Decisions with receipts: the method

Most marketing decisions are made from information you cannot check, and the only thing that eventually tells you they were wrong is the profit and loss statement, months later, after the money is already spent. By then the budget is gone and the lesson is a guess about which of ten changes moved the number.

A decision you cannot defend is a decision you cannot repeat and cannot learn from.

The failure mode is a broken feedback loop. An action is taken on a claim, the claim's provenance is not recorded, and the only signal that closes the loop is a lagging, aggregate, confounded financial outcome.

Attribution across that lag is underdetermined: many changes, one number, no isolation. The correction arrives too late and too coarse to update the rule that produced the decision.

oddly's numbers come from real commerce it has actually looked at, and the counts on the site are the counts in the corpus today, growing week over week. There is no invented adoption number: until a figure is real and live, the stat is hidden rather than fabricated.

oddly re-checks a connected store every 15 minutes and still changes nothing without you.

The corpus is served as aggregate counts only, with no per-store identity and no personal data, edge-cached, every count coalesced to zero with a retrieval stamp. A count that cannot be read reports zero and the stat is hidden, which is honest-null by construction.

The cadence is a scheduled trigger every 15 minutes that re-syncs the store and re-evaluates rules and playbooks. The loop proposes, it does not mutate. That cadence is stated as true because the trigger list says so at time of writing, not as a permanent property, and it is re-checked before use.

oddly can never quietly raise your ad spend, on any plan, and that limit has no upgrade. The downside you are underwriting when you connect an account is bounded to zero.

Nothing changes until you approve it. You see exactly what a change would do before it happens, and you can undo it inside the window.

Ad-spend mutation is decrease-only by construction. A connector that could raise spend fails the build rather than shipping, so the asymmetry is a build gate, not a policy toggle a quarter could reverse.

The mutate path is propose, verify, dispose: dry-run preview, then human approval, then a reversible-within-window apply, each action traceable to the rule that fired it.

The corpus is small today and oddly says so. A handful of opted-in stores means many peer groups are still forming, and that absence is stated, not papered over.

The observation of a given week cannot be re-run later, because the commerce that produced it has already moved on. The honest record is the observation stamped with its date, not a claim of permanence.

Observation is longitudinally irreproducible: a metric is the most recent observation carrying its own retrieval date, and a re-run observes a different world rather than reproducing the old one.

This is why the method fixes the receipt at observation time. You cannot go back and receipt a claim after the fact. The provenance is captured when the number is read, or it is lost.

oddly only tells you where you stand against a peer group when the group is real. If the cohort is still forming it says so, and no single store can be reverse-engineered from the comparison.

A cohort number is served only under a floor of k greater than or equal to 5 distinct merchants paired with the spend bucket, and each contributor is reduced to an opaque hash before aggregation, so the comparison never touches a store's identity. Below the floor the cell is served as an honest null, never zero and never an estimate.

When oddly has nothing to say, it says so, in one line, rather than filling the gap with a guess. "Cohort forming" is a real answer, not a softened estimate.

Absence is a first-class value with its own rendering: a sub-floor cell returns the honest-null dash and gets a line on the index with no page of its own, and a corpus count that cannot be read reports zero and hides the stat.

A surface that goes quiet where it has nothing to say is making a claim, that it has nothing to say, and that claim is stated rather than implied.

Every move oddly names comes from a rule you can read, applied to a signal you connected, and the same inputs always produce the same output, so there is no black box to trust.

When oddly explains a move in plain words, the words can be model-written, but the decision and every number behind it are not.

The rule catalog is the receipt: deterministic rules over connected sources, served from the registry so the page cannot drift from what the engine runs, with the decrease-only rule itself enforced as a build check.

The language layer is default-off and read-only over the verdict. It may rephrase two human-facing lines, never mutate a rank, a play, or a figure. The narration envelope enforces this by construction, because the writer never receives the claim.

Ask oddly the same question five times and the answer does not move. If an answer moves, it was never a measurement, and a number that runs your ad budget should not move.

On the published grounded eval, the oddly engine produced a byte-identical answer set across independent re-runs with zero fabricated claims, while language-model pipelines on the same closed-form store questions did not.

Carry the caveat always: the reproducibility is a structural property, because a calculator does not disagree with itself, and not independent brilliance. The interesting column is the models'.

oddly does not ask you to trust the pitch. Every strong claim points at the thing that backs it: a corpus number you can see, a mechanism you can read, or a stated absence.

If a claim has no receipt, it is removed, not softened.

A receipt is one of exactly three forms: a corpus number with a path to its source, a named mechanism enforced in code (a build gate, a scheduled trigger, an envelope), or a named absence.

The adversarial test is the definition. Not "is this plausible" but "would this finding survive a reviewer whose job is to refute it". That is why the mechanism receipts are build gates rather than promises: a promise is refutable, a failing build check is not.

oddly is largely built by its own operator, running continuously, and every change it makes to the product is held behind a human gate before it ships.

The person approving is approving specific content, not signing a blank cheque. The accountability layer oddly sells is the same one it runs on itself.

The operator is an autonomous build runner. It proposes changes as pull requests and cannot merge a gated class of change without a review approval, from an allowlisted approver, on the pull request's current head commit.

Consent attaches to content, never to a pull-request number: an approval on an older commit does not count. The gated classes are mechanical, not a judgement call, and they include a change to the operator's own instructions, so it cannot grade and release its own control plane.