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.

CEX-1 — DISMISSED ON ARRIVAL: the C7 allocation line's phase == "control" exclusion is dead code, but it sits behind a gather that already drops retired sleeves, so it changes no output

status: dismissed · raised AND killed in panel run 50 (2026-11-16) · class: would-be P0 safety on a pre-registered rule · never filed as open


What was claimed. f606336 changed meanrev's registry phase from "control" to RETIRED. The only production consumer of the string "control" is the exclusion in the C7 equal-risk allocation rollup (fleet_digest.py:959), so that exclusion can never fire again; its only tests fabricate the input (test_fleet_digest.py:82, :381). Meanwhile the robo has phase="live", so it would enter the pre-registered cross-sleeve allocation tilt despite being barred from real money twice over — while the rendered line still says "(control excluded)".

Every mechanism claim reproduces. sleeves.py:55-88 emits only "live" and "retired"; config/fleet.yaml:110 scopes the weight-sleeve branch to "weight sleeves (momentum, meanrev)"; _sleeve_stress_fraction:910-929 branches on contract == "order_intent" and falls through for anything else; test_robo_config.py:128-133 asserts evaluate_go_live(cfg) raises for the robo. Measured tilt against the real configs: momentum 0.153 / vrp 0.765 / robo 0.082, robo stress_frac 0.185688.

Severity collapses on four measurements.

  1. The guard is doubly dead and changes no output. fleet_digest.py:50-52 already drops retired sleeves at gather, so meanrev never reaches _stress_allocation at all. phase != "control" is dead code behind a dead path; deleting it changes nothing today.
  2. The C7 line is report-only with no consumer. Nothing reads stress_alloc / weights outside the two renderers (_stress_line, :1000-1003). The config/fleet.yaml amendment is itself pre-registered as report-only pending a correlation-evidence gate (≥126 overlapping days AND ≥1 joint stress observation — the robo has zero days).
  3. It cannot render for the robo for months. _stress_allocation:950-953 requires a non-empty equity_curve AND stress_frac > 0 AND len(prom) >= 2.
  4. The SGOV objection runs the conservative way. Marking 10% T-bills to an 18.57% equity crash over-states the robo's stress fraction, which under-weights it in an equal-risk tilt — on a report-only line, for a sleeve barred from real money.

On the pinned-config question: config/fleet.yaml:110 is silent, not violated. It describes the weight-sleeve branch's intended scope; it does not forbid a third contract, and the fall-through is unspecified rather than contradictory. Amending it would touch FLEET_KEYSTONES — a keystone-pinned, pre-registered allocation rule — which a routine may not propose on report-only evidence. That should have been the candidate's own kill.

The one true residue, routed not filed. "(control excluded)" at fleet_digest.py:1000-1003 is appended unconditionally and is now a false statement about a pre-registered rule, in the owner's only email. One conditional, plus the dead-branch deletion. Routed as a fold-in on the next fleet_digest.py touch — with a dated default, because "fold in on next touch" demonstrably failed twice for public_export.py inside this same week's delta: if it has not been folded in by the next weekly audit (2026-12-01), it becomes an ops/HANDS.md line.

Reopen conditions (any one)

  1. Any consumer reads stress_alloc / weights for anything other than rendering — including a human decision recorded as citing the equal-risk tilt. The row's whole severity turns on there being no consumer.
  2. A second never-promotable sleeve is registered — a new negative control, or any sleeve with no go_live: section — at which point the dead exclusion becomes a live correctness bug in a pre-registered allocation rule and the fix stops being cosmetic.
  3. _sleeve_stress_fraction is used for a sizing or allocation decision rather than a printed line, which makes the robo's silent fall-through denominator load-bearing.