← Neo
CONNECTV1

What changed, and why does it matter now?

CONNECT V1 gives Neo the ability to continuously watch what is changing outside a company, understand what is currently true inside it, and surface the connection when those two things materially intersect, before someone knows to ask.

CONNECT is the architecture. Neo is the operator who acts on it. It is not a separate product, an API, or an agent of its own.

01

Scouting

Models, articles, news, research and job-specific developments, watched continuously.

New models and releases, capability and pricing changes, benchmarks, context windows, availability. Industry and competitor news. Research papers, repositories, changelogs and provider documentation. The client defines the territory: AI infrastructure, pet products, SaaS competitors, suppliers, whatever the job requires.

Scouting finds things. CONNECT decides whether what was found matters to this company now.

02

Market and competitive intelligence

Continuous intelligence about competitors and markets, tied to what the company currently believes.

Competitor launches, pricing moves, new and discontinued products, positioning shifts, supplier changes, strategic announcements, infrastructure developments.

Not a competitor digest. The question is whether a development weakens something the company believes, invalidates an assumption, changes a cost, creates a deadline, or affects an open decision.

03

Community monitoring and chat

Neo operates inside a live community rather than observing it from outside.

Monitors conversations, answers members directly, holds room context and searches conversation history, recognises what the community has been discussing over time, welcomes new members, escalates what needs a human. Public notes inform him; private notes never enter the prompt.

A live Neo capability available alongside CONNECT V1. It is not part of the frozen external-event pipeline.

04

Continuous sensing

A deliberately defined external environment, watched on a schedule.

News, articles, pricing pages, provider updates, company announcements, technical releases, selected social sources. Scheduled collection with freshness checks, publication dates, lookback recovery if a pass is missed, and deduplication by URL and by content hash.

Collection is deliberately not relevance-filtered. Noise enters. Relevance is decided later, which is what makes a quiet result meaningful.

05

Live company context

A structured representation of what is currently true inside the organisation.

Priorities, assumptions, open decisions, pricing facts, positioning, product capabilities, constraints, commitments, relationships, campaigns.

Not a document archive. External change can only be judged against a usable representation of present reality.

06

Temporal awareness

Company facts are held with dates, sources and supersession rather than as flat truths.

A record carries what is believed true, when it became true, whether it is still current, where it came from, how confident it is, when it was replaced and by what.

Not “X is true” but “X was true from this date, on this evidence, until it was superseded.” Stale context cannot quietly stay active.

07

Connection discovery

The central capability: does this outside event matter because of something currently true inside?

External changes are compared against internal state and classified by how they relate: weakens, strengthens, invalidates, creates a deadline, changes cost, changes competitive position, or no relation at all.

Without requiring a human to formulate the connecting question first. This is the load-bearing difference.

08

Evidence-bound intelligence

A conclusion stays attached to the evidence that produced it.

A supporting claim must be tied to a span that actually appears in the source material. If the claim cannot be bound to evidence, the adjudication fails.

Unsupported claims are not softened into vague language to preserve an output. They are dropped.

09

Verification

Sources, facts and figures are checked before anything is surfaced.

The source URL resolves. The material is fresh enough. The supporting text exists and is retrievable. Any number, price or percentage cited is checked against the source it came from.

Model memory is never allowed to substitute for evidence.

10

Significance filtering

Separating “related” from “actually important”.

Whether a relationship exists and whether it deserves human attention are two different judgements, decided separately. Scoring components and the surfacing threshold are fixed before any result is seen.

Something can be entirely true and completely irrelevant. A true connection does not automatically become a surface.

11

Quiet by design

If nothing matters enough, Neo says nothing.

Output is capped rather than maximised. Weak relationships are rejected. Quiet days are expected and are not treated as a fault.

Silence is not a failure. The system fails commercially the day people learn to ignore it.

12

Human-controlled state

External events never silently rewrite what the company believes.

“The outside world changed” and “the company has changed its position because of that” are different events. Only the second updates internal state, and important changes require human confirmation.

A trust boundary, deliberately placed. Neo can challenge an assumption; he cannot quietly replace one.

Neo

Scoutingfeeding Neo

What CONNECT V1 is not

  • It does not act. No outbound contact, no pitching, no commitments, no spending.
  • It does not learn relevance from feedback. Evaluation labels are evidence, not training.
  • It does not predict what a named counterparty will do.
  • It does not rewrite company state from an external event without a human.
  • It is not a standalone API, a platform for other agents, or multi-agent infrastructure.

V2 is defined and not built: pattern memory, and whether what a company says publicly is still true given what is now true internally. V3 is undefined.

Where it actually stands

CONNECT V1 is being engineered and validated. The first dry run froze its company state, source list, prompts, scoring weights, threshold and declared expectations before collecting anything, and runs to a fixed window.

A null result is a legitimate result. If nothing surfaces, the threshold does not move afterwards, the sources do not widen, and the run is not repeated to manufacture something. Whatever it produces goes on this page, whichever way it falls.