The system of record for data decisions

AI is only as trustworthy as the data beneath it.

Threshold governs your data before your AI ever sees it, and keeps a permanent, queryable record of every decision that made it trustworthy. Which rule ran. Who approved the exception. What it looked like when it arrived.

Runs before your AI. Never runs your AI.

unit_cost· part CK‑008 run #412
$14.20 Why this number is this number
  1. Arrived $14.20, unchanged from source batch 2026‑09‑24 08:12 · northmark_parts
  2. Mapped Unit Cost → unit_cost declared mapping #12 · applies to 3 sources
  3. Checked Within the declared range rule #47 · validated 212 times · high confidence
  4. Flagged Up 62% against run #411 threshold 20% · surfaced without being asked
  5. Ruled Stands. Supplier repriced in September, confirmed against the signed contract. M. Reyes · 08:41 · in their own words
Recorded permanently Queryable by your AI
01Our stance

We run before your AI. We never run your AI.

Threshold ships knowledge access and grounding. It does not ship a model. Your organization connects its own AI, under its own credentials, inside its own security perimeter.

Your model Copilot, Claude, Gemini, or whatever you have already approved. We don't ask you to adopt a new one.
Your keys Inference runs on your credentials, billed to you, logged by you. Threshold never holds a provider key.
Your data Grounding happens through a read-only endpoint your AI calls. Your rows stay where they are.
Enforced We removed a shipped feature to keep this true. A constraint you won't hold yourself to isn't one.
02The gap

Your dashboard says $4.2M. Nobody can tell you why.

The pipeline ran. The number rendered. But the decisions that produced it live in scripts, spreadsheets, and one analyst's memory: which rows were corrected, which rule applied, who approved the exception. Ask why the number is what it is and you don't get an answer. You get a reconstruction.

BI and visualization

Renders the number beautifully. Cannot tell you which corrections stand behind it, or who signed off.

No decision trail

Warehouses and pipelines

Move and transform at scale. Capture the code that ran, never the human judgment that decided what correct means.

No judgment record

A chat tool and a spreadsheet

Answers fluently from whatever you hand it. Has no way to know whether the data was right to begin with.

No source of authority

You cannot audit a number. You can only audit the decisions that made it.

The premise Threshold is built on
03What you actually get

Four surfaces. One record.

Threshold is not a dashboard you check. It is where the work happens, and the record is what the work leaves behind.

A review queue

Everything needing a human call, in one place, with the context to make it. A cross-source conflict shows you both values and where each came from.

Decide it

A rules library

Every check your team has ever agreed on, who wrote it, when, and how many times it has held up since. Confidence you can see rather than assume.

Reuse it

A findings feed

What moved since last time, raised on arrival, before anyone thought to ask. Each one carrying what it was compared against and how sure that comparison is.

Catch it

A knowledge endpoint

The whole record, open to the AI your organization already uses. Read-only, provenance attached, no new chat tool to adopt.

Ask it
04One scoring language

One yardstick.

Four dimensions, one vocabulary, applied to every dataset that boards. That is what makes two clients comparable at all. A supplier file and a payroll extract get held to the same named standard, so "better" and "worse" finally mean something.

northmark_parts run #412 94
Completeness 96
Validity 88
Distribution 93
Uniqueness 100
Shape Pre-flight gate · passed Structure is not scored. It either matches what this source has always looked like, or the run stops before a single rule gets to touch the data.
05Safety architecture

Seven things Threshold will not do.

A system that governs data earns trust by what it refuses to do on its own. These seven are architectural, not policy. Each is a property of how the system is built, not a rule somebody has to remember.

01 Containment 02 Bounded authority 03 Human oversight 04 Observability 05 Immutability 06 Fail‑safe 07 Isolation
01
Containment

Nothing reaches the outside without permission.

Every external source Threshold may contact is named in advance. An address that isn't on the list is refused before the request is made, not after.

Refused No allowlist configured means an empty allowlist, never an open one.

02
Bounded authority

It never guesses what your data means.

Two fields become the same concept when a person says they are. A resemblance between column names is a tiebreak at most, never grounds to act.

Undeclared No fuzzy search and no similarity discovery exists anywhere in the system.

03
Human oversight

A person made every call.

Threshold proposes, a reviewer decides. Nothing is applied automatically at any confidence level, however many times a rule has proved itself. The reason gets recorded in the reviewer's own words, never picked from a dropdown, because the categories are exactly what gets lost.

No such setting Auto-apply is not a switch that defaults to off.

04
Observability

It tells you what it couldn't see.

Every check reports its own blind spots by name: the field that wasn't present, the comparison group too small to be fair, the source that couldn't be reached.

Not evaluated A quiet report is never mistaken for a clean bill of health.

05
Immutability

It can read your numbers. It can't rewrite them.

Threshold watches your output and reports what it finds. Observing carries no power to change what was observed, and what arrived is kept exactly as it arrived, so "what did this look like before we touched it" always has an answer.

Append-only There is no edit path and no delete path to build one from.

06
Fail-safe

It degrades loudly, never quietly.

When a source is unreachable, Threshold says so and carries on with what it can still check. Stale figures are labelled with their age rather than passed off as current. A failure in the watching layer can never delay or corrupt the governed work underneath it, and it never invents a result to fill the hole.

4 more runs needed Not enough history to judge means exactly that, not a number dressed up as one.

07
Isolation

Knowledge travels between clients. Data never does.

What one engagement teaches the next is a rule or a pattern, never a row, never a value, never a figure traceable to anyone. Comparisons across a book of clients ship as medians and distributions, or they don't ship.

Out of scope Cross-client access to raw data isn't deferred to a later release. It's absent.

+
And one that follows from them

Quiet the noise. You can't switch off the watch.

Silence an alert you have already dealt with, giving a reason and a date the silence expires. The check keeps running underneath, every time, and a materially worse recurrence comes back through anyway.

Still detecting Quieting suppresses the notification, never the detection. There is no mute.

Prefer to see this on your own data?

Bring one messy source and one question you keep answering from memory. We will run it and show you the record it produces.

Book a demo
06How it works

Three stages. Every source, every time.

Threshold sits at the point data enters your organization. Each stage produces trustworthy output and, just as importantly, the record of how that output came to be.

1Per source

Clean

One source at a time, through nine quality phases. Structural drift is caught before a single rule runs.

  • Schema drift caught pre-flight
  • Rules carry a confidence history
  • Reference tables drive validity
  • Row-level audit on every pass
2Per use case

Transform

Where the modeling happens. Joins, rollups and business logic become recorded decisions instead of a script in someone's folder.

  • Named operations, never a formula box
  • Existing Power Query logic imported
  • Governed inputs only
  • Lineage back to the exact run
3Across everything

Insight

The accumulated record, put to work. Threshold compares each arrival against what it already knows and raises what doesn't fit.

  • Checks fire on arrival, unprompted
  • Every finding carries its provenance
  • Benchmarks from your own book
  • Open to your AI over MCP
07Why it compounds

The fifth client is easier than the first.

Every other tool in the stack works inside one organization at a time, so what it learns at one client stays there. Threshold carries the lesson forward, which means the more data passes through it, the more it already knows.

01

Learn it once

Run Northmark's parts data. Build the rules, resolve the exceptions, record the reasoning. Ordinary work, captured instead of discarded.

02

Name the concept

Northmark calls it Part Number. Cedarline calls it SKU_name. Someone declares both to mean part_id, once.

03

Arrive already knowing

Cedarline's first run opens with Northmark's rules proposed at high confidence. You know most of what to look for before you have opened the files.

08Who it's for

Built for whoever handles everyone else's data.

Threshold pays off fastest where the same kind of work repeats across many sets of books, because that is where captured knowledge stops being a nice idea and starts being an asset.

Consultancies and SIs

Your modeling stops dying at the end of each engagement.

The cleaning logic, the joins, the business rules currently living in Power Query and one person's head become something the firm owns and brings to the next client.

Accounting and audit

Every judgment stays defensible years later.

When somebody asks why a figure was treated the way it was, the reasoning is attached to it, in the words of whoever made the call, with the date they made it.

Portfolio operations

One standard across every company in the book.

When two holdings report the same metric differently, you get a straight answer about which definitions diverged instead of a month of reconciliation.

In-house data teams

Institutional memory that does not resign.

What your longest-tenured analyst knows about why the numbers behave the way they do becomes something the team holds rather than something you hope stays.

09Connect your AI

Give the AI you already use a memory.

Point your organization's assistant at Threshold's endpoint and it can answer questions that have never had an answer before, grounded in the record, with provenance attached to every claim it makes.

Why did this part's unit cost change in the September run?

Who approved the correction on the revenue field, and what was their reasoning?

What is different about this client's data since the last delivery?

Every tool is read-only. Your AI can ask. It cannot change anything.

Knowledge endpoint · MCP
READget_rule_biography
READget_field_history
READget_knowledge_ledger
READget_benchmark
READlist_findings
Microsoft CopilotClaudeGeminiAny MCP client
Get in touch

See what Threshold knows about your data.

Tell us where the trail goes cold today: the number nobody can explain, the cleanup that gets redone every month, the client question you answer from memory. We will show you what it looks like recorded.