Skip to content

The Mass

The essay calls the product one living mass: governed truth and the bonded human–AI pair working it. That phrase is compressed enough to hide the invention.

The mass is a truth machine in operation. That is a functional description, not a new component beside the stone, the pair, or the rails. It is the whole loop by which partial and conflicting observations enter as evidence, meet the current record under scoped authority, become a reviewed change, and advance what the product may responsibly treat as true. The pair inhabits that loop. The rails govern it. Facets are cuts through which selected results leave it and new contributions return.

This is why the truth machine cannot be treated as infrastructure underneath Facet. Without it, presence produces fluent improvisation and thin surfaces have nothing trustworthy to show. It is what the mass does.

Truth here is operational, not metaphysical. It means the current best reviewed account on which the people around this stone may responsibly act. It can be incomplete. It can carry a conflict. It can say that a date is unknown, that two sources disagree, or that a conclusion remains tentative. It can later be superseded without pretending the earlier judgment never happened.

The machine earns the word truth precisely by refusing easier substitutes for it:

  • a source saying something does not make it true;
  • a value being stored in an important system does not make it universally authoritative;
  • a model extracting or inferring something does not make it reviewed;
  • approval of a change does not erase the evidence or uncertainty under it; and
  • the latest current record does not rewrite what was previously believed or communicated.

A conventional record often collapses all of these into one mutable value. The truth machine keeps them apart and governs the transitions between them. Its claim is not that the current account is infallible. Its claim is that the account is attributable, corrigible, and honest about why it presently stands.

The smallest useful picture is a cycle, not a stack:

sources and events in the world
immutable evidence · provenance · bounded source authority
pair reconciles against current truth at revision N
coherent proposed delta · cited evidence · remaining uncertainty
authority review + deterministic validation + exact expected revision
▼ atomic transition
current truth at revision N+1 + append-only change record
facet or frozen publication
answer · correction · acknowledgement · new observation
└──────────────────────────────► evidence

Nothing jumps from arrival to current truth. Nothing edits current truth from a facet. Nothing sent outward is reconstructed later from whatever happens to be current then. The cycle has one admission path and one transition boundary, regardless of how many sources and facets surround it.

The loop requires several meanings that ordinary applications routinely blur. Their storage layout is a product decision; their semantic separation is not.

Record The question it answers Governing rule
Evidence What did this source actually say or show? Preserve the original and its provenance; extraction is derivative; correction adds evidence rather than altering it.
Contribution What did a person outside the bond submit? Retain it exactly as received and attribute it; receipt does not make its claims current truth.
Current truth What may the product responsibly act on now? Change it only through a reviewed transition; carry explicit uncertainty and field-level ancestry.
Change What moved, from what to what, why, and under whose authority? Append the exact delta, cited evidence, rationale, approval, expected revision, and remaining uncertainty.
Publication What exactly crossed the boundary at this encounter? Freeze the outward artifact; keep drafted, approved, sent, acknowledged, and incorporated states distinct.

These are not five databases. They are five promises made by one mass. A Git repository, a database, an object store, or another durable substrate can keep them; none of those technologies supplies the discipline by itself.

Admission is not promotion. An email, form answer, spreadsheet, ERP row, call note, screenshot, or model-produced extraction enters as a claim from somewhere. The original survives byte for byte when the medium allows it. Its source, receipt, integrity, and extraction status travel with it. Failed extraction preserves the source rather than converting a tooling limit into lost reality.

Authority attaches to the claim at the narrowest useful scope. A customer may be authoritative about what they requested and not about what the shop can produce. A production system may be authoritative about what was posted and not about an event nobody posted. A person may be authoritative about a direct observation and not about a consequence outside their role. Calling something a system of record does not dissolve these boundaries; it names one source whose authority still has edges.

This is the first general move of the truth machine: preserve what arrived, then ask what kind of claim it is allowed to support.

Reconciliation is not copying fields from evidence into a cleaner table. The pair compares a new claim with the current whole and asks:

  • what, exactly, does this evidence establish;
  • where does its source have authority;
  • what agrees, conflicts, or remains unknown;
  • which current values would change;
  • what else in the coherent domain would that change affect; and
  • what uncertainty remains after the best available judgment.

The output is a proposed delta bound to the state against which it was reasoned, not a rewritten snapshot. Flow uses an exact account revision; another stone may express that concurrency boundary differently. Before and after values, cited evidence, rationale, affected records, and downstream consequences are first-class data. If one decision changes several related records, they travel as one change because partial truth is a new falsehood, not half a success.

The whole does not always mean an entire company. Each product must name its coherence boundary: the smallest unit across which a local answer can create a hidden collision. In Flow that unit is the customer account. In another stone it will be learned from the domain. The invariant is that a convenient screen, row, source, or task is never allowed to choose the boundary silently.

The Peer can do nearly all of the labour: extract, join, detect divergence, trace provenance, draft the delta, explain consequences, and present the decision at the right grain. That abundance does not confer authority. The declared authority reviews the proposed transition and grants the consequential judgment the product requires. In Flow, that authority is the human operator.

Review binds something exact. The reviewed proposal, evidence set, before values, and expected state are the thing that may apply. Flow binds the bytes with a hash; the integrity mechanism is a product decision. If current truth moved while the proposal waited, the transition is stale and nothing happens. The work remains available for deliberate rebase; it is never merged by optimism. If validation fails, nothing happens. If one mutation in a coherent change cannot apply, nothing happens.

An accepted transition then does two things together:

  1. it advances current truth to an ordered successor state; and
  2. it appends the complete attributable record of why that advance occurred.

That boundary is where evidence becomes authority. Conversation is not the boundary. A model stating a conclusion is not the boundary. An authority saying “looks good” without binding the reviewed change is not the boundary. The machine is trustworthy because there is one mechanical answer to the question what can make this true here?

A current value without ancestry is merely the latest assertion. Every consequential field must be able to name the evidence and reviewed transition under it. The current record carries the useful answer; provenance can recover the argument.

This changes what the pair can do. A grounded answer is no longer prose with a confident tone. It is a read from current truth that can name what supports each material value, distinguish shown meaning from stored representation, and say where the account is silent. The Peer does not have to reconstruct belief from raw files every session, and the human does not have to trust the Peer’s memory of how a conclusion was reached.

The append-only change record also makes the delta a durable product output. The machine can answer three different questions without laundering one into another:

  • What do we believe now?
  • What changed, and why?
  • What did we tell this person at that time?

That third question is why publications freeze. Regenerating yesterday’s message from today’s truth destroys the history of the encounter. Sent state and acknowledgement state remain separate for the same reason: producing, sending, receiving, accepting, and incorporating are different events.

The truth machine is not an autonomous pipeline with a human approval button at the end. The pair works inside the stone on the same evidence, current truth, and transitions. The Peer supplies attention across the whole: intake, comparison, retrieval, drafting, verification, and the unanticipated case. The operator supplies the judgment, scoped authority, authorship of asks, and permission for consequential acts that the product assigns to the human.

As judgments recur, deterministic code can take over their mechanical labour. A compiled read can replace a join the Peer was repeating; a drafter can assemble a proposal the pair was rebuilding. Promotion moves labour, never authority. The same evidence, review, and transition boundary remains. This is how the machine compounds without becoming an autonomous oracle: the practiced interior hardens while the pair’s attention moves to the frontier.

Remove either half and the product changes category. A ledger without the pair can preserve known workflows but cannot economically absorb unanticipated ambiguity. A pair without governed transitions can produce insight but cannot safely make it durable. The mass is their union, not an AI laid over records and not records hidden behind an AI.

When an encounter uses a facet, the cut is safe to keep thin because it does not perform reconciliation. A fact flows to it only after the mass has made its status and meaning explicit: current truth appears as current truth; a bounded source observation may appear as an observation when the distinction itself matters. Evidence is never silently promoted merely because it is useful to show. An ask reaches the facet only after its declared author has taken responsibility for the obligation. The facet projects the right slice for a person; it does not invent, adjudicate, or retain a competing account.

The facet is not an admission gate. A call, email, spreadsheet, direct system observation, or another existing channel can carry the same encounter. The mass admits the arrival as evidence under the same provenance and authority rules. A product-authored cut earns its existence by making a recurring boundary excellent, not by making every other route invalid. The mass must survive the cut being used, ignored, shared, changed, or deleted.

What comes back is equally disciplined. An answer, correction, file, approval, or acknowledgement is an immutable contribution and therefore new evidence. It may support a later change; it never edits current truth directly. This asymmetry is the clean seam between outside participation and inside judgment:

current truth or labelled observation ──► facet or existing channel
authored ask ──► facet or frozen publication
outside contribution ──► mass as evidence

The loop explains all six properties of a facet. It has no interior because the machine is behind the cut. It can be disposable because truth never moved into it and the evidence boundary does not depend on it. It shows only what is true or authored because those are the only two outbound paths. Thinness is not a visual style; it is the visible consequence of a complete reconciliation discipline elsewhere.

A Portable Grammar, Not a Universal Schema

Section titled “A Portable Grammar, Not a Universal Schema”

The mechanism is intended to travel beyond manufacturing coordination, but that does not make Facet a generic truth platform. The candidate transfer is a grammar:

evidence → reconciliation → reviewed delta → current truth → encounter

Every real stone must still learn and encode its own nouns, invariants, coherence boundary, source authorities, uncertainty vocabulary, consequential operations, approval policy, and publication states. Those are not configuration details to abstract away. They are the domain becoming legible.

This is the important kind of generality to test: the discipline may govern a particular practice, account, operation, or institution whose people act against durable and disputed truth, while every implementation remains shaped by the reality it reconciles. Flow has paid for the grammar once. A second stone must challenge it before local mechanisms become universal doctrine. A universal data model would erase exactly the authority and meaning the machine exists to preserve. A generic Facet runtime would be the portal reflex moved behind the cut.

The fit test therefore remains strict. If there is no durable shared truth, no meaningful disagreement between sources, no consequential transition, or no judgment worth attributing, use a script or an ordinary application. The truth machine is broad because evidence and reconciliation are broad, not because every product needs one.

Flow made this loop concrete before it became doctrine. Customer spreadsheets and calls and read-only ERP observations enter as evidence. Its coordinator contribution path is implemented under the same rule. On 2026-08-04 the coordinator and customer were each exposed to Flow, while the account continued to admit evidence through its existing channels. Authority is scoped: customer intent, execution reality, and ERP facts are not interchangeable, and none silently determines production priority. The pair reconciles against the whole account. A change names the revision it expected, cites its evidence, carries exact before and after values, preserves uncertainty, and either advances the account atomically or does not advance it.

That discipline prevented several plausible shortcuts. A posted shipment is an ERP event, not automatic proof that a customer need should be retired. A value disappearing from a current ERP snapshot does not erase the prior event. A generated customer package is not sent, and a sent package is not acknowledged. A coordinator answer is attributable evidence, not a direct edit. Each distinction was learned where collapsing it would have made the account easier to store and less true.

Flow is one worked stone, not the schema for every stone. Its candidate reusable result is the loop and its boundaries. Its public evidence now covers first outside contact as well as implementation and operator use. It does not yet establish adoption, a completed contribution cycle, business value, or transfer to a second stone; the case study keeps that limit explicit.

A product has a truth machine only if it can answer these mechanically:

  • Can a new source enter without changing current truth?
  • Does the untouched source survive a bad extraction or interpretation?
  • Is every source’s authority bounded, rather than inherited wholesale?
  • Can every consequential current value name why it is believed?
  • Can a correction preserve the claim and decision it supersedes?
  • Is the delta itself durable, with before, after, rationale, actor, evidence, and uncertainty?
  • Can stale or partially invalid work advance nothing while remaining available for review?
  • Are generated, approved, sent, acknowledged, and incorporated states kept distinct?
  • Does a facet distinguish current truth from a bounded source observation when it shows both?
  • Can an outside arrival use another channel without creating a second truth path or being discarded?
  • Does an outside contribution return as evidence rather than mutate truth?
  • Can every facet be deleted without damaging the current record or its history?
  • If the Peer disappears, do the last valid truth and every completed decision remain intact and readable?

If any answer depends on a person remembering to be careful, the machine is not complete. That is a missing rail or a second truth path. The facet is the shape of presence, but presence becomes a product only when the mass can keep what it learns true.