Add a timescale-matched stop-regime variant #262
No Assignees
Notifications
Due Date
No due date set.
Dependencies
No dependencies set.
Reference: Brummel/Aura#262
Reference in New Issue
Block a user
Delete Branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
The sole risk-regime variant
vol{length, k}computesstop = k · Sqrt(EMA(Δ², length))over engine cycles, i.e. m1 bars (crates/aura-std/src/stop_rule.rs).lengthonly smooths the minute-vol estimator; the stop's timescale stays one cycle, so an H1/H4 signal cannot get a stop matched to its holding timescale except by inflatingkby hand (random-walk scaling, ≈ √60 for hourly).Empirical illustration (external 8.5-y H1-breakout matrix): vol{60, k2} vs vol{3, k2} left per-trade cost drag and trade counts nearly unchanged (the smoothing does not change the magnitude), while raising k to 6 / 12 cut the drag ∝ 1/k and the trade population by ~4× — the lever that actually matches the stop to the signal is k, which is unintuitive when the knob named
lengthlooks like a timescale.A variant computing the vol over resampled bars (
vol_tf{period_minutes, length, k}) or an ATR-style range regime would express "stop at the signal's timescale" directly. Independent of the feature, the length-smooths / k-scales semantics deserve a sentence in the glossary risk-regime entry (see also the units issue).Design reconciliation (specify)
Spec: vol-tf-stop. The issue offers two shapes ("a variant computing the vol over resampled bars (vol_tf{period_minutes, length, k}) or an ATR-style range regime"); resolutions below are derived orchestrator decisions (recorded 2026-07-14).
vol_tf{period_minutes, length, k}: the same estimator as the sole existing regime (k · √EMA(Δ², length)), computed overperiod_minutesbuckets instead of engine cycles.Basis: derived — it is the minimal extension of the existing vol regime (identical semantics, one added timescale knob, so the
length-smooths /k-scales reading recorded in the glossary carries over unchanged), and it directly implements the issue's own framing ("stop at the signal's timescale"); an ATR regime introduces a second estimator family (range-based) and is additive later without prejudice.Resamplemold: internal accumulator keyed byctx.now()bucket, state across cycles, emit only on bucket rollover), not aGraphBuildercomposition.Basis: derived — today's vol stop composes generic per-cycle primitives, none of which read the clock; no time-bucketing primitive exists except the OHLC-shaped
Resample, so the composition cannot express wall-time bucketing; the fused node follows the provenResampleaccumulator pattern and the partial-bucket-never-emitted discipline (C2 causality).Firing::Anyports (Sizer, PositionManagement) hold the last pushed stop distance.Basis: derived — verified against the engine's push/firing model: both consumers fire every cycle off price alone and read
stop_distancefrom the last-pushed window value, and the R-latch samples the stop only at the entry cycle, so a coarser update cadence needs zero engine changes and cannot look ahead (a bucket's value exists only after the bucket closes).--stop-length/--stop-ksugar stays Vol-only by construction;vol_tfis reachable through campaign documents (the authoring surface for regimes). The default regime path stays the single-sourcedvol{3, 2.0}— a new variant is reachable only by explicit opt-in.Spec auto-signed (specify Step 6, autonomous run)
Spec: vol-tf-stop (a second risk-regime variant
vol_tf{period_minutes, length, k}computing the same vol estimator over completed time buckets — a new fused stop primitive in the Resample mold; emission on bucket rollover only, downstream holds by firing semantics; default regime and stored Vol documents byte-untouched).The signature is the Step-5 grounding-check PASS — nine load-bearing assumption clusters ratified against currently-green tests, including the Ema seeding semantics that anchor the vol_tf{1} ≡ vol correspondence pin (report in the session transcript). No human signed; the fork decisions live at
issues/262#issuecomment-3538. Status: design settled, spec signed — proceeding to implementation planning.Design reconciliation, planning-stage addendum (2026-07-14): plan recon exposed a manifest round-trip hole the spec had not scoped — the member-manifest stamp is
if let StopRule::Vol(stamping stop_length/stop_k) and the reproduce-path re-derive reconstructs only Vol, so a VolTf member would stamp nothing andaura reproducewould silently fall back to the default Vol stop. Derived decision: in scope — the stamp gains a VolTf arm (stop_period_minutes/stop_length/stop_k) and the re-derive reconstructs VolTf whenever stop_period_minutes is present. Basis: the C1 reproducibility discipline (a stored member re-runs as it ran); a silent default fallback is the exact wrong-without-error class this whole batch has been closing. Also scoped in as mechanical record-reality prose: the two further sole-variant/slot-label doc sites (authoring guide, design ledger) update alongside the glossary. Spec amended and re-gated before planning.Re-attestation: the manifest-round-trip amendment recorded in
issues/262#issuecomment-3541was re-run through the grounding gate on the final spec bytes — PASS (the stamp/re-derive shapes ratified by the non-gated synthetic reproduce sentinel plus the archive-gated regime-stamp e2es; the three prose sites verified byte-current). The signature ofissues/262#issuecomment-3540stands on the amended spec.Second plan-defect note for the audit trail (2026-07-14): the vol-tf plan's executor-arm code sample contradicted its own mirror-the-Fixed-arm instruction — it showed a bare
g.add(VolTfStop::new(..)), which does not satisfyGraphBuilder::add'sInto<BlueprintNode>bound; the real Fixed arm is abuilder().bind(..)chain. The loop shipped the node slice (committed as the first #262 slice) and blocked the arm task on the contradiction; the plan is corrected to the builder+bind form and the remaining tasks re-run. The node's unit-test harness snippet was likewise adapted by the implementer to the crate's real eval-harness idiom under the plan's own fallback clause — assertions unchanged, hand-verified.