Skip to content
Wonder Constant Labs
Research
Project LatticeActive

What survives when information becomes data

Now active through two live applied cases: how much of the meaning in a body of information survives translation into something computable, and can that loss be measured rather than assumed?

Research question

Turning information into data is never lossless. Can we characterise what a given representation keeps and what it discards — and put a number on the difference?

Overview

A conversation becomes a transcript. A transcript becomes tokens. Tokens become vectors. Something is discarded at every step, usually silently, and the analysis downstream has no way to know what it lost.

Lattice began as a question rather than a programme of work. Before committing effort we wanted to know whether the loss introduced by a representation can be characterised in a way that generalises — or whether it is only ever answerable case by case.

That question now has a concrete system to study rather than only a synthetic example: the RECON Intelligence System™, a live commercial product Wonder Constant Labs builds and operates from a founder-created system that predates the company. RECON compresses six dimensions of public-business evidence into a score, then restores explanation and action around it. It is the initiative's first real-world case.

A second real-world case sits alongside it, in a different medium: Claude Anvil, an internal instrument for specifying an engineering goal as a bounded work order, auditing what comes back against evidence rather than narration, and keeping what is learned. It is the same question Lattice asks about RECON, asked instead of an owner's intent as it is translated into a structured work order, an implementation, and a completion report — what survives that translation, and what has to be checked rather than assumed.

Approach

  • Sketch small, well-understood cases where the intended meaning is known by construction, so representation loss can be quantified rather than debated.
  • Survey where existing fields already answer this — information theory, compression, measurement theory, linguistics — before inventing anything.
  • Use RECON Intelligence System™, running under live product conditions, as a real-world case for locating where representation loss happens and whether useful context can be restored.
  • Use Claude Anvil, an internal work-order and audit instrument, as a second real-world case: locating where a returned implementation's claims diverge from what can actually be verified, and whether that gap can be caught before being accepted as done.
  • Decide honestly whether the question is tractable at a useful scale, and close the initiative if it is not.

Instruments

  • Claude Anvil

    Available

    An internal instrument that turns an engineering goal into a bounded, verifiable work order, audits the returned implementation against evidence rather than narration, and keeps a small ledger of reusable lessons. Active and in daily internal use: every work order it drafts and every audit it runs is a real operational result. Whether that discipline measurably improves representation fidelity is a separate, still-unpublished research question — not released as software.

Commercial product used as an applied case

A real-world case, not a finding — the product below is live commercial software from Wonder Constant Labs, doing the kind of representation compression this initiative is asking about under real operating conditions.

  • RECON Intelligence System™

    Project Lattice asks what survives when information becomes data. RECON is a commercial instance of exactly that problem: it has to decide which facts become variables, how unlike evidence gets normalised onto a common scale, what context survives when that evidence is compressed into a single number, and how explanatory and operational context can be restored afterward so the number is still trustworthy to act on.

    The system underneath RECON predates both Wonder Constant Labs and this initiative. Its current customer-facing product was redesigned and commercialised by the company. This initiative did not produce it.

    Read the product profile →Visit RECON ↗

Log

  1. June 10, 2026 · Concept logged

    Opened as an exploratory concept. Reading and scoping; no commitment to a full initiative.

  2. August 2, 2026 · Applied case: RECON Intelligence System™

    Wonder Constant Labs brought the founder's pre-existing RECON system into the company as Lattice's first real-world applied case. It already performed the kind of evidence-to-score compression the initiative asks about. The initiative moved from exploratory to in development because it had a concrete system to instrument, not because a result had been produced.

  3. August 23, 2026 · Second applied case: Claude Anvil — status moves to active

    Claude Anvil — an internal instrument for specifying, auditing, and learning from engineering work orders — joins RECON as Lattice's second live case: owner intent translated into a structured work order, an implementation, and a completion report, with its own translation loss to locate. Anvil is active and in daily internal use, producing a real operational result — a drafted work order, an audit, a recorded lesson — with every cycle it runs. The initiative moves from in development to active on the strength of that daily operational use, not because a research finding about representation fidelity has been produced.

  4. September 7, 2026 · RECON's redesigned commercial product recorded

    The current RECON product is now built and operated by Wonder Constant Labs through two public, self-serve experiences: one for business owners seeking a clear view of their own online presence, and one for professional teams evaluating local businesses for work. Owners use Score, Dossier, and Technical; Map and Scheduler extend the system for teams working across many businesses. A broader applied case is still a case, not a finding.

Known limitations

Stated before anyone has to ask.

  • Claude Anvil has not been evaluated by anyone outside its own use, and there is no controlled comparison against work orders written without it.
  • Nothing about how much either applied case — RECON or Claude Anvil — actually reduces representation loss has been measured, characterised, or reported.

Findings to date

No research finding published. Both applied cases are active and produce real operational output day to day — RECON's live scoring, Anvil's drafted work orders and audits — but operational output is not a research finding: nothing about how much either case actually reduces representation loss has been measured, characterised, or reported.