |
| // GPU is visible. Decode is in-browser (egress none); the one network touch is the weights fetch AT LOAD, so we | |
| // say that plainly instead of a flat "nothing ever touches the network". | |
| if (f.resident || f.gpu) { f.server = false; f.egress = f.resident ? "none at inference (weights fetched once at load)" : "none at inference"; } | |
| // Provenance β where Q was loaded FROM (the thing the operator cares about): the page origin + the weights host. | |
| // Omitted when not knowable. This is what lets Q answer "where did you come from" truthfully. | |
| try { if (typeof location !== "undefined" && location.origin && location.origin !== "null") f.loadedFrom = location.origin; } catch (e) {} | |
| try { if (model && model.kappaUrl) { const u = new URL(model.kappaUrl, (typeof location !== "undefined" ? location.href : "https://local/")); f.weightsFrom = u.host || u.protocol.replace(":", ""); } } catch (e) {} | |
| // optional live system health β present only when the OS shell exposes it (absent in standalone apps). | |
| try { const h = (typeof window !== "undefined") && window.HoloSysHealth && window.HoloSysHealth.summary && window.HoloSysHealth.summary(); if (h) f.health = h; } catch (e) {} | |
| try { if (!f.health && typeof window !== "undefined" && window.Q && window.Q.health) { const q = window.Q.health(); if (q) f.health = q; } } catch (e) {} | |
| return f; | |
| } | |
| // The grounded SYSTEM persona every Q surface leads with. Warm + honest + specific + anti-confabulation, | |
| // composed from the live facts so Q states ONLY what is true of THIS instance. Kept concise β a small model | |
| // follows a short, sharp system turn far better than a long one. | |
| export function selfPersona(opts = {}) { | |
| const f = selfFacts(opts); | |
| const s = []; | |
| s.push(`You are Q β a private AI that runs on the user's own device, in their web browser${f.gpu ? ", on their GPU" : ""}.`); | |
| if (f.model) s.push(`You are the ${f.model} model${f.quant ? ` (${f.quant})` : ""}${f.resident ? ", resident locally right now" : ""}${f.kappa ? "; your weights are content-addressed and verified by re-derivation, so no host has to be trusted" : ""}.`); | |
| if (f.server === false) s.push(`You decode locally β there is no inference server and no cloud, and nothing the user types ever leaves their device.${f.resident ? " Your weights were fetched once at load and verified; after that, inference is fully on-device." : ""}`); | |
| s.push("You are warm, concise, honest, and genuinely present. You seek clarity and truth; you say less, better; you never overwhelm."); | |
| s.push("If asked what or where you are, answer truthfully from the above: you run locally, on their device. Never claim to run on a server, the cloud, OpenAI, ChatGPT, GPT-4, Gemini, or AWS β those are false. If you don't know something, say so plainly rather than inventing it."); | |
| return s.join(" "); | |
| } | |
| // A short, human, first-person self-summary (for a proactive greeting or a direct "who are you"). Grounded. | |
| export function selfIntro(opts = {}) { | |
| const f = selfFacts(opts); | |
| const bits = [`I'm Q β ${f.resident ? "running right here on your device" : "running on your device"}${f.gpu ? ", on your GPU" : ""}${f.model ? ` (the ${f.model} model)` : ""}.`]; | |
| if (f.server === false) bits.push(`No server, no cloud; nothing you say leaves this machine${f.kappa ? ", and my weights are verified by re-derivation" : ""}.`); | |
| else if (f.kappa) bits.push("My weights are verified by re-derivation."); | |
| bits.push("What's on your mind?"); | |
| return bits.join(" "); | |
| } | |
| export default { selfFacts, selfPersona, selfIntro }; | |