Ghost

The problem

The work got faster. The company did not necessarily get smarter. Agents now ship more work than any team could, and they repeat yesterday’s mistakes at machine speed while they do it, because nothing is holding on to what yesterday taught anyone. A company is running more work than it can see, and the hard part is no longer doing the work. It is seeing it, tracing why, and keeping what it taught you.

The layer that is missing

The software stack is finished except at the top. Systems of record store what happened. Systems of work move it through pipelines and boards. Systems of action arrived with agents and do the work. Above that there is supposed to be something that turns outcomes and judgment into capability that compounds, and almost nobody ships it. Models reason. Agents act. Nothing owns the decision. That gap is what this is.

• Systems of record: ERP and CRM store what happened
• Systems of work: software moves it through pipelines and boards
• Systems of action: agents do the work, then forget they did it
• Systems of learning: turns outcomes and judgment into capability. The gap

The rethink: one platform, many canvases

Ghost is one intelligence fabric with different views into it. A canvas is not a separate app with its own database. It is a region of the same shared fabric, and an object that appears on one canvas is the same object everywhere it appears. Four surfaces, each answering one question, and none of them is Ghost. Ghost is what sits underneath and remembers.

• Zignal: what is changing outside the company. the market, the competitors, the sentiment, the research, the regulation, and whether you show up at all when a model answers a question about your category
• Hunt: where to act. which organizations, which people, what just moved, and the evidence behind the ranking
• Oppo: whether you are becoming what you said you would become. what got proposed, budgeted, committed and scoped, against what was actually spent, delivered and earned. It runs against my own proposals and decks today, and it is going into its first partner deployment now
• Craft: what to make. the campaign, the page, the message, the outreach, opened on something worth making rather than on a blank prompt
• Ask Ghost is the doorway rather than a fifth surface. You ask, and Ghost composes the answer from whichever canvases it takes

Drift: are reality and intention separating?

A finance tool can tell you margin fell. A proposal tool can tell you what you sold. A project tool can tell you delivery took fifty-one days. None of them knows those three facts are about each other. Drift is the question Ghost can ask because it holds all three: is what we are doing still what we said we would do?

• Promise drift: what was sold against what was delivered
• Financial drift: the forecast against the actuals, and where the margin actually went
• Scope and timeline drift: contracted against performed, expected against elapsed
• Strategy drift: the declared priority against where the hours and the money really went
• Market drift: an assumption the business was built on against what is observable now
• Evidence drift: something believed that is going stale or being contradicted
• Message drift: what is being said in the field against what is actually approved

What that sounds like

Take a proposal: a fixed implementation fee, a thirty day launch, a monthly retainer, an assumed twenty hours of support a month. Ghost reads those as commitments rather than as sentences in a PDF. Later the accounting system reports the hours the implementation really took, the project record reports the day it launched, and the support log reports the hours actually spent.

• The useful thing to say on the next proposal is not that margin is down
• It is that the last three of these took closer to fifty days than thirty, at roughly twice the implementation effort this one assumes
• That is a sentence a founder understands immediately, and no single tool in the stack can produce it

Everyone touches the loop. Nobody closes it.

This is the honest read on why the layer is still open. Plenty of good products own a piece of it, and each one stops before the piece that costs something.

• Model vendors remember, inside their own silo, and that memory exists to keep you on their model rather than to outlive it
• Agent platforms act. Almost none of them grade what the acting produced
• Memory and retrieval tools store and fetch, and never do anything with it
• Consultancies genuinely improve you, and then the judgment walks out the door at the end of the engagement
• Closing it means all four at once: remembering, acting, improving, and the record belonging to you

The rules we build by

The interface is calm, spatial, and deliberate, and a few laws are not negotiable.

• The user does not draw the canvas. You ask, Ghost draws
• Answers become objects, not chat bubbles. Chat is transient; the canvas remembers
• Every belief is traceable: source, date, evidence, confidence, and when it was last verified
• Ghost proposes actions. A human approves the consequential ones, and the outcome becomes learning

It does not get installed. It gets taught.

This is the part nobody can buy off a shelf, and it comes before everything above it.

• Every company already knows what predicts a win. It sits in a handful of people, not in a database.
• A system with nothing to learn from on day one is fluent, fast, and confidently generic.
• So the extraction is the work, and it happens in a room before a single agent runs. The long version is Alignment is the analysis.

Straight talk

Every mechanism below is built. The line is between what is already running on real work and what is wired up waiting for outcomes to land in it.

• Running: every claim carries its own source and how it was captured
• Running: signals scored and explained well enough that a person can argue with them
• Running: operator judgment is kept as a first-class record rather than a comment, and it is only starting to fill up
• Running: nothing consequential sends without a human yes, enforced in code and not in a setting
• Wired and waiting: outcomes at volume. The cases are open and they are closing, and the day enough of them land the scoring stops being an informed guess
• Next: the lift. Corrections are captured and kept, and wiring them back into the scoring is the work in front of me rather than something already sitting there
• By design, and staying that way: nothing learns across companies. Your record sharpens your system and nothing else. That is a wall, not a gap in the roadmap

AI does not get the keys

This is the first question anyone asks, so it gets answered before they ask it. Ghost runs beside the systems it reads, never inside them. Read-only by default, no autonomous writes, and a separate environment per deployment with nothing mixing between them.

• You own your data and your ledger. Exportable, portable, and yours on the way out
• We own the engine and the method, and never anyone else’s evidence or results
• Every recommendation shows its evidence, its confidence, and who approved it