Skip to content

ADR-0046 — Execution conflict and intra-bar fill model v1

Status: ACCEPTED

Date: 2026-09-26

Context

Decision Gate DG-F.2 (SCOPE.md, atom H03) must resolve execution conflict semantics — same-bar conflict resolution, intra-bar ambiguity, order cancellation/OCO, and partial fills — before H03 can unblock H04/H05 per CAPABILITY_DAG.md. H03 requires H01 (Order/Fill lifecycle) and G04 (resolved by ADR-0045).

ADR-0012 establishes one canonical execution engine shared across rules, supervised and RL policy families; it does not itself define conflict resolution. CORE_CONTRACTS.md §23 (ExecutionSpec) already names "cancel/replace rules" and "partial-fill model" as expected concepts without defining their shape; §26 (Position) names "quantity" and "average basis" without defining whether an instrument may carry more than one concurrent position.

Resolved collaboratively, this design turned out to reach further than DG-F.2's original four bullets: it also fixes parts of H01 (order replace as a state transition) and H04 (the position/ledger shape itself, since hedge-mode requires more than one concurrent position per instrument). This ADR is therefore authority for those parts of H01/H04 as well as H03 -- issues #143 (H01/H02) and #144 (H03/H04) must both follow it, not only #144. No new SCOPE.md Active Path step or Milestone is introduced by this broadening; it deepens the existing DG-F.2 design gate rather than adding a new one.

This ADR does not touch I06/I07 (Strategic/Execution RL, OPEN_DEFERABLE, Wave 11/12): per ADR-0011, execution RL "decides how to implement a desired trade/order objective" and belongs within the execution-policy boundary this ADR defines -- it will operate through these mechanics later, not redefine them now. Per ADR-0009, DecisionIntent is learner-agnostic; the decision logic that chooses prices, quantities, and when to move a level (trailing) remains a G01/G02 strategy-policy concern this ADR does not decide -- H01/H03 only execute whatever is injected, deterministically.

Decision

1. Intra-bar reconstruction

The canonical source of truth is trade-v1 (A01); a Candle is an accepted first-class derived representation (ADR-0005), not a replacement for the underlying trade stream. Accordingly, resolving what happens "within a bar" uses the real reconstructed trade path inside that bar's declared temporal bounds -- never a conservative bar-level convention (e.g. "assume the stop was hit before the target because both are within the bar's high-low range"). Events are ordered using the existing canonical order (exchange_ts, trade_id) (ADR-0040) -- no new ordering primitive is introduced.

Tie-break: if a single trade event's price crosses more than one conflicting trigger level simultaneously (there is no earlier/later order to resolve it), the more conservative outcome wins (e.g. a stop-loss trigger takes precedence over a take-profit trigger tied at the same trade).

Forward compatibility (not implemented now): this mechanism must be described in terms of "the finest-grained ordered event stream available within this bar's bounds," not hardcoded to always mean "query trade-v1 directly." A future capability to build a higher-timeframe bar from other, lower-timeframe bars (rather than only from raw trades) must be able to supply an analogous ordered stream without requiring H03/H05 to be redesigned. Designing that multi-grain source-resolution mechanism itself is explicitly deferred (OPEN_DEFERABLE, evidence-triggered, matching this platform's existing treatment of A12/L1-L3) -- this ADR only requires that today's implementation not preclude it.

2. Position model: hedge-mode (touches H04)

An instrument may carry two independent concurrent positions, one per side (LONG, SHORT), each with its own quantity, average basis, and separately realized PnL -- not one net position with a flag. This is "hedge-mode" as implemented natively by major crypto perpetual venues (Binance, Bybit), not a generic multi-asset concept: it is scoped to one instrument's two sides, and is unrelated to and does not activate H06 (multi-asset execution, OPEN_DEFERABLE, out of this Wave's scope).

A future aggregated cross-side view (e.g. one combined "net exposure" report spanning both sides) is not designed now and is not precluded by this shape.

3. Opening (entry) logics

Three mutually exclusive patterns, selected and injected by the strategy (G01/G02) per trading decision -- H01/H03 process whichever is injected generically, without hardcoding which pattern is "normal":

  1. Single: one entry logic, one side.
  2. Hedge: two independent entry logics, opposite sides, both intended to remain live simultaneously and independently (per §2's position model).
  3. OCO (One-Cancels-Other): two or more competing, mutually exclusive entry logics -- exactly one is meant to actually execute. Legs may target the same side (e.g. "enter long on breakout above X, OR enter long on pullback to Y") or opposite sides; direction is just a parameter of each competing leg, not a constraint the OCO mechanism itself imposes. Any fill of any member leg, including a partial fill, cancels every other member of the group -- OCO legs compete for the same underlying intent, so a partial commitment to one already resolves the competition.

OCO groups are dynamic: legs may be added to or removed from a live group as explicit, individually recorded operations, not only at group creation. OCO groups do not support sub-grouping (a group's cancellation cascade always reaches every other current member; a caller wanting a narrower relationship must use independent legs, per §4, instead of a group).

4. Exit (position-management) logic: independent multileg (touches H01, H04)

Once a position(-side) exists, its exit management uses a different mechanism from §3, regardless of which opening pattern created it: an open-ended set of independent legs (dynamic stop-loss and take-profit orders at up to N target levels), injected and updated by the strategy.

Each leg carries its own allocated quantity. Filling one leg, in whole or in part, never cancels or reduces any other leg -- this is the opposite cancellation behavior from §3's OCO groups, because multileg exit orders are not competing alternatives, they are a scaled/laddered plan where each leg is meant to execute independently as the position is worked down.

Invariant: the sum of quantities allocated across all currently active legs for one position(-side) must never exceed that position's current open quantity. This is a quantity-conservation constraint enforced deterministically by H01/H04, not a cancellation rule -- an attempt to allocate more than the remaining open quantity is refused explicitly, never silently clamped.

A "reduce-only" order (an exit order with no OCO peer) is the degenerate N=1 case of this same mechanism, not a separate construct: an independent leg with no siblings behaves identically to what "reduce-only, does not cancel with anything" already means.

An order belongs to exactly one of: an OCO group (§3), or the independent-legs set (§4). §3 and §4 do not overlap on the same order -- they apply to different, non-overlapping situations (competing entries before a position exists, versus exit management of an already-open position), and a strategy that wants OCO-style linking among exit orders can already express purely competing exits as an OCO group of exit orders, which is a valid but distinct use of §3, not a blending of §3 and §4 on one order.

5. Order replace / trailing mechanics (touches H01)

H01's Order state machine supports replacing a resting (unfilled) order's trigger/limit price in place, preserving its order_id and its OCO-group or independent-leg membership unchanged. This is the mechanic a strategy's trailing-stop or dynamic-target ExitPolicy decision (a G01/G02 concern, not decided by this ADR) is executed through.

For v1, only the price is replaceable. Replacing a leg's allocated quantity is not designed now and is not precluded by this shape (it would still need to respect §4's quantity-conservation invariant whenever it is designed).

Replace-vs-fill ordering: a replace instruction is itself decision-time evidence, causally anchored (ADR-0006) to the point in the §1 reconstructed chronological trade stream at which the strategy's decision was made. If the order's prior trigger price was already touched by a real trade earlier in that same chronological stream, the fill has already happened and takes precedence -- the replace cannot retroactively un-happen it. No new rule is required for this: it is a direct consequence of processing the reconstructed stream in canonical chronological order (§1).

Excluded

This ADR does not define or authorize:

  • the multi-grain (bar-from-bars) intra-bar source-resolution mechanism itself (§1) -- deferred, OPEN_DEFERABLE;
  • an aggregated cross-side portfolio view for hedge-mode positions (§2) -- not designed now, not precluded;
  • leg quantity replace (§5) -- not designed now, not precluded;
  • H06 multi-asset execution, K07/K09, J02/J04-J08 -- all remain out of SCOPE.md's Wave 4 scope, untouched by hedge-mode being single-instrument;
  • I06/I07 Strategic/Execution RL state, action or reward contracts -- those remain OPEN_DEFERABLE; this ADR only fixes the execution primitives they will later operate through, per ADR-0009/ADR-0011;
  • H02's cost/synthetic-fill model itself (fees/slippage) -- H02 applies its own existing, separately-resolved model to whatever fills this ADR's mechanics produce, regardless of OCO/multileg structure.

Consequences

  • Issues #143 (H01/H02) and #144 (H03/H04) must both implement per this ADR, not only #144 -- #143 owns the order-replace state transition (§5) and the reduce-only/independent-leg order shape (§4); #144 owns the hedge-mode position/ledger shape (§2), OCO group mechanics (§3), the quantity-conservation invariant (§4), and the intra-bar reconstruction/ conflict resolution (§1) composed with H01's state machine.
  • H03 can be credited RESOLVED in docs/architecture/OPEN_DECISIONS.md and docs/product/CAPABILITY_DAG.md once this ADR is accepted, per the governance-boundary rule that only an explicitly designated governance issue may make that edit (#147, not this one).
  • H04 (which requires H01, H02, H03) remains blocked until #143 and #144 are both credited complete, not merely once this ADR is accepted.
  • A future Execution RL design (I07, deferred) inherits a substrate that already supports hedge-mode, competing entries, and independent scaled exits -- it will select/parametrize within these mechanics rather than needing new execution primitives invented for it later.