Valistry

Ask a question. The ledger answers.

Spend, coverage, unit cost, a spike or a saving, in plain language. The plan that produced the answer travels with it, and what the ledger cannot answer is said rather than guessed.

  • Never invents a number
  • The plan travels with the answer
  • Unknown stays unknown
Ask Valistrysample org

Why did spend spike on the 21st?

  • Draft reply agent$349
  • Unallocated$39
  • Ticket triage$7
Open Spend Explorer

Which agents cost more without improving outcomes?

  • Draft-reply agent$2.91
  • Triage agent$0.19
  • Unallocated
Open Unit Economics
Ask about spend, coverage, unit cost, a spike or a saving⌘K

Why Ask Valistry

The model writes the sentence. It never works out the number.

Two of the three layers a question passes through never run a model, and the one that does is given the result and nothing else.

A question goes in as a sentence, and comes back as an answer with the plan that produced it.

Read · deterministic

The question becomes a plan

One measure, a grouping, up to three filters and a window, validated before anything runs. No model reads your question.

Run · deterministic

The plan becomes figures

The plan runs over the reconciled ledger in integer micros. Every figure, row and piece of evidence in the answer comes from this step.

Write · a model

The figures become a sentence

A model is handed the result and nothing else, and may only re-phrase it. What it writes is checked against the figures before you see it.

  • 0 invented numbers

    Computed, not generated

    Every figure in an answer exists in the result the query returned. A sentence that introduces one is refused before you see it.

  • 2 of 3 layers

    Run without a model

    Reading the question into a plan and running that plan over the ledger are both code with tests behind them.

  • 1 plan per answer

    Traceable

    The plan, the window and the records travel with the answer, so a wrong answer is traced to a query and not to a mood.

The guardrails

What it will never do.

Each of these is a rule in the code, not a line in a prompt. A sentence that breaks one is dropped and the computed sentence stands in its place.

  1. Never compute

    Work out a figure of its own.

    Every number is arithmetic in integer micros, checked by tests. The model is handed the result and may only re-phrase it.

  2. Never conclude

    Say what the figures do not.

    Coverage is not called healthy, a window is not called representative, and a ratio the engine did not compute is dropped.

  3. Never fill a gap

    Guess at what is missing.

    A period with nothing before it is refused rather than compared. An unmeasured outcome is named, never estimated.

  4. Never reach

    Read outside the asking organization.

    Every query is scoped by tenant in the store's own interface, and no prompt or response content is read at all.

Where it belongs

An answer is only as good as what it is standing on.

Every cost tool now has an assistant. The question is what happens to the plan, the evidence and the next step once the sentence is written.

Handed a copy

An assistant that has been given your numbers can write a good sentence about them. Three things do not come with it.

  • The plan stays behind, so a figure that looks wrong cannot be traced to the query that made it.
  • The evidence class stays behind, so a rule reads like a reading and an unknown reads like a zero.
  • The next step is in another tool, so the person who has the answer is not the person who can act on it.

Standing on the ledger

Ask Valistry is the ledger answering, on the screen where the decision is made.

  • The plan travels with the answer: the measure, the grouping, the window and the records it ran over.
  • The evidence class travels with the number, and what the ledger cannot answer is said rather than filled in.
  • Every answer ends in the screen that acts on it: the explorer, the scenario, the rule or the policy.

The same records make the monthly report the finance leader signs, so the answer and the report can never disagree.

Move from dashboard hunting to decision-making.

A 30-minute demo on your own estate, or connect one source and ask the first question.

Every phase writes to the one ledger this answers from, which is the whole reason the answer holds.

Next: the sources it reads