The entry enum is extended by decision, one mechanism at a time

EntryKind is a closed enum (M3.2, R4.2's sibling): a kind that is not a member raises at construction. Six members exist — two apparatus kinds, ALWAYS and COIN_FLIP, and four mechanism kinds, BREAKOUT_HIGH, BREAKOUT_LOW, CLOSE_ABOVE_MA, CLOSE_BELOW_MA. Each is one branch of signal() in occams/engine/day_boxed.py, which reads bars [0, first) and never the box itself; one row in the compiler's GEOMETRY table, which is what refuses a buy-stop below the market (M3.4, the v1.8 defect by name); one row in REQUIRED_PARAMS; and one component of the spec hash (D4, ADR-0007).

The enum is closed so that the space of statable mechanisms is enumerable. That is what makes "what did we search?" answerable (N6) and what the search correction can be honest about (R4.4): k counts cells inside one kind, and the overlap gate (R4.9) sees every question on an axis. A kind added on a whim is a family added to the search without a record.

On 2026-09-12 the first price_daily prepare (DRAFT-004) showed the closed enum's limit: every mechanism it can state on daily bars is at or below zero gross on the definition partition against the declared floor, and the best of them — a close below its moving average — fires on half the days of a bear market and says little about why the next days should revert. The floor is the author's declaration and is not this ADR's subject. The enum is. The question is not whether to open it but how a kind is added so that the enumeration stays honest.

Decision (proposed)

The rule

An entry kind is added only by an ADR that states, before any prepare on it:

1. The mechanism it exists to test, in one sentence, with both interpretations — why prices alone would carry an edge after this signal, and what a null would mean.
2. **Its formula over bars [0, first)** and its integer or percentage parameters. Parameters belong to the Hypothesis (fixed in the template or swept as an axis); the kind is the family.
3. Its geometry row: whether it yields a level and where that level sits relative to the market at decision time, or whether it fires as a market entry (NONE). This is the row M3.4's refusal reads.
4. Its look-ahead test: a fixture on which the signal day's own bar would fire it and the prior bars would not, asserting it does not fire.
5. Its Pine obligation: to_pine (M11.1, cap-blocked) must derive it from the spec when that row opens. Until then the Python signal is the one source (R8); no kind may exist only in Pine (M11.3).
6. Its auditor family (added at adoption, found by building M3.9): a kind belongs to a family in occams/costs/auditors.py or its fills cannot be audited, and unaudited means refused (M5.7). The three kinds below form the reversal family and share the trend family's auditors: overnight gap, opening auction, halt.

A kind proposed after reading the measurement partition is refused (D13). Reading the definition partition is permitted — it is the calibration set — and this ADR was written after reading it; the kinds below are stated as mechanisms with a prior reason, and their parameters are not chosen here. A kind is never tuned into existence.

Every question on a new kind is a question on the same axis and the same partition as the ones before it, so the overlap gate refuses it until it states its distinction (R4.9). The distinction is the mechanism sentence from point 1. That is the accounting for choosing among kinds: not a Bonferroni term, but a recorded, refusable declaration per family.

Three kinds this ADR adds, if adopted

All three are long-side market entries on the next open (GEOMETRY row NONE for both sides), fire from bars [0, first) only, and carry a time exit and a percentage stop like the existing moving-average kinds.

Each is a conjunction the existing kinds cannot express: DOWN_RUN is not a moving-average relation at all; RETURN_BELOW is a magnitude; PULLBACK_IN_TREND needs two lookbacks in one entry, which Entry can carry but no current kind reads. Their mirrors (UP_RUN, RETURN_ABOVE) are not added: the persistence mechanism was prepared as candidate B on 2026-09-12 and was worse than random entry in every cell; a second persistence family would be a family added without a reason.

What does not change

• The two apparatus kinds stay two. The null for any new kind is the geometry-matched random entry the beats-null check already draws; a kind needs no control of its own.
• The spec hash covers the kind and its params as it covers the existing ones (D4). A question on a new kind is a new family with a new hash.
• Side.SHORT on these kinds is constructible but not tradeable here: no borrow capability exists in Capability (ADR-0018), so a short spec would be refused at compile for an undeclared capability the venue does not offer. Shorting is a capability decision, not an enum one.

Consequences (if adopted)

• occams/spec/spec.py: three members. occams/spec/compile.py: three GEOMETRY rows and three REQUIRED_PARAMS rows; to_engine refuses PULLBACK_IN_TREND with short >= long by name.
• occams/costs/auditors.py: a reversal family; family_of refuses a kind with no family by name instead of failing on a lookup.
• occams/engine/day_boxed.py signal(): three branches, reading _on_basis rows as the moving-average kinds do so a split loads as an action. The position-boxed engine reuses signal unchanged.
• Tests: one look-ahead fixture per kind; a hash test that a new kind's spec differs from every existing kind's at equal params; the M3.4 geometry test extended to assert a NONE row admits only market orders.
• occams/proposers/price.py: a mechanism sentence per kind, both interpretations, a falsifier — the text in the table above is the draft of it. question prepare --entry gains the three names.
• TASKS-v4.md: a row M3.9 — entry kinds added by ADR-0037 under M3, done when the five points above are met for each kind; a note on M11.1 that to_pine owes three derivations.
• CONTEXT.md: entry kind is a family; mechanism is what a kind exists to test; a kind is added, never discovered.

Considered options

• Keep the enum closed and change nothing. The honest default, and the one the lab's own falsifier points at: two nulls, no promising third. Rejected only conditionally — if the author holds that the closed enum's four mechanisms are the whole of what daily bars can say, the lab should close on the third null rather than add families to avoid it. This ADR exists so that a third question, if there is one, is stated as a mechanism and not as a way around a verdict.
• An open expression entry — a formula string evaluated over bars. Rejected: the hash would cover free text, which is the free-text kind M3.2 refused; a formula language is a garden of forking paths with no enumeration and no per-family record; the look-ahead test cannot be written once for all formulas.
• Add the usual indicator set at once — RSI, MACD, Bollinger, rate of change, and their thresholds. Rejected: each is a family that will be searched; adding ten at once is adding ten searches with one ADR's worth of reasons. One mechanism, one kind, one reason.
• Gap kinds (GAP_DOWN, GAP_UP): open below the prior low, above the prior high. Rejected for now: the signal contract reads bars before the box and never the box's own open; a gap is known only at the box's open. Admitting it means a second signal contract ("at the open") with its own fill and look-ahead rules — a separate ADR if wanted.
• Re-declare the floor instead. The floor is the author's (M0.15, DRAFT-004); no value is proposed here, and changing it is a recorded decision of a different kind. This ADR is orthogonal to it: the kinds above would be worth stating at any floor.

Adoption

Adopted 2026-09-12 by the author with all three kinds; M3.9 opened and built the same day. A prepare on a new kind is a reading of DRAFT-004's successor, and registration remains the human act.