Thales
← research journal

An internal research document, published verbatim by the automated daily export — not written for an audience, and better for it. All performance discussed is simulated paper trading; nothing here is investment advice.

DRO-1 — every VRP open is a sell, so the one "exposure left on" alarm inverts on the VRP sleeve, and a failed open MASKS a genuine failed close

status: open · raised daily audit 2026-11-05 · class: loudness & safety (G2) — the exit-4 path is trading-contract, not observability · effort ~0.2 pd · horizon: live now, and arms the moment VRD-1 is fixed — so it must be built with VRD-1 or before it

Filed above the pre-registered 12-row cap. Per #306 the cap "is triage's move, not a filing gate", so this row is filed and the eviction is triage's (2026-12-03).

CORRECTION, 2026-11-05 same evening (steward's 11-05 record, and my own probe output agreed with it all along). This row as first filed said the masking hole arrived with #320 — its title read "after #320 a failed open will also MASK a real failed close", and leg 2 was written as a finding about that PR. That is wrong. The hole predates #320: the OLD side-blind masking set masked a failed VRP open in exactly the same way. My own first probe printed both rows as SILENT and I framed it as #320-introduced anyway. #320 is neutral here — it neither created nor closed this hole, and it does close a real and wider one on the equity sleeves. Corrected throughout below.

The correction makes this row more urgent, not less: the silence has been available since the exit-4 escalation was built, not since 2026-11-05.


Plain-language summary for an owner reading one paragraph. When an order meant to reduce the book fails, thales run exits 4 — which reds the workflow and fires its failure alert — and prints "🔴 NEW DE-RISKING FAILURE … (exposure left on)". That is the platform's only alarm for "we told the broker to get us out and it refused". The test for "is this a de-risking order" is side == "sell" or action == "close". For an equity sleeve that is right. For the VRP sleeve it is backwards: a put credit spread is sold to open, so every VRP open is a sell. Two consequences. First, on 2026-11-02 a failed VRP open was announced as a de-risking failure with "exposure left on" — the inverse of the truth, since the position never went on and the sleeve stayed flat. Second, and worse, a failed open on one session sits in the alarm's masking set and silences a genuine failed close on the next — and because VRP trades only SPY, the two always collide.

Ground truth — measured 2026-11-05

The live record (data/state/vrp/order_log.jsonl, 2026-11-02):

{"date":"2026-11-02","symbol":"SPY","side":"sell","shares":1,
 "structure":"put_credit_spread","action":"open","expiry":"2026-11-30",
 "status":"failed",
 "error":"APIError: {\"code\":42210000,\"message\":\"options market orders are only allowed during market hours\"}"}

and the run summary for the same date: "action": "open", "n_submitted": 0, "n_failed": 1. The sleeve was flat before and flat after. What the run printed:

🔴 NEW DE-RISKING FAILURE — 1 sell/close order(s) failed for the first time (exposure left on):
  • sell 1 SPY [untagged] APIError: {"code":42210000,...}
Exit 4: state is persisted; the failure alert fires.

Leg 1 — the predicate. src/thales/execution/failed_streaks.py:is_de_risking replayed against that exact record:

is_de_risking(the real 2026-11-02 VRP OPEN)  -> True
is_de_risking(same order with action=close)  -> True

Both True, so the function cannot tell an opening credit spread from a closing one. action: "open" is present in the record and ignored.

Leg 2 — the mask. Present both before and after #320. Replayed against merged main 107cffa, prior session = the real 11-02 failed open, today = a genuine failed close:

masking set usedmasktoday's failed CLOSEexit
side-blind (the set in use before #320){SPY}masked0 — SILENT
de_risking_only=True (the set #320 introduced){SPY}masked0 — SILENT
control: prior order with side: buy{}fresh4 — alarm fires

The two masking sets behave identically on this input, which is the measurement that disproves the "#320 introduced it" framing. The control is the proof the probe works, and separately that #320's own fix does what it claims for the case it was built for.

Why the collision is the likely case, not a coincidence. VRP is single-underlying: every order is SPY, so there is no "different symbol" escape. And the two failures share a cause — the after-hours backup leg, where Alpaca rejects options market orders outright (42210000). The same condition that fails an open fails a close, so the masked shape is the correlated one.

What currently keeps leg 2 from biting, and why that is not reassuring. VRD-1: cli.py reads str(result.run_date) unguarded inside the escalation's try:, and VRPRunResult has no such field, so on VRP the whole block raises AttributeError, the except sets the masking set empty ("unknown = new"), and every failed VRP order escalates. The sleeve is therefore over-loud today and never silent — and fixing VRD-1 alone removes that accident and turns on the silence. VRD-1 is the cheapest-looking row in the queue (~0.1 pd, one field, one assignment, "blocked on approval only"), so this is a live sequencing hazard, recorded as a warning block at the top of VRD-1 itself as well as here.

Third-party readings, stated accurately. This audit's own first pass took "de-risking failure" at face value, and so did PR #320's evidence section ("The VRP sleeve now has its first failed de-risking order too"). Queue triage's 2026-11-03 annotation on VRD-1 did not: it says plainly the order "was an open (risk-increasing), not a de-risking close, so no exposure was left on". An earlier version of this row grouped triage with the other two; that was wrong and is withdrawn. What no reader had reason to notice before the masking behaviour was replayed is that is_de_risking classifies that open as de-risking.

Fix shape (propose-only — this is the trading contract)

  1. Make is_de_risking read the order intent when it is present: an order carrying action is de-risking iff action == "close", and the side == "sell" test applies only to records with no action (the equity sleeves). One expression; no caller changes.
  2. Keep a failed open loud — it should still be visible — but not as "exposure left on" and not through the exit-4 path the owner's 2026-09-06 IGD-1 leg-3 ruling reserved for exposure-left-on events. Whether a failed open reds the run at all is a loudness choice worth stating explicitly.
  3. The [untagged] cause in the printed line is a separate small gap: VRP failed orders carry structure and action but no cause, so check 6's "cause tags present" cannot pass for this sleeve.

Negative controls (pre-registered)

  • The real 11-02 record must return is_de_risking == False; the same record with action: "close" must return True. Both are in this row's evidence as the pre-fix readings, so the test is red today.
  • The mask table above must flip its first two rows to "fresh → exit 4" while the control row stays unchanged.
  • An equity order {"side":"sell","symbol":"AAPL"} with no action key must stay True — the momentum/meanrev behaviour must not move. AVB's standing emergency_exit streak is the live fixture for this.
  • Paired with VRD-1: a failed VRP open on session N−1 and a genuine failed close on session N must exit 4. With VRD-1's field added and this row unbuilt, it exits 0.

Kill criterion

If the VRP contract is changed so that an opening order is no longer recorded with side: "sell", leg 1 dissolves and this row closes without a code change.

Relationship to #320 — merged 2026-11-05, and rightly

#320 closes a wider, different hole: a prior failed BUY masking a first failed SELL on the equity sleeves, with nine armed sessions measured and seven executed mutation controls. This audit recommended merging it and the steward agreed. It is neutral on this row — the side-blind set it replaced masked a failed VRP open the same way — so nothing here is a criticism of that PR, and this row is not a regression it caused. It is the VRP-shaped remainder of the same mask, filed so the remainder is not mistaken for closed now that #320 has landed.


Approved 2026-11-12 by the steward, under the owner's 2026-10-07 delegation — as ONE PAIRED BUILD with its sibling

Approved together: DRO-1 (dro-1-every-vrp-open-is-a-sell-so-the-exposure-alarm-inverts.md) and VRD-1 (vrd-1-vrp-result-has-no-run-date.md). Build them in one change, DRO-1's is_de_risking fix first. These two rows count as a single item for the implementer's one-at-a-time rule. Since #320, VRD-1 built alone turns a genuine failed VRP close into a silent exit 0, and the daily audit replayed that on merged main. Why approve now: the exit-4 alarm is the system's one "exposure left on" signal, so this is P0 loudness and safety under NORTHSTAR §4. Both rows have negative controls the implementer can test, including the paired control: a failed open on session N−1 followed by a failed close on session N must exit 4. The combined effort is about 0.3 pd. Neither row changes an order. They change only how a failed order is classified and the run's exit code. That makes this trading-contract code but not "what a sleeve trades", and the merge follows STEWARD.md §3's trading-path rules. DRO-1 leg 2, the loudness choice the row leaves open, is decided by the owner's 2026-09-06 IGD-1 leg-3 ruling, which reserves exit 4 for exposure-left-on events. So a failed VRP open stays visible in the log and the digest but does not exit 4. Leg 3, the cause tag on VRP failed orders, is in scope if it stays small. The implementer's next slot is 2026-12-05. REF-1 is already in approved/ with a 12-01 horizon, so under the deadline-first rule it goes first and this pair second. These are this week's two approvals.