Your company knows more than it remembers.

Real-world example

A London-based growth equity firm spent three years building a knowledge base of 14,000 documents, deal memos, board minutes, market maps. An internal audit found the majority of documents were stale, contradictory, or disconnected from any actual decision made in the past twelve months. Associates had stopped using it. The firm had built a library, staffed it, filled its shelves, and it quietly died while everyone was looking at their inboxes.

Foundation-AI was built to solve exactly this. Not just store knowledge, but keep it alive.

Foundation Media Intelligence, more than a fact-checker. It's a full perception and provenance layer.

Before that Corp A press release enters Foundation-AI's memory, Foundation Media Intelligence runs a battery of checks: domain legitimacy, AI-generated disinformation detection, deepfake scoring on any embedded images. But what Foundation Media Intelligence produces is more than a pass/fail, it produces a Media Passport.

Think of a Media Passport like a wine's provenance certificate: it records where the information came from, what checks it passed, what transforms were applied to it, and what confidence level it earned. That passport is permanently attached to the piece of information. It travels with it everywhere, into the knowledge base, across PACT contracts, into any contribution packet that references it. When an analyst asks a question six months later, the system knows not just what the information says but how much to trust it and why.

Every piece of information gets an authenticity score from 0.0 to 1.0. A score of 0.3 means "treat with skepticism." A score of 0.95 means "this is reliable." That score, and the full provenance chain behind it, can be verified by any authorised party through Midnight, without revealing the underlying content. Proof of trustworthiness, without exposure.

Why this matters, the disinformation problem

In 2024, the cost of creating a convincing fake video of a CEO making a statement dropped to approximately $0. A fake press release is indistinguishable from a real one to a human reader skimming hundreds of documents a day. An AI-generated analyst report looks identical to a real one. If Foundation-AI simply ingested everything it found on the internet without verification, its knowledge base would gradually fill with fabrications, and the system would confidently cite them as evidence.

Foundation Media Intelligence exists to prevent this. Every piece of content has to carry provenance before it earns a trusted place in the knowledge base. Importantly, this check verifies provenance and manipulation status, not factual accuracy. A genuine press release from a real company can still contain projections that prove wrong. The live system proves passporting, fingerprinting, disclosure tagging, sensitivity tagging, and gate enforcement; production deepfake and manipulation workers attach scores as those model lanes come online. Contradiction detection and confidence decay handle the harder question of whether the content is actually true.

Current status: the Media Passport schema, content fingerprinting, the disclosure and sensitivity tagging, and the ingestion gate are live in production. The automated deepfake and manipulation-scoring models are being rolled out, so some passports today carry a "not yet evaluated" score rather than a numeric one.

Real-world example: why the authenticity score changes everything

A Scout discovers a video on social media showing a Japanese manufacturing CEO announcing a major factory closure. This would be significant news. Without Foundation Media Intelligence, it enters the knowledge base and becomes a fact. An analyst asks about the company's capacity. Foundation-AI cites the video as evidence. The analyst acts on it.

With Foundation Media Intelligence: the video enters the media passport lane first. If the production deepfake worker is enabled for that deployment, it can attach a manipulation score and authenticity score. If not, the passport says the video is not yet evaluated for that lane, forcing corroboration before the content can carry high evidentiary weight. Either way, the system surfaces the trust boundary instead of treating the video as ordinary evidence.

Foundation Media Intelligence didn't prevent the fake from entering the system. It prevented the system from treating the fake as trustworthy evidence.

What Foundation Media Intelligence extracts from a video or audio file

Clips: Specific segments extracted automatically, for example, every time a company name or topic is mentioned. A Scout set to monitor a company's earnings calls will receive an automatic clip of every moment that company is referenced, across every call Foundation Media Intelligence has processed.

Highlights: A curated set of the most significant moments, with timestamps and topic tags. A two-hour analyst day gets reduced to a highlights reel: key announcements, guidance changes, Q&A moments, and anything where sentiment deviates from the baseline.

Transcript: A searchable, timestamped text version of everything spoken, so that knowledge from video and audio enters the same searchable knowledge base as documents. What the CEO said in a Tokyo investor briefing becomes searchable alongside what was written in the annual report.

Authenticity check on the audio itself: Modern voice-cloning technology can produce convincing audio of someone saying something they never said. Foundation Media Intelligence checks for the telltale patterns of synthesised speech, detecting AI voice generation the same way it detects AI video generation.

Real-world example. Tokyo

Your analyst in London asks Foundation-AI: "What's Corp A's satellite strategy?" Foundation-AI already knows the facts, it has crawled Corp A's announcements, partner filings, and press releases. What it needs from an external AI is not a lookup, but reasoning about those facts. So what it actually sends is something like: "A major telecoms group has significantly increased infrastructure investment and signed multiple partnerships in the satellite connectivity space over the past 18 months. What does this pattern suggest about competitive positioning in this sector?"

No company name. No identifying details in the query itself. It is worth noting a real limitation here: in some cases, a sufficiently specific combination of facts — a unique capital figure, a sector, and a timeframe — could still allow an informed observer to identify the company even without a name. Foundation-AI's anonymisation substantially reduces disclosure risk; it does not eliminate it entirely for highly specific or uniquely identifiable situations. The system's default is to generalise facts precisely to reduce this risk. The reasoning comes back. Foundation-AI plugs Corp A's name back into the answer. Your analyst gets the insight. The external AI never learned whose facts it was reasoning about.

Every external AI call gets a receipt: who asked, what facts were sent (stripped), what reasoning came back, what trust score was assigned. Logged permanently. Cannot be altered.

Every single call to an outside AI passes through a 10-step process. No exceptions, no shortcuts:

Foundation-AI uses four sensitivity levels for outgoing information:

• Level 0. Blocked: The request never leaves Foundation-AI. Full stop.
• Level 1. Public only: Only publicly available information is sent. No anonymisation needed because nothing is sensitive.
• Level 2. Anonymised: All names swapped, all details stripped. This is the default for most questions. The outside AI reasons about abstract entities, never real ones.
• Level 3. Raw (disabled): Sending real data with special approval. This option is turned off in production. It exists but cannot be activated without an explicit operator decision.

The brain's spending limits, enforced every 8-hour cycle

Cortex is powerful, but it operates under strict budget limits it cannot override:

• 50 calls to outside AI per cycle, talking to AI costs money; the system rations it automatically
• 100 external web lookups, so it can't accidentally overwhelm a data source
• 20 knowledge refreshes, focused updates, not blind re-indexing of everything
• 25 service restarts per cycle, plus a hard cap of 3 automatic kills per hour, so it can fix failing components but never thrash them
• 30 minutes maximum per cycle, prevents runaway processing from consuming resources

These limits are enforced by a separate guard (the Circuit Breaker) that Cortex cannot negotiate with or override. Even the brain has a supervisor.

Real-world example: Cortex predicts, not just detects

A San Francisco VC firm uses Foundation-AI to track portfolio companies. Cortex notices the quality of answers about a Series B company in their London portfolio has been declining by 2% per day for a week. No alarm has triggered yet. But Cortex projects: "At this rate, answers about this company drop below acceptable quality by Tuesday." It flags it today, before anyone notices, with a specific recommendation: schedule a fresh data crawl this weekend.

This is the difference between a thermometer and a weather forecast. Most systems give you a thermometer.

The incorruptible guard

At the highest level, Foundation-AI can modify its own agent configuration, playbooks, and logic, but not its compiled core systems. The Circuit Breaker prevents that. What "self-patching" means in practice: the system can update how it behaves, not what it fundamentally is. This is powerful, and requires an equally powerful constraint. Think of how Tokyo's metro platform screen doors work: they don't ask the driver's permission to close, they don't check a policy file, they don't take exceptions. A sensor says "train has left" and the doors close. Period. Foundation-AI's circuit breaker works the same way, a completely separate process written in C++, compiled, with no configuration file. If any part of Foundation-AI tries to touch a file it's not allowed to touch, the circuit breaker kills that process immediately. Not after review. Not after logging. Immediately. It doesn't read policies. It doesn't listen to arguments. It just watches and acts.

You cannot ask a powerful system to reliably limit itself. So Foundation-AI built a simpler, dumber system whose only job is enforcement.

Current status: in production deployments the top "Self-Patcher" tier is off by default. Autonomy runs up to the Autonomous level, and self-patching is gated behind an explicit operator opt-in, so the system proposes and applies configuration changes only where an operator has switched that tier on.

The five hard invariants that always trip a critical, kill-priority alarm

Most of Stasis's thresholds are adaptive, they learn what "normal" looks like for a specific deployment. A London office running heavy Japan market research will have different baseline traffic patterns than a San Francisco seed fund. Stasis learns both. But five things are sacred regardless of context. If any process leaks a redacted name into an external output (redaction_leak). If a governance receipt fails its integrity check (receipt_integrity). If private embeddings are exported off-node (embedding_denied). If a raw exception carrying sensitive data reaches an external surface (raw_exception_disabled). If generated content escapes the safety filters (unsafe_output_escape). In any of these five cases, Stasis raises a critical, non-adaptive violation, no questions, no learned thresholds, no second chances, and the independent Circuit Breaker enforces the kill. These rules are the same in Tokyo, London, and San Francisco.

Real-world example: research that spawns itself

A Tokyo-based fund creates a Research Scout to investigate Japanese fintech licensing changes after the FSA signals new rules. While running, the Scout discovers three specific companies worth monitoring individually, a Kyoto payments startup, a London neo-bank expanding into Japan, and a San Francisco crypto firm seeking FSA approval. It spawns three sub-scouts, each with their own schedule, their own state, their own job. Those sub-scouts can each spawn further sub-scouts, up to 50 total descendants from one root. Every spawning step requires approval from Cortex. The approval takes milliseconds, but it can't be skipped.

19 places a Scout can send its findings

The Foundation dashboard, the chat interface, email, SMS, Discord, Slack, Telegram, Microsoft Teams, Dropbox, Google Drive, OneDrive, Notion, Airtable, a custom web address, an automated callback, your calendar, your CRM system, a structured data export, or a file download. The dashboard, chat, email, file export, and the major cloud-drive and CRM destinations are wired today; the rest light up through connectors as you enable them. A Scout isn't just a researcher, it's a researcher with a publishing deal that covers every channel you use.

The tombstone, preventing re-learning

A London analyst's Scout spends two weeks tracking rumours that a San Francisco startup is about to raise a Series C. The rumours turn out to be false, the company is actually preparing to wind down. Foundation-AI forgets the Series C information, but leaves a tombstone: "We investigated this. The evidence was wrong. Don't revisit." Three weeks later, a different Scout stumbles across the same rumour circulating on LinkedIn. The tombstone is designed to stop it from being automatically re-promoted as fresh intelligence. A human reviewer can override a tombstone if they believe the original assessment was incorrect. Without tombstones, systems endlessly rediscover and re-store noise they have already evaluated and discarded. The trade-off is that a wrong forgetting decision becomes sticky. The audit log preserves every tombstone decision for human review.

Real-world example: the playbook

A San Francisco fund's Foundation-AI instance keeps surfacing a particular Tokyo financial blog as a credible source for semiconductor analysis. Analysts correct it on separate days, the blog has a known bias toward domestic suppliers. Once the same kind of correction repeats (the system's pattern threshold is three of a kind), it proposes a playbook bullet: "When analysing Japanese semiconductor supply chain, weight sources from [blog X] at 20% of normal, track record is poor." On approval, that bullet applies to every agent handling semiconductor questions across the whole system. A repeated pattern of corrections → a system-wide playbook update.

The model and expert lanes, and exactly what each one does

The Senior Analyst (Claude Opus). The most powerful AI available. Used for deep analysis, synthesising research into final reports, and reasoning through complex strategy questions. Every call goes through the privacy border first, it never sees your company's real name.

The Fast Reader (Claude Haiku). A smaller, faster model. Used for understanding what you're asking, identifying which companies are mentioned, and quickly classifying incoming documents. Think of it as the intake coordinator who reads everything first and decides where it should go.

The Librarian (BGE-M3). Runs entirely inside Foundation-AI on a dedicated graphics processor. Converts every piece of knowledge into a mathematical shape so it can be searched by meaning, not just keywords. When you ask about "Japanese telecom strategy," it finds relevant information even if those exact words never appear. Never communicates with the outside world.

The Research Assistant (local vLLM lane). Also runs entirely inside Foundation-AI. Reads and tags incoming documents, summaries, key facts, and company mentions. Does all of this without ever sending anything outside. Completely local for supported workloads.

The first two talk to outside AI servers, but through the full 10-step privacy process. The last two never leave the building at all. This means Foundation-AI can do most of its work without ever talking to an external AI, and when it does, your secrets are fully protected.

Now: learning on rails

Every one of those self-improvements — a playbook update, a model-routing change, a promoted memory, a refined prompt — passes through a governed gate before it can touch the live system: propose → offline-evaluate → staged rollout → auto-rollback → quarantine. A change that regresses in a canary is rolled back automatically; a rejected lesson is quarantined so it cannot be re-proposed. The whole loop is measured on a live instrument panel — one counter per learning lane — so "the system is improving safely" is a number you can read, not a promise.

And it is deliberately earned, not flagged. Most lanes run shadow-first: they record and score every real decision but do not change it until that lane has demonstrably earned the right to, one tier at a time. Pruning stale memory and promoting reused knowledge are already live; the rest record in shadow today and graduate to acting only when the evidence says they should.

The key difference: store vs. compile

A database stores things and retrieves them. Alexandria compiles knowledge, like a compiler turns raw code into something a machine can execute. When a new analyst report arrives about London-listed Japanese tech companies, Alexandria doesn't just file it. It identifies every company mentioned, checks whether they have existing profiles in the library, adds cross-references in both directions, and flags any summaries that are now outdated in light of this new data.

The result: when you ask about Corp A's London ambitions, you don't just get documents that mention Corp A, you get the whole connected map of what Foundation-AI knows about Corp A, its UK portfolio companies, its relevant competitors, and every analyst note that has ever touched those relationships.

The "I Don't Know" Principle. Foundation-AI's most important design decision

Most AI systems must produce an answer. Every time. Even when the evidence is thin. The result is a confidently worded guess that looks like knowledge.

Foundation-AI's deliberation council is built differently. Any model can abstain, to say "the evidence is insufficient" or "this falls outside my competence." A London fund asks: "Is now the right time to enter the Japanese retail REIT market?" Two experts abstain. One is bullish. One is bearish. The council doesn't pick a winner, it surfaces the disagreement: "Two experts lacked confidence to opine. The remaining ones are split. Here is exactly where and why they diverge."

An honest map is more useful than a falsely confident answer. This is the single most important design decision in the entire platform.

Real-world example: catching strategy drift. Tokyo

A San Francisco VC fund has spent six months building a go-to-market strategy for Japan, anchored on the belief that the FSA's stance on foreign fintech entrants remains open. Foundation-AI's Strategy Twin has been tracking that belief: "FSA posture toward foreign fintech, confidence: 78%."

Over three weeks, Foundation-AI quietly notices: an FSA policy working group published new language around "domestic data sovereignty" (−6%), a Tokyo-listed competitor quietly withdrew its foreign partnership application (ambiguous, −4%), and an internal briefing note from the London team contradicted the timeline assumptions (contradiction flag, −8%). Confidence is now 60% and falling. Foundation-AI surfaces this: "Your Japan entry thesis rests on a regulatory assumption that has weakened by 18 points in three weeks. Three signals suggest the environment is shifting. Here they are, in order of significance."

You didn't notice. The black box did.

What the Digital Twin actually does, a plain-English walk-through

Imagine Kenji, 34, living in Shibuya, Tokyo. He's a Corp A customer in a future production deployment built on the bounded five-layer proof. Foundation-Corp A holds his Digital Twin under his consent rules. Here's what that means in the target architecture:

Layer 1. Identity: Kenji's identity is verified, his name, age, address, and payment details have all been confirmed by Corp A. But his Digital Twin doesn't store his passport scan. It stores a proof of his passport, a mathematical certificate that says "this person is real, verified, and creditworthy" without revealing any of the actual details. Like a nightclub stamp on your wrist: it proves you passed the door check, without telling anyone inside what your ID said. This proof is routed to the Midnight connector boundary so authorised parties can verify it once the deployed connector read-back is live.

Layer 2. Context Graph: The twin knows Kenji has been a Corp A subscriber for 8 years, recently searched for family health insurance (twice), reduced his commute frequency (inferred from location data he's opted into), and increased his spend on grocery delivery services. These are observed facts in episodic memory, building a context graph of who Kenji is becoming.

Layer 3. Belief State: Foundation-Corp A's Cortex holds beliefs about Kenji: "Kenji is likely planning a significant life change, confidence: 71%." "Kenji's next major purchase is likely insurance or property-related, confidence: 63%." "Kenji is receptive to wellness offerings, confidence: 58%." These confidence scores reflect the system's Bayesian weighting of observed signals, not a claim of predictive accuracy. They decay daily without fresh evidence. Each belief has a confidence score, a decay rate, and a contradiction flag that fires if new behaviour disconfirms it.

Layer 4. Intent Model: Stasis detects life-stage drift. Kenji is transitioning from a single young professional to early-family life. The twin models this transition and predicts the next set of needs, not by guessing, but by matching Kenji's behavioural trajectory against millions of similar trajectories, locally, without his raw data ever leaving the node.

Layer 5. Agency: Kenji's target twin can act inside pre-approved limits. It can contact specialist Foundation nodes on his behalf to request quotes, retrieve information, or compare options. It can route payments only through authorised x402-over-MCP provider connectors once provider and payout receipts are live. It is Kenji's representative in the Federation, working within the rules Kenji has set. Kenji sees recommendations appearing on his screen. He doesn't see the twin working for him in the background.

The simplest way to understand Foundation-AI

A senior analyst leaves your London office and takes three years of Japan market knowledge with her. A new hire joins your San Francisco team and spends her first month rebuilding context everyone else already has. Your Tokyo team builds a strategy on a regulatory read that quietly became outdated two months ago. Foundation-AI doesn't eliminate these problems entirely, but it means your organisation's knowledge compounds instead of leaks. What the London analyst knew gets preserved, structured, and surfaced when the San Francisco hire needs it. The regulatory read gets flagged the moment the evidence shifts. The Tokyo team sees the drift before it becomes a crisis. The system never sleeps, never changes jobs, and never forgets to update the wiki.

And for the individuals who interact with the system, the customers, the patients, the buyers, the travellers. Foundation-AI is building toward something more ambitious: a Digital Twin. The target is a sovereign, receipt-enforced representation that works on a person's behalf across the Federation. The bounded five-layer proof described above is live; deployment-specific human-twin claims still depend on the tenant surfaces, connector receipts, and compliance gates attached to that deployment.

That is not a feature. That is a different relationship between people and the systems that serve them. And it is built into the architecture itself, not written in a terms-of-service document.

Real-world example: an exception becomes precedent

A renewal comes up for a London customer. An agent proposes a 20% discount. Policy caps automated discounts at 10%, so the exception routes for approval. The agent gathers the context that justifies it, two SEV-1 incidents last quarter, an open escalation, a near-identical exception granted to another account in the same sector, and routes it to Finance. Finance approves. In most systems that reasoning evaporates: the opportunity record shows "20%," and nothing more. In Foundation-AI, the inputs, the approval, and the rationale are sealed as a durable precedent. Months later, when the next renewal in that sector hits the same wall, the agent doesn't start from zero. It finds the precedent, cites it, and proposes the same structure, with the receipts to back it.

★ This is not a concept, it is running now

Ibex APAC runs Foundation-AI in production today. Every grant application, partnership proposal, MOU framework, and budget document that Ibex submits externally is routed through Foundation-AI before it leaves the building. The system reads the draft, cross-references it against the Alexandria knowledge base of prior submissions and governance policies, flags any commitments that require Managing Director sign-off, checks for conflict-of-interest signals, and ensures the proposal reflects the current strategic posture.

As of today: live services. verified knowledge items. Daily governance cycles. Every proposal Ibex submits has a permanent audit trail. The system never sleeps, never forgets to check, and has never missed a governance cycle since deployment.

The foundation is not a future product. It is running now.

The Collateral Review Gate — Foundation-AI checks its own claims

One of the least obvious but most important components in Foundation-AI is the Collateral Review Gate. It validates every outward-facing statement the system produces against a capability matrix — a registry of what the system can actually do, verified by live acceptance tests.

There is a prohibited patterns list. Claims the system catches itself making and will not publish. Three examples from the actual list:

• "Human digital twin" — blocked unless the specific implementation is confirmed in place and tested
• "Privacy border fully enforced" — blocked because universal claims are unprovable
• "Self-healing immune system" — blocked because it implies a complete and finished system

Governance applied to the system's own narrative. Foundation-AI will not make claims it cannot prove. That constraint applies to itself as much as to anything else.