Entries & beliefs

Everything in Ledger is an entry. This page is the full anatomy: the three kinds, the expectation, evidence, settlement, and the rules that keep the record honest.

Three kinds, two shapes

The split that matters is decisions vs. memory. A decision stakes a claim about the future and owes the workspace an answer. Lessons and observations are memory — there's nothing left to find out, so they arrive with nothing open. Ledger enforces the difference: only decisions can carry an expectation, and a decision's lesson can't be written at creation — it's written at settlement, after the evidence, never before.

The expectation

A decision's What we expect field is written at creation, before any evidence exists. It comes in two forms:

Measurable — when the outcome can bind to data. A claim names its metric from the Metrics tab, a comparator (>=, <=, >, <, =), a numeric target, and an optional unit. Pilot conversion · >= 12 % · by Mar 31 is something reality can check mechanically: a reading lands on the metric, the number either clears the bar or doesn't.

Most decisions move more than one number, so an expectation holds up to ten claims. Another number we expect to move adds one of those. Something this shouldn't cost adds the other kind — a number you're not trying to improve but refuse to wreck, written as this metric shouldn't drop more than X% below the N days before this decision. That's how you say "more CTA clicks, without losing newsletter reads" and have both halves count.

Any claim can be marked keep an eye on this one without committing to it. It's reported when you settle, but it isn't part of the promise — for the number you suspect matters and aren't ready to stake anything on.

Freeform — when it can't. Room decisions, market calls, organizational changes: "Participation in the benefits program up noticeably by Q4." Less checkable, still checkable — a freeform expectation written in advance is worth far more than a precise one reconstructed after the fact.

An expectation must say something checkable — a target or the freeform text. Ledger rejects an empty one, because an expectation that predicts nothing can never be wrong, and an entry that can never be wrong teaches nothing.

Evidence

As reality reports in, it attaches to the entry as evidence under What happened — a metric reading or a manual note ("pilot conversion held at 12% through week 3"). An entry can accumulate several pieces of evidence before settlement, and evidence can still be added after settlement — the record of what happened stays open even once the lesson is written. The one place evidence can't go is a superseded entry: it belongs on the successor.

Settlement

Settling is the compounding event: the evidence is in, the expectation closes, and you write What we now believe — the distilled lesson, which is required. Restating the number isn't a lesson; the lesson is what the workspace should do differently because of it.

You answer did the prediction hold? with one of three verdicts:

• Confirmed — it held.
• Mixed — some of it held and some of it didn't. The honest answer when a decision moved the number you wanted and cost you something else, which is why an expectation can hold more than one claim.
• Missed — it didn't hold.

Settlement happens exactly once. The transition is open → settled, it locks the entry, and there is no unsettle. If a lesson later proves wrong, the correction is a new entry that supersedes this one.

Being wrong is a fine outcome. An expectation that missed, settled with an honest lesson about why it missed, is worth more to the workspace than a hit — it's the entry that stops the next person from making the same confident mistake. The framing that matters is workspace-level honesty: how much of what we believed turned out to be true, and what we learned where it didn't.

What can change, and what can't

The line runs through settlement. Before it, an entry is a proposal; after it, the claim is part of the record.

Superseding is how the record corrects itself. A reversal or refinement is recorded as a new entry that backlinks the old one; the old entry flips to superseded and drops out of the lessons, but keeps everything it ever held. Follow the chain in either direction from the entry detail — what replaced this, and what this replaced. Both open and settled entries can be superseded: an open one when the decision was reversed before evidence arrived, a settled one when a later result overturned the lesson.

Retracting is for entries that shouldn't be in the record at all — a duplicate, a hallucinated metric, a draft that never should have landed. Retract pulls the entry out of the stream and the lessons; it isn't a purge, and Restore brings it back. Pulling a bad entry from the record is a different act from rewriting a prediction, which is why retract exists and in-place editing of settled claims doesn't.

The part you can't rewrite is the part you can trust. An expectation, once settled, reads exactly as it was written before the evidence came in — so when the lessons segment says "we tried this twice and it regressed accuracy both times," nobody has to wonder whether that history was tidied up afterward.

The stream

The Decisions tab is the whole record, reverse-chronological, with a calibration stat on top — how often settled expectations came true — and one segment row:

• All — everything.
• Open — decisions awaiting evidence, each showing a days-open counter. This is the triage view: an expectation open for 60 days either needs its evidence chased or its deadline was fiction.
• Settled — decisions that closed.
• Lessons — lessons and observations.
• Superseded — the corrected record.

Once an entry has a lesson, the lesson is its headline in the stream — the change it came from becomes the receipt underneath. The memory is the product; the decision is its provenance.

Lessons

The Lessons segment answers "what do we know here?" — every lesson the workspace holds, newest-settled first, each citing the entry that earned it like case law citing cases. Superseded entries have dropped out, because a newer entry speaks for them now.

Two readings it's built for: a new teammate's day one ("here's what this workspace has learned, with receipts"), and any moment someone says "didn't we already try that?" — now answerable with a link instead of a memory.

Where — the Compass link

Entries can link the Compass pages they touch — the workflow, system, or people involved. That's what makes the memory navigable: pull every entry that ever touched a workflow before you touch it again. It's the link zv1's consult-before-acting runs on — before proposing a change, it checks the record for prior attempts on the same Compass page and how they settled.