- What we claimed
- That the constraint model and its checker shared "no module, helper or constant".
- What was actually true
- Computing the artifact hashes to publish them showed both live in ONE file, with an identical sha256, sharing the input data and the constraint parameters.
- What changed as a result
- Two rows moved from "not shared" to "shared". The conclusion — that no such roster exists — is unaffected, because it rests on a pigeonhole argument verified by hand. What was wrong was the strength of the independence claim, not the result.
Company
We build software for answers that have to be right.
ProServIQ is an independent engineering practice, founded in 2026 and working with clients anywhere. We take on problems where a wrong answer costs money, a signature or a reputation, and we build every solution with a second, independent check beside it. This page is who we are, the rules we work by, and the dated record of where we have been wrong.
- Founded
- 2026
- Works with
- Clients anywhere
- Practice
- Product and verification engineering
- Products
- CLAVIS, NEXIO live · CIPHER in build
- Engagements
- Fixed fee · you own the source
- Checks published
- 7, in the register
The name
Pro · Serv · IQ
Three syllables, and each one is a commitment we can be held to.
- Pro
Work someone puts their name to
Professional in the sense a signature implies. Every engagement ends with a written argument for why the answer is right — not only a system that produces one — so the work can be defended by whoever inherits it.
- Serv
Your problem first, our products second
We start from what you need, not from something we want to sell. Where one of our products fits, we say so; where none does, we build what does. What we build for you is yours, as source, and we keep nothing back when the work ends.
- IQ
Intelligence that can be checked
Not a model’s confidence. A second computation of the same answer, written separately so that it could disagree — and a result that says established, divergent or unestablished rather than a confident guess.
What we do
Three kinds of work, one way of doing it
Every piece of work we take on, whoever it is for, ships with a second instrument that can disagree with it.
Products we build and run
- CLAVISLiveStatutory income tax (India)
- NEXIOLiveAgent output, checked against the world
- CIPHERIn buildCausal measurement over high-frequency time series
Work we do for clients
We build the system you need with its checker included, build only the checker for a system you already run, or put a gate in front of what you store. Each is scoped in writing and priced as a fixed fee.
The engagements →Checks we publish
7 so far, numbered and dated — most of them of systems we did not build, and several of our own work found wrong. Each can be re-run from its published method without asking us anything.
The register →03 · Corrections
Where we have been wrong, dated
A practice arguing that systems should be checked by something able to disagree with them has to publish the occasions its own work was found wrong. An empty section here would mean the checking had stopped.
- What we claimed
- Time to prove infeasibility: 0.049 s.
- What was actually true
- That figure came from a different machine and had been copied into prose rather than read from the run. The correct figure on the stated environment is 0.089 s.
- What changed as a result
- The page now renders every timing from the generated state file rather than from written copy, so the same class of error cannot recur silently.
- What we claimed
- That an SMTP 250 response establishes a mailbox exists and a 550 establishes it does not.
- What was actually true
- RFC 5321 does not support either reading. A 250 means the server accepted the forward-path for delivery; a 550 can be policy rather than absence; and a server may accept and reject later.
- What changed as a result
- Restated as accepted, rejected, and indeterminate — which is the three-state shape the rest of the site argues for, and which the original copy had itself flattened to two.
Two of the three above were found by the act of preparing evidence for publication — computing a hash, reading a figure back off a run. That is not a coincidence, and it is the argument for publishing at all.
Standards
Four rules that decide arguments here
Not values. Rules — each one has settled a real disagreement about what to build or what to publish.
01
Three states, never two
Every check reports established, divergent, or unestablished. The third is not a failure to reach an answer — it is an answer, and it is the one most systems throw away. A confident zero where the honest report is "could not be established" is the defect this practice exists to name.
02
Independence is a claim, so it carries evidence
Saying two instruments are independent is worth nothing on its own. Every check we publish carries a table of what the two share — module, constants, inputs, and the written rules themselves — because a disagreement only proves what the shared parts allow it to prove.
03
A refusal names what would change it
A system that declines without saying what was missing is an outage with better manners. Every refusal we build states the fact it could not establish and what would establish it.
04
The measurement is checked before the finding is
A harness that measures the wrong thing returns a confident number, and a confident number stops the search. We have caught eight of our own doing exactly that. Check 0027 is the record.
Rule 01 is the one the other three rest on. The plate shows what a two-state report does with a result nobody could establish: it has nowhere to put it, so it quietly calls it one of the other two.
The boundary
What we will not claim
These are not disclaimers. They are the edges of what this practice is competent and licensed to do, and each has already stopped a piece of work.
We do not represent taxpayers, or anyone else, before any authority.
Representation is professional work under the relevant statutes. We are not registered for it, and building software adjacent to a profession does not confer its licence.
We do not rank investments or assess their suitability.
That is investment advice and requires registration to give. Our instruments state facts about statutory instruments and compute their effect exactly; they do not tell anyone what to buy.
We do not publish a check we have not re-run from the method as written.
A result that cannot be reproduced from the published method is a claim, not a check. Where we hold material we cannot yet publish in full, the register says so rather than presenting a summary as an entry.
We do not describe a finding about a third party until it has been put to them.
Check 0030 concerns a computation we did not build and is held unpublished for exactly this reason. The register states its state rather than omitting it.
Work with us
The quickest way to find out whether we are the right people is one conversation about your problem. It is free, and it ends with a plain answer.
Found something wrong here?
A correction to anything on this site is the most useful thing you can send us, and it goes into the register above under its own date.
contact@proserviq.com