📕 Vademecum MFF

👋 Welcome

MFF (MarcoFLY Framework) is an epistemic control layer on top of generative AI: it makes visible how confident a model is about what it says, statement by statement, and helps you verify it. This app applies the framework automatically to every message, across dozens of models from different providers, with YOUR API keys (BYOK).

The two modes: “MFF Framework” injects the epistemic protocol (labels, shields, state card) and consumes MFF credits only when the framework is actually applied. “Simple chat” (PLAIN) is free conversation without the framework and without credits — but with ALL platform features: attachments, web search, peer review, read-aloud, export.

What MFF is NOT

• It does not guarantee truth: labels reflect what the model declares to know; the final check stays with you.
• It does not replace experts: medical, legal, tax or financial decisions need a professional.
• It does not zero out hallucinations: it reduces them and makes them recognisable, but the model can be wrong even with a green badge.

🚀 Quick start — 3 steps

1. 1Set up an API key (or use the included MFF credits). Go to Settings → AI provider keys: paste a key — the fastest path is “Connect OpenRouter” via OAuth, one key for hundreds of models, free ones included.
2. 2Create a session from the Dashboard: choose mode (MFF or PLAIN), domain, model — and the defense modules if you want. Every choice is explained in the sections of this vademecum.
3. 3Write your question: the answer streams in, with labels. Use 📎 to attach images and documents, 🌐 to request a web search before the answer. Every message has 📋 copy, 🔊 read-aloud, ⚖ peer review and regeneration.

🔓 Free keys, provider by provider

Each chapter below is linkable (🔗 icon): from the PWA Settings you land straight on the provider you are configuring. Reading key: 🆓 start without a card · 🌱 minimal entry cost · 💳 credit required — declared, never hidden.

🏷️ The epistemic labels

Every statement carries its declared commitment level. They are not decoration: they are a contract — and they persist in exports. The golden rule: the tag PRECEDES the content, one tag per block.

Suffixes: where it comes from and how confident

After the label, source suffixes may appear: [doc] official document · [test] testimony · [comm] official statement · [log] technical log · [emp] empirical data · [lit] scientific literature · [leg] legal source · [cert] verifiable certification. Then [mem] = training memory not verified in session (always to be checked) and [p:75%] = the model’s estimated subjective probability. The glossary at the bottom explains them all.

🛡️ The shields — defense modules

In MFF mode you can enable up to 9 epistemic modules that steer the model’s behaviour. L1 is always on; the default set L1+L2+L3 is great for most tasks. The cards below are the SAME ones you see in the session creation wizard.

💬 Sessions: every choice and what it implies

A session is a complete conversation with one model: it has its own title, model, mode, domain, language and shields, set at creation. Every wizard choice changes something concrete in the engine.

The domain does two things

First: it defines what 🟢/🔵/🟡 mean in that context — in the scientific domain “CERTAIN” means replicated, published results; in the creative domain the bar differs. Second: it sets the default temperature, from 0.2 (Legal: conservative, repeatable) to 1.0 (Creative: maximum variability). Advanced settings can override it.

The operating mode

MFF-E: rigorous epistemic analysis · MFF-G: cuts noise and redundancy · MFF-X: interprets and explains without altering · MFF-EX: combines them (the central mode) · MFF-EGX: adds the discipline of synthesis · AUTO: picks per turn. PLAIN: no framework, no credits.

The session language is a choice, and it stays

It is separate from the interface language. Switching the UI language does NOT translate answers already generated: the epistemic label belongs to the text that model produced — translating under the labels would falsify them.

Inside the conversation

• Each new message carries the session context: follow up without repeating. For new topics open a new session (clean context, lower costs).
• Under every answer: 📋 copy · 🌐 sources · 🔊 read-aloud · ⚖ peer review · 🔄 regenerate. Your own messages can be edited and resent.
• “▸ Continue” resumes an answer truncated by the budget; in sessions beyond 100 messages “⤒ Load earlier messages” walks back through history.
• All sessions live in the Archive: reopen and continue where you left off. From there you can also delete them, one by one.

🧠 Choosing the model

The picker always shows the provider’s full lineup, marking which models are actually served right now: the truth comes from the live catalog, not a hand-written list. An unserved model is visible and marked, not hidden.

Capabilities come from the catalog too: vision → image attachments enabled; reasoning → extended thinking with L5; image generation → the model joins the generation routes. The model name decides nothing: what the provider declares does.

Free versus paid

:free models cost nothing, with three trade-offs: tighter context and document budget, no 🌐 web search, and provider daily quotas (a 429 on a :free often just means “retry later or switch :free”). Paid models run on YOUR key: you pay the provider’s list price, MFF adds nothing.

If the provider fails: declared failover

On recoverable errors (rate limit, overload, retired model) the engine tries the reserves: first the fallbacks you set in advanced settings (up to 3, in order), then the policy chain across your keyed providers. Everything is DECLARED: answer, labels and card report who actually served. The turn restarts from scratch on the reserve model — what you read is always from a single model. Non-recoverable errors (invalid key, exhausted balance) do not fail over: they are shown, and the MFF credit comes back.

🔌 Providers and routing

MFF talks to 17 providers, in two families. YOU choose the session provider at creation — there is no hidden routing: what you choose is what answers, and any fallback is declared (see failover).

Aggregators: one key, many vendors

• OpenRouter (sk-or-… key): hundreds of models, :free included; also connects in one click via OAuth. On paid models it uses its native web_search. Small markup (~5%) versus direct.
• OrcaRouter (sk-orca-… key): zero-markup aggregator with its own independent key — it does not use your OpenRouter credit.

Direct: the shortest path and native capabilities

OpenAI · Anthropic · Google · Groq · Cerebras · Mistral · DeepSeek · xAI · Perplexity · Z.AI · NVIDIA NIM · SambaNova · Hyperbolic · Cohere · Cloudflare Workers AI. No markup, and the vendor’s native features: Anthropic’s built-in web search, Perplexity’s native citations (shown among sources), the generous free tiers of Groq, Cerebras, NIM and SambaNova.

Image generation is a platform capability, not a chat-model one: if the session model cannot generate images, the engine routes to the first of your keyed providers that can, and declares it.

🔑 Configuring API keys

1. 1Go to Settings → AI provider keys and open the provider card.
2. 2Paste the key: MFF validates it IMMEDIATELY with a real test query — a key that fails the test is not saved.
3. 3The key is encrypted server-side (AES key-ring with rotation) and is never shown in clear again, never put in a URL, never sent to the browser. The model NEVER sees it: requests leave from the server.
4. 4Each key shows a traffic light (active · invalid · paused) and, in Expert view, the 🎯 allowed-models whitelist: session and failover will never leave the list.

Where to create the key, provider by provider

SEARCH keys (for the 🌐)

Separate from model keys: Serper, Tavily, SerpAPI, Brave — in Settings, Web search section. These too are validated with a test query. Elect the default with the ⭐: the engine tries it first, the others remain reserves in the fixed order Serper → Tavily → SerpAPI → Brave. With your own key you only pay your provider; without one, you use MFF’s included quota (limited and shared).

🌐 Web search and sources

1. 1Press 🌐 next to the composer: a mini-model extracts the proposed search query and SHOWS it to you. You can fix or rewrite it — the query is your decision, not a blind automatism.
2. 2The search runs on the MWAL chain: your default key first, then the reserves (Serper → Tavily → SerpAPI → Brave). If an engine fails, the hand-off is declared.
3. 3Results enter the context as a grounding block: the answer can cite them, and sources stay in the message’s 🌐 chip — they survive reloads and end up in exports.

• Honest limit #1: search snippets are by construction a cache — for minute-by-minute data (quotes, the exact time) the source can be pertinent yet stale.
• Honest limit #2: on :free models the 🌐 is disabled for cost containment.

⚖️ Peer review between models

A SECOND model, from a different provider, re-reads the answer and shows you its judgement next to the original. It is protocol level L5-B: convergence of two independent agents can promote a 🔵 PROBABLE to 🟢 CERTAIN; divergence forces 🟡 or 🔴. Two models sharing neither weights nor channel, agreeing on the same statement, are worth more than one.

1. 1At session creation, in Advanced settings, choose the peer model: the list draws from your keyed DIRECT providers, deliberately different from the session provider.
2. 2In chat press ⚖ under the answer you want re-read: the review streams in and stays anchored to that message, with its convergence verdict.
3. 3Alternatively, the auto-trigger (checkbox in Advanced) runs the review on EVERY answer: a mode for critical sessions — it doubles the peer provider’s consumption.

Reviews are persisted: you find them on reload and in exports; those anchored to messages older than the loaded page remain reachable at the top of the transcript. If the ⚖ appears dimmed, the session has no peer model: it is chosen at creation. It works in PLAIN too — and in PLAIN it consumes no MFF credits: you only pay your peer provider.

Epistemic implication: if you regenerate an answer, reviews stay anchored to the message they referred to — a review never migrates onto a text it never read.

🗂️ State card and transferable memory

The session collects in a 🗂 card what has been established and on what grounds, extracted from the labels already declared — not a summary written by a model. Every statement carries its declarant: after a failover you read WHO actually declared what, never a convenient attribution.

It serves two purposes: reading it, and taking it elsewhere. “Copy as memory” produces a text to attach to a new session — perhaps with another model — without dragging the whole transcript. The recipient reads that those labels were declared by someone else: statements to discuss, not verdicts to inherit. This is also the correct way to “switch model”: not mid-session, but with the card travelling.

🪙 MFF credits

Welcome bonus

Every new account receives 500 credits progressively: 50 at signup, +450 at the first answer where the framework is actually applied. The progression is minimal anti-fraud; past your first MFF message, you have the full 500. In the free beta credits cannot be purchased; balance and every movement are in Settings → Credits.

💰 Optimising costs

Models charge per token (~4 characters in English), both input and output. Every answer in MFF shows its numbers in the technical footer: tokens, timings and — when the price list is known — the message cost estimate and the session running total. If the provider’s usage never arrives, the estimate is marked with ~, never passed off as data.

1. 1Use :free models for drafts and exploration (trade-offs: reduced context, no 🌐).
2. 2Short sessions + a transferred card beat mile-long conversations: history beyond the cap gets truncated by the engine anyway — the card is the memory that travels without paying context.
3. 3Keep reasoning on Low until the task demands more: thinking tokens are billed like any others.
4. 4🎯 whitelist on your key: no automatism can pick an expensive model on your behalf.
5. 5Cheap fallbacks in Advanced: failover respects your list.
6. 6PLAIN for casual use: zero credits, same features.

⚙️ Settings and diagnostics

The Settings page has two views, selectable from the menu at the top and synced across your devices: “Basic user” shows the essentials (profile, keys, balance, voice, privacy); “Expert user” adds diagnostics, technical AI preferences, 🎯 whitelist and movement history. If this vademecum mentions something you can’t see, it almost certainly lives in the Expert view.

Diagnostics (Expert view): no invented numbers

• BYOK probe: a REAL test query on each of your keys, with outcome and latency.
• Images panel: declares WHICH of your keys would generate an image right now, and with which model.
• Model health: telemetry of what you actually used (time to first token, errors, 429s).
• Field census: what provider APIs expose and MFF does not translate yet. Missing data is shown as missing, never passed off as zero.

What to do when something is red

• “Invalid” key: revoked or expired on the provider’s site — regenerate it there and paste it again. MFF cannot repair it.
• “Paused” key: the breaker stopped it after repeated errors (three “invalid” in a row = a quarter hour) to stop burning requests. It reactivates by itself, or immediately via the “Reactivate” button.
• Failed probe on a key that worked yesterday: look at the code before deleting — a rate limit or a provider outage passes on its own.

📱 Installing MFF (PWA)

MFF is a Progressive Web App: it works in the browser, but installed it gives you a dedicated icon, full screen without the URL bar and instant start. AI features stay identical (connection required: there is no offline inference).

Android (Chrome / Edge / Brave)

1. 1Open the app in the browser: the “Install app” banner appears → Install. If it doesn’t: ⋮ menu → “Install app” / “Add to Home screen”.

iOS (Safari)

1. 1Open the app in Safari (on iOS PWAs require Safari) → Share icon → “Add to Home Screen” → Add.

Desktop (Chrome / Edge / Brave)

1. 1On the right of the address bar the install icon appears (monitor with arrow) → Install: the app opens in its own window.

🔧 Troubleshooting

• 429 error on a :free model → it is the provider’s daily quota, not a failure: wait a few minutes, try another :free, or use a paid model with your key.
• Answer interrupted midway → “▸ Continue” under the message resumes where it stopped. If the provider went down, declared failover already tried the reserves; any error stays visible and the credit is refunded.
• Slow answer → extended-reasoning models are slow by construction: for simple tasks use a fast model (Groq, Cerebras) or lower the reasoning effort in Advanced.
• Red or paused key → see “Settings and diagnostics”: regenerate at the provider if invalid; wait or “Reactivate” if paused.
• Session won’t create → the message under the button says why (model unavailable, credits exhausted in MFF mode, session expired). Reload and retry; PLAIN needs no credits.
• UI showing stale elements → refresh the PWA: hard reload (Ctrl+Shift+R / Cmd+Shift+R); if installed, close and reopen the app.

❓ FAQ

📖 Glossary

Every technical term of the framework, explained in plain language. Merged from the public guide’s glossary (74 entries in 7 categories).