Hindsense Hindsense

hindsight, engineered

Catch your AI’s mistakes before your client does.

For consultants, engineers and scientists who sign their work. Hindsense plugs into your AI assistant and catches invented citations, guessed names and off-target answers before they reach a deliverable.

The citation checker works today. The full suite is on its way, and we’ll send you one note when it’s ready.

You asked one question. It answered a nearby one.

You ask about the risk to one receptor under one set of conditions, and the AI gives you a thorough answer to a broader question. It reads well, so you only notice on the third read, or your reviewer notices for you. Hindsense sits between the assistant and your deliverable and catches that kind of mistake while it’s still cheap to fix.

The four mistakes it’s built to catch
  • The nearby question. The AI answers a broader version of what you asked. It sounds thorough and misses the point.
  • The confident guess. A limit, clause number, role or address is stated without hesitation, and some of it is invented.
  • The default-safe conclusion. The AI hedges to the most conservative reading. On your work, needless caution still has to be justified.
  • The draft you can’t defend. The prose is competent, but ask what a paragraph rests on and nothing is there to show a reviewer or an auditor.

Keep your assistant. Hindsense works behind it.

Use Claude, ChatGPT, Copilot or whichever assistant you already trust. Hindsense checks your request before the AI starts work and checks the draft before it’s handed back to you. Findings come back in the same conversation, so there’s no new window and nothing to paste.

How it connects

Hindsense works with any assistant that can call an external skill. It runs when you ask, or when the assistant is about to hand you a finished draft, and it answers with a verdict and a reason in the same conversation. Behind the scenes it is a JSON API: the assistant sends the material to check, and Hindsense sends back its findings.

What it checks

  • Getting the question right. It sharpens a vague request into a clear objective before any work starts.
  • Sources and facts. It blocks citations, clauses and names that nothing actually supports.
  • Drafts that read like your work. It puts the answer first, flags needless caution, and raises problems for you to decide rather than rewriting them quietly.
  • Files and records that hold up. Word files that open in Word, references that don’t break, and decisions that a later session can’t quietly reverse.
See every check, including the environmental consulting pack

Objective and evidence

  • Sharpens a vague prompt into a real objective before the AI runs on it
  • Refuses a citation whose source couldn’t support that kind of claim (live)
  • Blocks a standard, statute or guideline reference with nothing behind it (live)
  • Won’t fill in a name, role or contact detail that was never confirmed (live)
  • Points you to the original source rather than an article summarising it
  • Checks that units, datums and reference conditions match before a number is compared to a limit
  • Re-derives a pass or fail rather than trusting the one already typed into the table

Drafting that reads like your work

  • Answers the question first, then supports it
  • Checks factual claims from first principles rather than pattern-matching a familiar answer
  • Flags an unnecessarily cautious conclusion
  • Raises a substantive problem for your decision rather than quietly rewriting what you meant
  • Catches passages repeated across sections, and content sitting in the wrong section
  • Tests that every commitment in a plan names a trigger, an action, an owner, a timeframe and a record
  • Writes for somebody who already knows their field

Files that survive the round trip

  • Checks a Word file actually opens in Word
  • Repairs cross-references so section, table and figure numbers update
  • Keeps a spreadsheet’s inputs, workings and results separate, with consistent units
  • Builds photo records with captions that describe what’s there

A project record that holds up

  • Keeps every document you supplied, so you’re never asked for it twice
  • Treats the version you edited and sent back as the one that governs
  • Records decisions so a later session can’t quietly reverse them
  • Keeps ideas still being explored separate from conclusions you can act on
  • Distinguishes what has been built from what is live
  • Produces the AI-declaration record your quality system asks for at sign-off

Environmental consulting pack (in development)

  • Reviews underground service plans before intrusive works
  • Drafts regional geology and hydrogeology sections a site owner can picture
  • Interprets soil and groundwater results against the right assessment criteria
  • Assembles site inspection photo records

Work in a different field? Tell us in the waitlist question.

What you get back

A clear verdict and the reason for it: pass, pass with flags, or refuse. Each finding points to the claim it concerns and says what’s missing, for example:

Refused. A clause-level claim has no source. Add one from the legislation register, the standards body or the regulator’s own guidance.
Want to see the raw API response?
POST /v1/skills/citation-discipline/invoke 
"verdict": "refuse",
"findings": [
  {
    "target_id": "c1",
    "status":    "verified",
    "severity":  "info",
    "message":   "Verified against provided source structure.
                 Reconfirm the exact clause number in the
                 in-force version before signing."
  },
  {
    "target_id": "c2",
    "status":    "unsupported",
    "severity":  "blocker",
    "message":   "No source supplied for a clause-level claim.
                 Preferred publishers: state or commonwealth
                 legislation registers, the standards body
                 itself, or the regulator's own guidance page."
  },
  {
    "target_id": "c3",
    "status":    "named-but-flagged",
    "severity":  "warning",
    "message":   "Source is present but not from the preferred
                 publisher for this claim kind."
  },
  {
    "target_id": "s1",
    "status":    "named-but-flagged",
    "severity":  "warning",
    "message":   "Internal specific supplied without a verification
                 channel. Do not expand initials, usernames, or
                 email local-parts to a full name unless verified."
  }
],
"entitlement": { "ruleset_version": "2026.08", "update_available": false }

Why it exists

Ask an AI how many days of the week contain the letter “d” and it may say three. It recognises a familiar puzzle and reaches for a familiar answer instead of reading the question. Hindsense started as a fix for that habit, and every check since has come from a real mistake in real work.

Read how this started

You stay in charge

Hindsense flags problems; you make the call. It won’t sign your work or stand in for your review.

What it does and doesn’t do
  • It doesn’t replace your judgement. It surfaces findings for your decision and doesn’t sign the deliverable.
  • It doesn’t read the source end to end yet. The live checker confirms the source you supplied is plausible and from the right kind of publisher for the claim. Deeper retrieval comes later.
  • It doesn’t guess at people. A name or role that can’t be tied to a verification channel stays flagged.
  • The full product isn’t released yet. The checks are built and tested. Accounts, keys and the wider suite are still in development, and there’s no release date yet.
Common questions
Does AI actually make things up in professional reports?

Yes. Language models generate plausible-looking citations, standards references, clause numbers, dates, addresses and names that do not exist, and they do so with confident wording. The failure is systematic, not a bug awaiting a fix. Which is why sitting a guardrail between the assistant and your deliverable is the fix, not switching models.

Is this different from just asking the AI to be careful?

Yes. An assistant told to "always cite your sources" will happily cite a fabricated one. Told to "think carefully" it will produce longer prose in the same shape. The Hindsense guardrails are external checks with their own logic. They interrogate the prompt before the AI runs, and evaluate the draft against tested standards after, rather than trusting the AI to police itself.

Which AI does it work with?

Any assistant that can call an external skill file. The engine is a JSON API (the assistant sends the material to check, the engine returns findings). The same guardrails have been used against several frontier models in real work.

Is this shipping today?

The guardrails exist, they have been tested against dozens of real jobs, and they work. The citation-discipline engine is live behind a stable JSON contract. What is being built is the product wrapper (accounts, keys, the way the rest of the suite is packaged and distributed). That is what the waitlist is for.

Who is it for?

Professionals whose work is judged on whether the outcome is right and defendable (consultants, engineers, scientists, analysts, planners, technical specialists). If you use AI to research, draft or review, and a wrong answer costs you real time or real credibility, the waitlist is aimed at you.

What happens to my draft when the engine checks it?

The engine evaluates the material in the request and does not retain the draft after the response. Waitlist signups store your email address for one release notification, as set out in the privacy notice. No marketing platform, no sharing, no sale.

How this began

I wanted the AI to care when it got something wrong.

As a graduate, early in my career, I made a mistake at work, and I can still remember exactly how it felt. I was furious with myself and convinced I was about to be fired. The error was fixed and my career survived, but the feeling never left, and it became the standard I have held myself to ever since: check the detail, test the reasoning, and read it once more before it leaves your hands.

Years later I started using AI in my work, and one thing frustrated me more than any of its limitations. It could make a confident mistake and then, when I pointed it out, reply with some version of “you’re right to call that out, I missed that” before moving straight on. The correction was welcome, but the mistake appeared to cost it nothing, and it seemed no more determined to catch the next one before I did.

So I asked a slightly strange question: could I make an AI care about the quality of its work, and hate getting a detail wrong? The first attempt failed almost immediately. Asked how many days of the week contain the letter “d”, it answered three, because it had recognised a familiar puzzle and reached for a familiar answer instead of reading the question. That small failure taught me the real lesson: telling an AI to be careful is not enough; care has to be built into a process that checks from first principles, verifies the source and rereads the work before handing it over.

One instruction grew into a set of guardrails for citations, units, prompts and review, each born from a real failure and tightened whenever it missed something. Hindsense is where that led. It does not ask you to trust the model more; it gives the model a disciplined way to check its own work, then keeps testing whether that discipline holds. Because “you’re right, I missed that” is useful after an error, but it is far better not to miss it in the first place.

Hindsight, before you need it.

One note when the first release is ready. Nothing before then.

Join the waitlist