Guide

How Unisona works

Written for someone brand new — to Unisona, to AI assistants, and to trading. No prior knowledge assumed. Every term is explained the first time it appears.

What Unisona is

Unisona is an AI assistant that remembers your work, and a stock-market terminal, in one place. You talk to it in plain English; it can look things up, read and write documents, and — if you want — watch the market with you.

Two things make it unusual, and both are worth understanding before you start:

  • It keeps its memory on your side. Your chats, notes and watchlist are stored as plain files rather than locked inside someone else's product. You can run the whole thing on your own computer if you'd rather.
  • It isn't tied to one AI model. Unisona is a system that uses AI models, not a wrapper around a single one. Which model answers you is a setting you control, and swapping it doesn't cost you your history.
What it is not. Unisona does not predict the market, and nothing in it is financial advice. The trading side is a tool for watching, reasoning and recording — its own results are an open question we measure in public rather than a claim we make. Never trade money you can't afford to lose.

What it does

Four surfaces. You can use any one of them and ignore the rest.

💬

Chat

The main way you use Unisona. Ask questions, get help writing, have it search the web, read your files, or explain what the market is doing. It remembers earlier conversations, so you don't re-explain yourself every time.

How to use it →
📈

Trader

A live chart terminal. Watch prices, read charts, and — once you connect a brokerage account — place orders. It starts in paper mode, which means practice money.

How to use it →
👁

Watchlist

The list of symbols you care about, down the left side of the trader. Everything else — charts, alerts, what the assistant talks about — follows from this list.

How to use it →
🔔

Alerts

Tell Unisona to watch for something and it will message you when it happens — on the site and by email — so you don't have to sit and stare at a screen. Pro

How to use it →

Getting started

1. Make an account

Go to the sign-in page and either continue with Google or create an account with an email address and password. You'll get a confirmation email — click the link in it to prove the address is yours. That verification matters later: Unisona will only email you alerts at an address you've confirmed.

2. Pick a plan (you can start free)

There are three, and the free one is a real product rather than a trial.

PlanPriceWhat you get
Free$0 Chat with memory, your own charts and watchlist, a paper (practice) trading account, and a view of a live demo account. If you connect your own brokerage account, you can trade through it on the free plan too.
Pro$20/mo Everything above, plus the full trading terminal, alerts and an automatic trade journal, portfolio analysis in chat, paper options orders, wide multi-model research, the Creator Suite, and the ability to plug in your own AI provider keys.
Pilot$200/mo Everything above, plus the autonomous AI trader (it acts on its own signals rather than only showing them to you) and a say in what gets built next.

Current plans and exact wording live on the pricing page — that page is the authority if anything here drifts.

3. Your first five minutes

  1. Open Chat and ask it something real — "explain what a stock index is", or "summarise this document" with a file attached. That's the whole interface; there's nothing to configure.
  2. Open Trader. You'll see a chart in the middle and a list of symbols on the left. Click any symbol to load its chart.
  3. Add something you care about: + Add symbol at the top of the left rail, type a ticker (say AAPL), pick it from the results.
  4. Go back to Chat and ask "what's on my watchlist, and does anything look weak today?" It already knows your list — you don't have to tell it.
You have not risked any money at this point, and you can't accidentally. Trading requires a brokerage account you connect on purpose — see Connecting a broker.

Using the chat

Type in the box at the bottom and press Enter. That's it. A few things worth knowing:

It can actually do things, not just talk

When it needs information it doesn't have, it uses a tool — a real action, like searching the web, reading a file you uploaded, fetching a live stock price, or checking your watchlist. You'll see a small marker in the reply (for example 🔧 trader_watchlist) telling you which tool ran. That marker is your receipt: it means the answer came from looking something up, not from memory.

It knows your market setup

If you ask anything about trading, the assistant already has your watchlist, current prices, and the trading engine's own read on each symbol. Ask "should I be worried about SMH?" and it will reason from the same numbers the terminal is showing you, and tell you when its figures were taken.

It will not place orders. Ask it to buy something and it will decline and point you at the trader. Deciding to trade is yours; the assistant's job is to help you think.

Getting better answers

  • Say what you want to end up with. "Explain X to someone who's never traded" beats "explain X".
  • Give it the material. Upload the file, paste the error, name the ticker.
  • Push back. If an answer is too vague or too technical, say so — it will adjust.
  • Ask it where something came from. A good answer can tell you.

Choosing which AI answers

The picker above the message box switches models. Different models have different strengths and speeds; your conversation carries across when you switch. On Pro you can also add your own provider keys — see API keys.

Using the trader

The layout

  • Left rail — your watchlist, and behind the second tab, your alerts.
  • Middle — the chart or charts. The dropdown in the toolbar chooses how many you see at once.
  • Right — the order ticket, where a trade is actually placed.
  • Bottom — your positions, orders, order history, and the journal.

Reading the chart

The default view is a candlestick chart. Each candle covers one slice of time (five minutes by default — change it with the timeframe dropdown). One candle tells you four things at once:

  • The thick body spans the open and close — where price started and ended in that slice.
  • The thin lines above and below (the wicks) reach the high and low.
  • Colour tells you direction: closed higher than it opened, or lower.
  • The bars along the bottom are volume — how many shares changed hands. Big volume means a lot of people acted; it makes a move more meaningful.

Price runs up the right edge, time along the bottom. Hover anywhere to read the exact values under your cursor.

The watchlist

+ Add symbol to add, the on a row to remove, and drag rows to reorder. Click a symbol to make it the active one — the chart, the details panel underneath, and the order ticket all follow. Your list is saved to your account, so it's there on your phone too.

Alerts Pro

Switch the left rail to Alerts. Press + New alert and choose:

  • A signal fires — the engine's own read on that symbol turns bullish or bearish.
  • Price reaches a zone — price comes within a set percentage of a support or resistance level (see the glossary).
  • Washout confirmed — the engine's stricter "this setup is real" verdict.

Stay quiet for is how long an alert waits before it's allowed to fire again, so one jumpy morning doesn't send you thirty messages. Leave Email me when it fires ticked and you'll get an email as well as an on-screen notice; untick it for on-screen only. You can pause an alert with its checkbox or delete it with the ✕.

Paper versus live — the important distinction

PaperLive
MoneySimulated. None of it is real.Real. Losses are real.
NeedsNothing — it's on by default.Your own brokerage account, connected on purpose.
ForLearning the terminal, testing an idea, seeing how a strategy behaves.When you've done the above and decided.

The badge in the toolbar always tells you which mode you're in. Check it before you press anything. Paper trading is genuinely useful — the mechanics are identical, so you can learn every button without a single dollar at risk.

What the autopilot does and doesn't decide

Unisona has a scanning engine that reviews your watchlist about once a minute and forms a view on each symbol: a direction, a confidence, the levels it thinks matter, and a suggested stop and target. On Free and Pro, that is all it does — it shows you what it sees, and every order is placed by you.

The autonomous trader, which acts on those signals by itself, is a Pilot-plan feature and is off unless you deliberately turn it on. Even then it works only through a broker account you connected yourself.

An engine having high confidence is not the same as being right. Confidence describes how well the current picture matches a pattern it recognises — not the odds of making money. Treat it as one more input.

Connecting a broker

A broker (or brokerage) is the licensed company that actually holds your money and executes your trades — Unisona never does. To place a real order you connect an account you already have. Two are supported.

Before you start

  • You need an account with the broker already open. That's their signup, not ours.
  • Start with a paper account at the broker. Both options below offer one free.
  • Everything happens in Settings → Connections.

Alpaca (simplest — paper only)

Alpaca is an online broker with a good practice mode. Unisona connects to Alpaca paper accounts only, which makes it the safest place to start.

  1. Sign up at alpaca.markets and sign in.
  2. Switch to Paper Trading (there's a toggle — make sure you're not in the live account).
  3. Find API Keys and generate a new key. You'll be shown two strings: a key ID and a secret key.
  4. Copy the secret immediately. Brokers show it once. If you lose it, delete the key and make a new one — that's normal and costs nothing.
  5. Paste both into Settings → Connections → Alpaca, and save.

Interactive Brokers (IBKR — paper or live)

IBKR is a large traditional broker. Unisona supports both its paper and live accounts, which means it's also the one where a mistake can cost real money. Use its paper account until you're comfortable.

  1. In Settings → Connections, choose IBKR and follow the steps shown there — the exact flow depends on how your IBKR account is set up, so the page walks you through the current one rather than us printing a version here that goes stale.
  2. Choose the paper account when asked, unless you have specifically decided otherwise.
  3. Confirm the account number shown in Unisona matches the one you intended. A paper account number looks different from a live one — check it.

Disconnecting

The same Settings page removes stored keys. Do that if you stop using Unisona, if you think a key leaked, or any time you're unsure — reconnecting takes two minutes.

API keys, explained

If you've never used one: an API key is a long random string that acts as a password for programs instead of people. When you give Unisona a key for another service, you're saying "this app may act as me over there".

Where they show up here

  • Broker keys — let Unisona read your positions and place the orders you ask for.
  • AI provider keys Pro — let you use your own account with an AI company, billed to you directly, instead of Unisona's.

Neither is required to start. You can use the chat and the paper trader without ever entering a key.

Treat them like passwords, because they are

  • Don't paste a key into a chat message, a screenshot, a support email, or a public issue — including into Unisona's own chat box.
  • Use a paper/read-only key when the service offers one.
  • If a key might have leaked, revoke it at the provider. Revoking is instant and free; recovering a drained account is neither.
  • Rotate them occasionally — delete the old, create a new, paste it in.

What Unisona does with them

Keys are stored encrypted, scoped to your account, and used only to make the requests you asked for. They're never shown back to you in full, and they're never sent to an AI model. Deleting a connection deletes the stored key.

Trading words, in plain English

Ticker (symbol)
The short code for something you can trade. AAPL is Apple.
Share
One unit of ownership in a company.
ETF
A single ticker that holds a basket of many things. SPY tracks the 500 largest US companies, so buying it spreads your money across all of them.
Position
Something you currently own. "A position in AAPL" means you hold some Apple.
Order
An instruction to buy or sell. A market order says "now, at whatever the price is". A limit order says "only at this price or better" — it may never fill.
Fill
An order actually being executed.
P&L
Profit and loss. Unrealised is on paper while you still hold it; realised is locked in once you've sold.
Day P&L
How much you're up or down today specifically.
Stop (stop-loss)
A price at which you've decided to sell to cap a loss. Deciding it before you buy is the single most useful habit in this list.
Target
The price at which you plan to take a profit.
Support
A price level where buyers have repeatedly stepped in — a floor price has bounced off before.
Resistance
The opposite: a level where selling has repeatedly appeared, a ceiling.
Volume
How much was traded. High volume means a move had real participation.
Volatility
How much price jumps around. High volatility means bigger moves in both directions.
VIX
A widely-watched measure of expected volatility in the US market. Low is calm, high is fearful.
Extended hours
Trading before the market opens or after it closes. Fewer people are trading, so prices move more erratically.
Paper trading
Practice with simulated money and real prices.
Long / short
Long is owning something, hoping it rises. Short is betting it falls — riskier, because a rise has no ceiling.
Drawdown
How far an account has fallen from its highest point. The number that tells you what a strategy actually felt like to live through.

If something goes wrong

The chart is empty or the prices look stale.
Markets close. Outside trading hours you're seeing the last prices from the previous session, which is expected — the toolbar shows whether the session is open. If it's mid-session and still stale, reload the page.
I signed in but the trader sends me to the Watch page.
The full terminal is a Pro feature. Watch is the free version — real charts, watchlist and news. Pricing has the details.
My alert didn't email me.
Check, in order: is the alert enabled (the checkbox)?; is Email me when it fires ticked on that alert?; is your email address confirmed?; and has its quiet window elapsed? Alerts only go to verified addresses. Also check your spam folder once and mark it "not spam" so later ones arrive.
I connected my broker but positions are empty.
Usually one of three things: the connection is to a paper account with nothing in it, the keys were pasted with a stray space, or the broker takes a few minutes to activate a brand-new key. Re-check the account number shown in Settings against the one you meant.
The assistant won't place a trade for me.
That's deliberate and can't be turned off. It can research, explain and analyse; placing orders is done by you in the trader.
An answer looks wrong.
Say so in the chat and ask where the number came from. If a tool marker (🔧) is shown, the figure came from a live lookup; if not, treat it more carefully. You can also report it with the camera button in the header, which attaches a screenshot.
Something else.
The FAQ covers accounts, privacy, pricing and self-hosting. The reference library below has the technical documentation.
Library

Reference library

Past the guide above: the full technical documentation — product guides, architecture decisions, and research, newest first. You don't need any of this to use Unisona. Type to filter live, pick a category, or switch to For builders for the internals. Press Enter to ask Unisona's grounding search for an answer instead of a document.

Filtering searches every document below. No match? Press Ask to search Unisona's guides, or just ask in Unisona Chat.
📘MD Agents & Contributor Guide Required-reading operating guide: the per-lane one-PR monoworkstream rules, git workflow, route map, and how the convergence agent fleet is coordinated. GuidesLivingUngroundedJun 2026 🔌MD unisona.ai — AI Provider Chain Documentation Provider inventory, keys, PCSF-backed routing and decode/serving params; provider status table stale (marks Anthropic/OpenAI/xAI active when only Gemini is funded). Setup & integrationsReferencePartialJun 2026 📘MD unisona.ai Skills & Capabilities Reference for live skills, agent personas, and provider chain; persona table and provider-status column are stale. GuidesReferencePartialJun 2026 📘MD Security Policy Required-reading security policy: secret handling, local-first verification, fixed vulnerabilities, and agent-safety rules with honest enforced/policy status flags. GuidesLivingGroundedJun 2026 🏛️MD unisona.ai — Architecture (Current State) Canonical current-state architecture snapshot with file:line evidence and an honest known-divergences section. ArchitectureLivingGroundedJun 2026 🏛️MD Architecture Decision Records (ADRs) Index and authoring rules for the append-only ADR log; approval gate and status values. Each ADR listed exactly once (0008 collision reconciled 2026-07-01). ADR · IndexLivingGroundedJul 2026 🧪MD Living Research Canon — Convergence 12 Curated per-component reference trail for the Convergence 12; uses dead 'unisona.ai' branding and some 'beat-zstd' framing is aspirational. ResearchReferencePartialJun 2026 🧪MD BENCHMARKS — External Marks Registry Living registry of external benchmarks unisona.ai runs: HumanEval/MBPP Run, SWE-bench/LongMemEval Partial, ARC-AGI/HLE/OSWorld Planned, with evidence ledgers. ResearchLivingGroundedJun 2026 🏛️MD The Convergence Loop — North Star briefing The immutable design contract every part of unisona.ai is built against: one loop (Observe, Remember, Reason, Act, Verify, Converge) and four core objects. Explains why the product refuses architectural sprawl. ArchitectureReferenceUngroundedJun 2026 📘MD How unisona.ai Remembers — The CSF Memory & Retrieval System Remember-stage guide: append-only JSONL + CSF memory, tiers, IDF-weighted retrieval, with a LongMemEval recall@5 benchmark table. GuidesLivingGroundedJun 2026 🧠MD Σ₀ Ouro Coder — the local coding agent Single source of truth for the local Ouro-1.4B + Σ₀ QLoRA coder: adaptive Q-exit depth, serving, honest scope, training status. Models & trainingLivingGroundedJun 2026 📈MD Σ₀ EV Gate — the trader's Converge stage The live expected-value gate that decides ENTER/SKIP on trades; Grok/Claude are graded council members, not gates. Backed by convergence_ev.py + tests. TradingLivingGroundedJun 2026 📋MD Unisona 1.8 — the personal reasoning cockpit goes public Release note for unisona.ai 1.8: per-loop-stage cockpits (Explore, memory decay, cockpits, autowork steering, verify surfaces) consolidating the 1.7.1xx line. ReportsLivingPartialJun 2026 📘MD unisona.ai Chat — Product Definition & Operating Contract Product contract for the member-facing operator console: grounded, tool-wired, fast chat; includes the fast/deep Ouro serving decision and measured 3070 latency baseline. Some 'unisona.ai' phrasing. GuidesLivingPartialJun 2026 🏛️MD Convergence IO — the Typed Primitive Stack Index of the P1-P10 governance primitives (DCF, NAP, AAPF, PCSF, CCF, CEG, dilation); how they compose and an honest note that not every primitive is on the live JS hot path. ArchitectureReferencePartialJun 2026 📘MD Act — do it with real tools The fourth stage: real tool calls (web, documents, workspace files, market data, repo/GitHub) with honesty-about-actions — Unisona only claims what actually ran. GuidesLivingGroundedJul 2026 📘MD Contributing to unisona.ai Contributor guide: fork/branch workflow, monoworkstream lane rule, ES6/2-space code style, WCAG-AA and i18n checklists, pytest + PR process. GuidesLivingUngroundedJun 2026 📘MD Converge — record the outcome and improve The sixth stage: convergence records (hypothesis + evidence + result + confidence) that turn one interaction into lasting capability, improving via experience not retraining, then closing the loop. GuidesLivingGroundedJul 2026 📘MD Explore Feed — single-pane, PCSF-ranked content As-built spec for the Explore page: one PCSF-ranked content stream reusing the provider leaderboard, with data flow, API contract, and known gaps. GuidesLivingGroundedJun 2026 📘MD Observe — take in the world The first stage of the convergence loop: how Unisona takes in the world (your message, attachments, web/market/repo data via real tools) under the External Reality Rule. Cross-links to the other five loop stages. GuidesLivingGroundedJul 2026 📘MD Privacy Governance — unisona.ai Policy on never showcasing private/personal data; prohibited vs approved demo data, handling rules, and commit/CI enforcement. GuidesLivingUngroundedJun 2026 📘MD Reason — decide what to do The third stage: the model-agnostic reasoning that plans the response and decides which tools to use; any LLM (Claude/GPT/Gemini/Grok/local) plugs into the same loop. GuidesLivingGroundedJul 2026 📘MD Remember — recall what matters The second stage: Unisona’s durable, private, append-only memory + CSF archive and retrieval — why a returning user never re-explains themselves, and why it improves by remembering, not retraining. GuidesLivingGroundedJul 2026 📘MD Trader Guide (user) End-user path from zero to a running autopilot: signup + email gate, BYOK Alpaca paper connect, the account-mode and Off/Intraday/Champion switches, what the autopilot does when on, how to stop/disconnect, troubleshooting. Mirrored as public/trader-guide.html. GuidesLivingGroundedAug 2026 📘MD unisona.ai — What It Can't Do (Yet) Evidence-tagged audit of current limitations (no live web fetch default, loop not end-to-end in serving, autowork branch reuse, weak local kernel ~10% pass@1, 4 real skills). GuidesLivingGroundedJun 2026 📘MD unisona.ai Chat — Brand Guidelines Brand, color-token, typography, component, and voice spec sourced from site.css; correctly retires 'unisona.ai' as user-facing. GuidesLivingGroundedJun 2026 📘MD Verify — check it against reality The fifth stage and safety mechanism: nothing accepted without evidence — grounding, fact-check, and the drift canary, applied before a result becomes an input. GuidesLivingGroundedJul 2026 📘MD Accessibility (WCAG 2.1 AA) Standards WCAG 2.1 AA checklist, keyboard/screen-reader testing guide, and PR-review a11y list; uses dead 'unisona.ai' branding in prose. GuidesReferenceUngroundedJun 2026 📘MD Action Pooling and Action Batching Operating-method policy for batching low-risk actions and holding purchases/account/hardware actions; legal-asset and hardware-expansion loops. GuidesReferenceUngroundedJun 2026 📘MD Chat Capability / Benchmark Eval Recipe How to drive dream-chat through a prompt suite by hand and score it — the fast path plus the gotchas (cold-start provider errors, the 30s preview cap, mid-run crash recovery). GuidesReferenceGroundedJul 2026 📘MD Contributor Covenant Code of Conduct Standard Contributor Covenant v2.0 code of conduct: expected behavior, unacceptable behavior, and reporting via GitHub issue. GuidesReferenceUngroundedJun 2026 📘MD unisona.ai — UX & Wording Standards UX and wording standards for all pages: spacing/typography/color tokens, tone, section-naming, WCAG-AA contrast checklist. Some 'Lantern' example copy is stale. GuidesReferenceUngroundedJun 2026 📘MD unisona.ai Convergence Loop The 12-step operating/release-decision method for a convergence loop (inspect→map claims→validate→fix→promote/hold/reject), with promotion states. GuidesReferenceUngroundedJun 2026 📘MD unisona.ai — Canonical Character Cast Canon visual+voice descriptions for the four mascot characters (Lantern, Blinkbug, Eclipse, unisona.ai); symbolic identity, not live personas. GuidesLoreUngroundedJun 2026 🔌MD ChatGPT Custom Connector Setup for MCP Setup guide for wiring ChatGPT to the OAuth2/PKCE MCP server; stale-branded 'unisona.ai', assumes live OpenAI keys, and awaits an unresolved Cloudflare tunnel fix. Setup & integrationsReferenceUngroundedJun 2026 🔌MD Connect Claude & ChatGPT to the MCP Server OAuth2/SSE connector setup for wiring Claude and ChatGPT to the MCP server; stale-branded (lantern-os.net) but the OAuth flow is a stable procedure. Setup & integrationsReferenceUngroundedJun 2026 🔌MD Google Sign-In Setup How to configure 'Continue with Google' sign-in (client id/secret, redirect, session). Setup & integrationsReferenceGroundedJul 2026 🔌MD MCP Client Setup Guide to connect ChatGPT, Claude Desktop, Grok, and Cursor to the unisona.ai MCP server via OAuth or Bearer token. Setup & integrationsReferenceUngroundedJun 2026 🔌MD Patreon OAuth 2.0 Integration Setup guide for the Patreon OAuth login: app registration, env vars, tier-to-role mapping, endpoints, and troubleshooting. Setup & integrationsReferenceUngroundedJun 2026 🔌MD unisona.ai ↔ MCP / live project context How keystone-context.js injects live repo state (open issues/PRs via gh, MCP tool inventory, branch) into the chat system prompt, provider-agnostically; context-injection not tool-calling. Setup & integrationsReferenceGroundedJun 2026 🔌MD unisona.ai Dream Journal — Lightweight Deployment Quickstart for the standalone 229MB Dream Journal Docker image and its dream/chat API endpoints; mentions Grok interpretation which is no longer a live provider. Setup & integrationsReferenceUngroundedJun 2026 🔌MD unisona.ai MCP Connector Safety contract + verification runbook for the local-first MCP connector: evidence classes, held items, and a promotion gate before any agent trusts the surface. Setup & integrationsReferencePartialJun 2026 🔌MD User Profiles — Local-First Backend API+storage reference for the local JSONL/CSF user-profile system with Patreon roles; uses dead 'unisona.ai' branding in prose. Setup & integrationsReferenceUngroundedJun 2026 📈MD Σ₀ Collapse Certificate — the $2k investing plan Adversarial Σ₀ audit of the $2,000+$20/mo momentum-tilted DCA plan (investing-2k-plan.html): strict evidence classes, six collapse modes (incl. survivorship, now measured), and the minimal honest statement of what the plan does and does not prove. TradingLivingGroundedJul 2026 📈MD Kalshi Trading API Specification REST endpoint reference for the Kalshi trading API; stale-branded 'unisona.ai'; overlaps the fuller docs/trading-api-reference.md. TradingReferenceUngroundedJun 2026 📈MD Portfolio Integration Setup (IBKR) Setup guide for connecting IBKR/Alpaca portfolio data; system serves mock data until IBKR Gateway runs. Cites a port-5050 trading service. TradingReferenceUngroundedJun 2026 📈MD Trading API Reference Endpoint reference for the 60+ Kalshi/trading REST routes in routes/trading.js: markets, swipe deck, convergence model, monitor, crypto, history, AI trader. TradingReferenceGroundedJun 2026 📈MD unisona.ai Trader — Deep Analysis (2026-07) Deep analysis, competitive comparison, and improvement plan for the trading surface. TradingReferenceGroundedJul 2026 📈MD Unisona Sharpe Certificate Certifies the mathematics the chat + Σ₀ trader use to optimize client risk-adjusted return (diversification, tangency, vol-targeting, fractional-Kelly, regime-gating), separating proven theorems from empirical claims pending measured evidence. E5: single-stock momentum rejected — measured survivorship-free. TradingProposedGroundedJul 2026 🏛️MD ADR-0006: Dual-Boot 4177/4178 Worktree Topology ADR: two servers from two git worktrees (4177 stable auto-deploys master, 4178 dev); matches current dual-boot setup. Status: Proposed. ADRLivingGroundedJun 2026 🏛️MD arXiv Recent-Research Corpus Local corpus of post-cutoff AI/LLM arXiv papers (abstracts + metadata) on drive F:, retrieved by BM25 to ground chat answers in research published after the model's training cutoff. Strengthens the Remember stage. ArchitectureLivingGroundedJul 2026 🏛️MD Changelog Canonical reverse-chronological release history (Keep a Changelog); actively maintained, top entry is today's 1.8.3, with per-entry verification evidence. ArchitectureLivingGroundedJun 2026 🏛️MD Code → Convergence Core Mapping Canon architecture map of code to the four objects + six-stage loop; carries a real-time status banner correcting stale-branded body. ArchitectureLivingPartialJun 2026 🏛️MD Surface Boundary — Core Loop vs. Extensions Living registry classifying every public HTML surface as CORE (loop stage) or EXTENSION (module/flag); 18:30 ratio, gated by test:boundary. ArchitectureLivingGroundedJun 2026 🏛️MD The Convergence Agent (the one agent) Canonical spec of the single Convergence-Core product agent; each capability tagged wired/coded/designed with honest gaps. ArchitectureLivingPartialJun 2026 🏛️MD Third-Party Notices License-attribution registry for third-party code ported into unisona.ai (currently Aider repo-map clean-room); permissive-only porting policy. ArchitectureLivingGroundedJun 2026 🏛️MD unisona.ai — IP Register & 2-Year Buildout Plan Defensive-publication IP register + 8-quarter roadmap; honestly tags each invention [implemented]/[design]; most milestones planned. ArchitectureLivingPartialJun 2026 🏛️MD unisona.ai Cockpit — the Human-in-the-Loop Spine Built+tested spine for the personal-AI flow: editable approved/proposed Profile, ActionGate holding send/schedule/submit/spend, and smallest-useful-question surfacing. ArchitectureLivingGroundedJun 2026 🏛️MD Σ₀ Local-Model Adapter Implemented local-model registry contract: VRAM-gated selection, selfConverges wrap, resolveLocalLead pin precedence; feeds loop-reasoner. ArchitectureLivingGroundedJun 2026 🏛️MD AAPF — Agent Action Provenance Format Spec for the append-only ActionRecord/ProvenanceLedger audit trail (P3); implemented+unit-tested in Python, a reference contract not on the live JS chat path. ArchitectureReferenceGroundedJun 2026 🏛️MD ADR-0001: Record Architecture Decisions Foundational ADR establishing the append-only ADR practice and the owner-approval gate; status Proposed, awaiting Alex's approval. ADRReferenceGroundedJun 2026 🏛️MD ADR-0002: Single Convergence Core — Reject Architectural Sprawl ADR mandating one Convergence Core and extension-over-addition; forbids parallel subsystems. Status Proposed, pending approval. ADRReferenceGroundedJun 2026 🏛️MD ADR-0003: One canonical CSF module Records the v2 decision to keep exactly one CSF module (pack/unpack, legacy read-only); honestly flags the zstd-19 codec claim vs zlib-bound reality as open debt. ADRReferenceGroundedJun 2026 🏛️MD ADR-0004: Append-only JSONL + CSF Archive as the Only Memory Systems ADR fixing memory to append-only JSONL logs plus one CSF archive; bars a third store. Status Proposed, pending approval. ADRReferenceGroundedJun 2026 🏛️MD ADR-0005: Models Are Interchangeable — Provider Abstraction with Fallback ADR (Proposed) making providers config-not-code with a fallback chain; describes intended Gemini→Claude→OpenAI→Ollama routing (only Gemini/Vertex funded today). ADRReferencePartialJun 2026 🏛️MD ADR-0007: Monoworkstream — One Open PR Lane per Agent ADR for the one-open-PR-lane-per-agent workstream rule with options, trade-offs, and evidence; status Proposed, awaiting approval. ADRReferenceGroundedJun 2026 🏛️MD ADR-0009: One routing contract — cloud-primary for coding, local as verified backstop Accepted ADR: coding routes cloud-primary when a cloud key is reachable, local is the verified offline backstop, resolved in one pure route-contract resolver. ADRReferenceGroundedJun 2026 🏛️MD ADR-0012: Nested adaptive Reason Architecture Decision Record — Nested adaptive Reason — Q-exit within-model × fidelity-escalation across models. ADRReferenceGroundedJul 2026 🏛️MD ADR-0014: unisona.ai desktop Architecture Decision Record — unisona.ai desktop — a thin signed launcher over the one Convergence Core (not Electron). ADRReferenceGroundedJul 2026 🏛️MD ADR-0015: Qwen-teacher verified distillation into Ouro Architecture Decision Record — Qwen-teacher verified distillation into Ouro — teacher proposes, execution verifies. ADRReferenceGroundedJul 2026 🏛️MD ADR-0016: Provider-agnostic OSS auth (local + Google + Discord + Patreon) on req.session.user Architecture Decision Record — Provider-agnostic OSS auth (local + Google + Discord + Patreon) on req.session.user. ADRReferenceGroundedJul 2026 🏛️MD ADR-0018: Split delivery into a hosted multi-tenant web tier + a full local Core Architecture Decision Record — Split delivery into a hosted multi-tenant web tier + a full local Core. ADRReferenceGroundedJul 2026 🏛️MD ADR-0019: IBKR connectivity via the Client Portal Web API through a local gateway (read-only) Architecture Decision Record — IBKR connectivity via the Client Portal Web API through a local gateway (read-only). ADRReferenceGroundedJul 2026 🏛️MD ADR-0020: IBKR live order placement Architecture Decision Record — IBKR live order placement — gated, dry-run by default. ADRReferenceGroundedJul 2026 🏛️MD ADR-0021: Serving substrate Architecture Decision Record — Serving substrate — retain the Ouro/Σ₀ custom looped transformer, not a stock runtime. ADRReferenceGroundedJul 2026 🏛️MD ADR-0022: Per-user IBKR connection via self-service OAuth 1.0a Architecture Decision Record — Per-user IBKR connection via self-service OAuth 1.0a. ADRReferenceGroundedJul 2026 🏛️MD ADR-0025: RLVR + 'dreaming' continual weight updates, double-gated by verification Architecture Decision Record — RLVR + 'dreaming' continual weight updates, double-gated by verification. ADRReferenceGroundedJul 2026 🏛️MD ADR-0026: Ternary (1.58-bit) as the distillation target format for the ≤8 GB serving artifact Architecture Decision Record — Ternary (1.58-bit) as the distillation target format for the ≤8 GB serving artifact. ADRReferenceGroundedJul 2026 🏛️MD ADR-0027: One-click broker connect Architecture Decision Record — One-click broker connect — per-user OAuth2 (Alpaca now, IBKR next). ADRReferenceGroundedJul 2026 🏛️MD ADR-0028: Managed-strategy mode Architecture Decision Record — Managed-strategy mode — Sharpe-mandate acceptance gate, tax-aware. ADRReferenceGroundedJul 2026 🏛️MD ADR-0029: The IBKR streaming Challenger Architecture Decision Record — The IBKR streaming Challenger — a return-seeking duel against the champion book. ADRReferenceGroundedJul 2026 🏛️MD ADR-0030: The Spiral — a verified-cascade convergence loop as the owned local reasoning core Accepted (Alex Place, in-session, 2026-07-22). The architecture and the Phase-0 build ADRReferenceUnreviewed 🏛️MD ADR-0031: Spiral Coder — ARC-AGI-2 budgeted-efficiency track as the SOTA target Accepted (Alex Place, in-session, 2026-07-23). Extends ADR-0030(0030-spiral-verified-cascade-harness.md) ADRReferenceUnreviewed 🏛️MD ADR-0032: Real-money trading onboarding — how a user goes from paper to live, safely - Status: Proposed (awaiting Alex's approval) — design only; no live trading is enabled by this ADR ADRReferenceUnreviewed 🏛️MD ADR-0033: Apex trader — maximize profit at the highest ACCEPTABLE risk (tail-capped 3× overnight composition) - Status: Proposed (awaiting Alex's approval) — design only; nothing in this ADR arms or trades ADRReferenceUnreviewed 🏛️MD ADR-0034: MoE for the Σ₀ core — adopted as soon as feasible Accepted (Alex Place, in-session, 2026-07-24: "we are using MoE — as soon as is feasible", ADRReferenceUnreviewed 🏛️MD AGI v1.10 — the White-Box Verified-Honesty frontier design literature and patents. It does not authorize a training run; ADR-0024's kill-gated program and ArchitectureReferenceUnreviewed 🏛️MD CCF — Capability Claim Format Spec for the P4 Capability Claim Format primitive: time-boxed claims, honesty tracker, and a capability gate; implemented and unit-tested but not on the live JS path. ArchitectureReferenceGroundedJun 2026 🏛️MD CEG — Convergence Execution Graph Spec for the typed graph substrate (nodes/edges/optimizer/executor) the convergence-io stack plugs into; implemented+tested in Python, not on the live 4177 JS chat path. ArchitectureReferenceGroundedJun 2026 🏛️MD Convergence Routing Architecture Explains the deterministic intent router + pattern cache; provider-chain section lists Claude/OpenAI/Grok as live though only Gemini is funded. ArchitectureReferencePartialJun 2026 🏛️MD CSF Format Specification Canonical binary spec for CSF-Pack v0.8 (one zstd-backed format); documents v2 consolidation and read-only legacy path. ArchitectureReferenceUngroundedJun 2026 🏛️MD D — Time Dilation Field Reference spec for the dilation field primitive (D in CEG): field function, grounding-policy bridge, dwell caps, swap guard. ArchitectureReferenceGroundedJun 2026 🏛️MD DCF — Data Classification Format Convergence-io P1 primitive: every datum carries propagating class labels so downstream CCF/NAP gates decide use. Python-tested reference; JS chat path doesn't import it. ArchitectureReferenceGroundedJun 2026 🏛️MD Distillation is a Deferred Last Resort — Verify-Gated, Never the Target ADR (Proposed, awaiting Alex): preserves LoRA distillation as a guard-railed last-resort escape hatch to North Star 'never retrain', with collapse tripwire. ADRReferenceGroundedJun 2026 🏛️MD Grounding Knowledge Base — Provenance Manifest Tracked provenance record of 14 real-world industry reference manuals (NIST, NASA, FAA, etc.) with measured SHA-256, pages, licenses; PDFs gitignored. ArchitectureReferenceGroundedJun 2026 🏛️MD MCP and Dream Chat tool parity Reference spec for the shared tool surface: tool-runner.js as canon, MCP bridge, gates (CHAT_TOOL_EXEC), receipt shape, contract tests. ArchitectureReferenceGroundedJun 2026 🏛️MD NAP — Negative Authority Profiles Spec for the P2 denial-form authority gate: hard denials override capability claims, compose with deny-lists; implemented+tested in Python, not on JS chat path. ArchitectureReferenceGroundedJun 2026 🏛️MD PCSF — Provider Capacity State Format Spec for the P4 provider-capacity router (states, circuit breaker, quota, fallback chain); honestly notes the live JS chat path routes providers itself. ArchitectureReferenceGroundedJun 2026 🏛️MD The Σ₀ White Paper — a small AI you actually own The plain-English version of our model design. Every claim here has a precise, sourced twin in ArchitectureReferenceUnreviewed 🏛️MD Worldwide-Patent Corpus (prior-art / IP grounding) A local, offline BM25 corpus of worldwide patents that lets the unisona.ai chat ArchitectureReferenceUnreviewed 🏛️MD Σ₀: Small-Model Reasoning via Persistent Test-Time Verification We describe Σ₀, a program to build a small (~1.5B-parameter) local language model that ArchitectureReferenceUnreviewed 🏛️MD ADR-0008: The end product is a personal AI wrapper — capabilities are Tools + Skills in the one loop ADR defining unisona.ai as a local-first personal AI cockpit; capabilities added as Tools+Skills not subsystems. Status Proposed, awaiting approval. ADRDraftPartialJun 2026 🏛️MD ADR-0011: Own a proprietary Σ₀ base model Proposed (not approved): fork a PLT looped-coder into a unisona.ai-owned Σ₀ base, adapter-only, council+CSF native; staged, each stage gated on on-box evidence. ADRDraftPartialJun 2026 🏛️MD ADR-0013: Subsystem register + one-loop gate — every surface names a loop stage or is scheduled for extraction ADR (Proposed, awaiting Alex): every public surface is registered CORE (one loop stage) or gated EXTENSION in lib/surface-registry.js, enforced by test:boundary + audit:orphans; consolidates three colliding 0001 drafts from PR #1813. ADRDraftGroundedJul 2026 🏛️MD Convergence Hooks: Σ₀ Loop Across All Steps Design of the hook framework routing every action through the six-stage loop into a ConvergenceRecord; git integration wired, kernel/test/API/dashboard hooks still planned. ArchitectureDraftPartialJun 2026 🏛️MD Superfleet Swarm — Design Design for scaling the Convergence Core into a bounded local worker swarm; Phase 0-1 marked done, Phases 2-7 (supervisor/executor/researcher/mesh) still unbuilt. ArchitectureDraftPartialJun 2026 🏛️MD ADR-0017: Surprise-gated decoding Architecture Decision Record — Surprise-gated decoding — grounded intervention driven by the CSF/web surprise signal. ADRProposedGroundedJul 2026 🏛️MD ADR-0023: The default profile foregrounds the loop Architecture Decision Record — The default profile foregrounds the loop — extensions are opt-in. ADRProposedGroundedJul 2026 🏛️MD ADR-0024: Σ₀ frontier training program Architecture Decision Record — Σ₀ frontier training program — honesty-native pretraining. ADRProposedGroundedJul 2026 🧪MD GPU Training Pipeline Diagnosis — Why Ouro Fine-Tunes Weren't Landing Grounded diagnosis of five dispatch→CUDA bugs (all fixed on master) and the finding that Kaggle's pre-Ampere GPUs lack bf16, so Lightning A10 must be the primary target. ResearchLivingGroundedJun 2026 🧪MD Hardening the Reasoning Loop Against Collapse CSF-native anti-collapse defense-in-depth plan (19 items, red-team gaps); working ledger with reconcile notes marking landed fixes and remaining G1 gap. ResearchLivingGroundedJun 2026 🧪MD The Convergence Blueprint — one loop, everything we know AGI-scope synthesis organized by the Observe·Remember·Reason·Act·Verify·Converge loop: 2026 best practice + what's in-repo + the honest gap per stage. Thesis: rent the six capabilities we lack, own the Verify/grounding stage. Includes a worked Verify example (the survivorship-momentum grounding result). ResearchLivingGroundedJul 2026 🧪MD The Question Machine — Bidirectional Consolidation of the Loop Built+tested math core: backward costate + forward-backward consolidation surfacing the seam as a question, CAP/NAP gated; channels still stubs. ResearchLivingGroundedJun 2026 🧪MD Theorem C3 — Σ₀⁻¹ Prevents Permanent Freeze (all A) Machine-checked proof that the anti-collapse operator can't latch a permanent freeze; freeze-prevention closed for all A, contraction still scoped. ResearchLivingGroundedJun 2026 🧪MD Σ_G — The Grounding Ledger Composes three measured results — the price of fresh verified truth, the schedule grounding must arrive on, and the budget it imposes on safe self-improvement — into one accounting frame. Companion to the Collapse Certificate. ResearchLivingGroundedJul 2026 🧪MD Σ₀ — The Collapse Certificate Stability-certificate theory for convergence collapse; freeze-prevention proven for all A, non-normal contraction still a draft route. ResearchLivingGroundedJun 2026 🧪MD Σ₀ — The Collapse Certificate, Explained Plain-English explainer of the collapse/grounding math: PROVEN freeze theorem, MEASURED anti-collapse operator, HEURISTIC trigger; non-normal proof still open. ResearchLivingGroundedJun 2026 🧪MD Σ₀ Worklist — Council, Evals, and the Coder Swap Living backlog from the 2026-06-29 Σ₀ session: council-as-route+verify thesis, done items (unified council, SWE harness), and prioritized next steps (Qwen swap, real SWE number, RAG). ResearchLivingPartialJun 2026 🧪MD Claude Code Multi-Agent Spawning — Validation Verification record: Claude Code (harness/SDK) can spawn/coordinate/merge sub-agents (GA); raw Claude API cannot auto-spawn. Notes multi-agent cost is high. ResearchReferenceGroundedJun 2026 🧪MD Frontier Directions — 2026 H2 The deep-research + design roadmap of frontier directions for the second half of 2026. ResearchReferenceGroundedJul 2026 🧪MD Grounding the Convergence IO Oracle in Real Cross-Border Regulation Deep-research validation mapping convergence-io primitives to in-force regulation (OFAC/DORA/GDPR); 24/25 claims confirmed on primary sources. ResearchReferenceGroundedJun 2026 🧪MD Lemma L2 — the one-step anisotropy lift (closed form, proven) Closed-form + machine-checked proof that one Σ₀⁻¹ covariance bump breaks the collapse freeze trigger for all A; explicitly does not touch Theorem-1 contraction for non-normal A. ResearchReferenceGroundedJun 2026 🧪MD OSS Landscape Baseline — 2026-07-06 The 8-agent / 178-lookup audit establishing that unisona.ai is assembly of proven OSS components plus one accountability layer nobody ships — the honest moat baseline. ResearchReferenceGroundedJul 2026 🧪MD Prediction Markets as External Grounding / Reward Signal — Literature Survey Cited literature survey on using prediction-market prices as a calibrated grounding signal; verified references, viable-as-grounding verdict. ResearchReferenceGroundedJun 2026 🧪MD The Convergence Oracle — design Design for the answerability-first 'oracle above the 42 machine' — place/price/answer/learn discipline for grounding questions, not a cosmology claim. ResearchReferenceGroundedJul 2026 🧪MD The Convergence Oracle — time-banded observer slices Built+tested grounding capability: locates a question in cosmic time, returns cited knowns and honest unknowns per era; CLI + 6 tests. ResearchReferenceGroundedJun 2026 🧪MD Theorem 1 (non-normal) — the spectral-dichotomy extension Proves the contraction half of Theorem 1 for non-normal A via the Riesz spectral dichotomy; machine-checked (600-matrix sweep + pytest), scoped to the local linear Jacobian. ResearchReferenceGroundedJun 2026 🧪MD Σ_θ A/B/C Continual-Update Harness — Cloud L4 Spec Specification for the Σ_θ A/B/C continual-update harness on Cloud L4 GPUs — the acceptance-gated update loop. ResearchReferenceGroundedJul 2026 🧪MD Σ₀ Collapse Certificate as an Analogy for QM–GR Incompleteness Speculative essay: QM–GR incompleteness as metaphor for Σ₀ collapse; explicitly not a derivation; AI-safety section is the grounded keeper. ResearchReferencePartialJun 2026 🧪MD Σ₀ Frontier Training Brief The brief for the honesty-native pretrained frontier model — scope, targets, and the training program (pairs with ADR-0024). ResearchReferenceGroundedJul 2026 🧪MD Σ₀ Honesty Benchmark — results Honesty benchmark results with competitor comparison — measured numbers for the honesty/hallucination evals. ResearchReferenceGroundedJul 2026 🧪MD Σ₀ Model Design (serving layer) The serving-layer model design — the Fable-max deliverable plus the orchestrator wiring for the Σ₀ stack. ResearchReferenceGroundedJul 2026 🧪MD Σ₀ Real-Data Grounding — the Parrot Attractor on a Real Log Runs the committed Σ₀ collapse-certificate + surprise monitor on a real conversation log; honest real-vs-synthetic provenance for each result. ResearchReferenceGroundedJun 2026 🧪MD Σ₀ Verified arXiv Reading Pack — July 2026 A curated, web-verified set of arXiv papers grounding the Σ₀ / serving / honesty work, with one-line relevance notes per paper. The reading list behind the research canon. ResearchReferenceGroundedJul 2026 🧪MD Σ₀-K1 — The first buildable convergence kernel (spec) Frozen spec for the Ouro-1.4B convergence kernel; honest that the core model is weak (pass@1 0.1); components built+tested, loop closure not live. ResearchReferencePartialJun 2026 🧪MD Explore as a Σ₀-Ranked, Profile-Fed Content Machine Design proposal to collapse Explore into one ranked card feed via PCSF leaderboard + interest log; web-grounded on Meta/HSTU recsys patterns. ResearchDraftPartialJun 2026 🧪MD unisona.ai-Σ₀: Self-Converging Local Kernel + Chat Harness for Sustained Work Design for wiring depth/halt/surprise telemetry through serving→chat→autowork; honest about Stage-0 parity gap and refuted 'ahead' claims. ResearchDraftPartialJun 2026 🧪MD Σ₀ V10 Video Scoring — Theory-to-Implementation Mapping Maps collapse-certificate math onto the V10 video scorer; the doc itself flags most guarantees as aspirational/not-computed-at-runtime. ResearchDraftAspirationalJun 2026 🧠MD Σ₀ Continual-Training Loop — the closed flywheel Offline opt-in harvest→verify→train→eval→promote loop for the local Ouro coder; gates built+tested, full-HumanEval rerun still pending. Models & trainingLivingGroundedJun 2026 🧠MD models/lantern-sigma0-coder — directory note Note documenting the retired lantern-sigma0-coder model and what its directory is now repurposed for. Models & trainingReferenceGroundedJul 2026 🧠MD Ouro Multi-Source Session Dataset How to build the private multi-source Ouro training dataset (Claude/Codex/ChatGPT tool trajectories) from local sessions; provenance/redaction rules and Kaggle mount path. Models & trainingReferenceUngroundedJun 2026 🧠MD Serving Architecture 2026: Fast Default + Deep Research Mode FAST cached vs DEEP Σ₀ Q-exit serving-mode decision with decode params and benchmark contract; calls itself unisona.ai (stale brand). Models & trainingReferencePartialJun 2026 🧠MD When to use FAST vs DEEP mode Decision guide for FAST vs DEEP serving modes; honest that connector DEEP only swaps decode params and benchmarks are unrun TODO. Models & trainingReferencePartialJun 2026 🧠MD Σ₀ Coder Serving Architecture — Staff Portfolio Update (v2) Peer-review plan for a swappable Ouro coder serving topology (cloud floor + ONNX embedded spike); honestly flags adaptive engine unmeasured and Ouro not yet the live coder. Models & trainingReferencePartialJun 2026 🧠MD Σ₀ Convergence Adapter — spec, data, eval, GPU program Proposed verification-gated fine-tune spec teaching a model native Σ₀ discipline; generator landed, eval + adapter not yet built. Models & trainingDraftPartialJun 2026 🎬MD Creator Dashboard — Learning Pipeline Research Honest reframe of a viral-editing pipeline; documents that 'viral' Σ₀ features are title-text proxies and the 5000-shorts study was synthetic/retracted. Creator suiteLivingGroundedJun 2026 🎬MD Editing & Analysis Model Research — Improvement Plan Research + progress log for the creator dashboard's analysis/editing models (shot detection, retention curves, Whisper, novelty scoring), with A1-B4 items marked done in code. Creator suiteLivingPartialJun 2026 🎬MD Editing Discovery Engine Creator discovery engine that mines recurring editing role-sequences into a playbook; structural core shipped+tested, performance gated as insufficient_data. Creator suiteLivingGroundedJun 2026 🎬MD Σ₀ Open-Video Research Loop Design + status of a delete-after-analysis flywheel that learns observable editing priors from open-license video; explicit honesty boundary and empty seed store, no live corpus yet. Creator suiteLivingGroundedJun 2026 🎬MD CaptionEngineV3 Spec for the event-driven caption engine: signal-derived event taxonomy, 50+ data-defined styles, ffmpeg effects; behind a default-off flag. Creator suiteReferencePartialJun 2026 🎬MD Creator Dashboard — Design & User Guide Design/user guide for the creator upload+analysis page (/create.html); accurate feature/API spec but 'Production Ready' framing and 'Lantern' branding are stale. Creator suiteReferencePartialJun 2026 🎬MD Creator Intelligence — Subsystem Architecture Architecture of src/creator-intelligence: data-first modules with an insufficient_data honesty contract (no invented metrics) and per-video vs population truth split. Creator suiteReferenceGroundedJun 2026 🎬MD ExportValidator Spec for the ffprobe-backed hard gate that blocks a short-form export failing resolution/fps/codec/duration/caption checks; flag on by default, blocks rather than passes when ffprobe absent. Creator suiteReferenceGroundedJun 2026 🎬MD Facecam Detection + Top-Locking How the creator render pipeline detects a facecam by its seam (not skin) and locks it to the top band; before/after corners and real-clip ffmpeg verification. Creator suiteReferenceGroundedJun 2026 🎬MD Research Dataset Schema Row-shape spec for the Creator Intelligence dataset (GeneralShort/GamingShort/EditEvent), storage layout, opt-in ToS-respecting collection, and sufficiency thresholds; ships empty. Creator suiteReferenceUngroundedJun 2026 🎬MD SafeZoneDetectorV2 Spec for a multi-frame safe-zone detector that protects facecam/HUD/crosshair when reframing gameplay to 9:16; fixes the V9 zone-discard bug, flag-gated, honest confidence. Creator suiteReferencePartialJun 2026 🎬MD VariantEngineV2 Spec for a creator variant engine (A–E edit recipes + honest insufficient_data ranking) behind the LANTERN_CI_VARIANT_V2 flag; implementation steps listed. Creator suiteDraftPartialJun 2026 📋MD Memory Recall Benchmark (LongMemEval) — methodology + how to run Re-runnable harness measuring unisona.ai's live memory retrieval on LongMemEval vs Letta/Mem0/Zep; committed fixture is synthetic, real dataset gitignored, incumbents wired not run. ReportsReferencePartialJun 2026 🗒️MD Research notes — 88 dated entries Every dated research note, newest first: experiments, negative results, and design records from 2026-06 to 2026-08. Working notes, not polished guides. ResearchLivingGroundedAug 2026
Project

Project & releases