Mover v3 research plan and protocol draft from the verified 2026-09-24 SOTA sweep (not frozen)

[seathatflowsinourveins]

Records the verified 2026-09-24 extreme-mover SOTA sweep as a dated research plan and a v3 protocol draft. It is not frozen, and no data or outcome has been computed. The plan builds on the mover studies merged in #162: v1 and v2 had no pass in any family, and the holdout was never read.

Source. A multi-angle sweep covered literature, repositories, point-in-time data, exits and sizing, and high-rate execution. It ran 80 agents, and two refuters checked every proposal: one on facts, one on fit. 5 proposals survived, 24 were contested and 8 were killed. Only survivors and weakly refuted contested items are used, and the contested ones are labelled. Killed and dropped items are listed with reasons. The raw sweep file is not committed because it contains host paths; its sha256 and counts are recorded.

Contents

• blueprints/us-equities/mover-v3/README.md. The plan covers:
  • why continuation is not supported, with MAX/lottery and attention-buying sources;
  • hypotheses H1-H6: lottery-spike reversal, catalyst-backed vs pure-attention movers, overnight vs intraday legs, a halt-conditional exit (gated on paid data), a days-to-cover crowding filter (gated on data), and broker-held vs client-side exit parity on paper;
  • the data plan, split into what is available now and what is gated on user-only paid data;
  • the automatic exit and leverage policy, inside leverage-schedule-v1 and the 1x/2x/4x ladder gates;
  • the high-rate execution design, with a call-cost table for the 200/min and 1000/min tiers. 1000 round trips/min is unreachable on one account, and throughput counts as capacity, never as a target;
  • the simulation-fidelity plan.
• protocol-draft.json. Status draft_pending_independent_pre_outcome_review. It contains:
  • an exposure registry;
  • a prospective holdout: the first 252 sessions after freeze, because no historical window is clean;
  • one Holm family of 37 at α = 0.05, plus a lineage sensitivity check over the 1,012 earlier rule-exits;
  • MDEs for each stage;
  • the v1 cost table reused by sha, fixed for the two lenient items Mover studies: preregistered v1 and v2 (no pass in any family), data corrections D1-D6 #162's review found (monotone tiers, and capacity-capped notional plus a square-root impact term);
  • a holdout gate with an access log.
• Catalog decisions. The supplement catalogs/us-equities/mover-v3-sweep-20260924.json is registered through scripts/catalog_decisions.py --write --supplement …#/repository_decisions. It adds 7 repository records, all already indexed, and keeps papers, APIs and vendors as non-repository data.

Independent pre-outcome review. Two lenses reviewed the draft: protocol soundness, and source fidelity with evidence honesty. They raised 20 findings, 3 of them high:

• exposure of the development and validation windows via the broad-universe study was understated;
• H6 could not pass by construction;
• a trailing_stop and a separate OCO on the same shares is not allowed by Alpaca's order rules.

All 20 were fixed in 0b924e7. One alternative was rejected because registering the Qullamaggie scanner breaks test_catalog_partition_matches_canon.

Verification (on 0b924e7, rebased on main at aa6fc79)

• validate.py, validate_catalogs.py, landscape.py --root . (844 repositories), catalog_decisions.py --check and evidence_manifest.py --check all pass.
• verdict_review_gate.py --base origin/main reports "no verdict rows changed".
• 282 unit tests pass in the touched areas.
• The only failure is the host-only test_effort_default_guard.test_shipped_guard_is_verbatim (a stale ~/.claude hook).

Before freeze

• A re-review of the fixed draft.
• A 2017-2019 cost re-measure and fees dated for 2017-2020.
• Direct reads of four citations still marked unverified.

🤖 Generated with Claude Code

---

Update, 2026-09-24 ~09:00Z: reduced to a core; not freeze-ready

• Full draft. Three more pre-outcome review rounds on the full 35-item draft made 75 fixes but did not converge: 4, then 3, then 4 high findings. protocol-draft.json is now superseded_by_core_draft_20260924 and is kept unchanged as the record.

• Core draft. protocol-core-draft.json covers H1 (MAX/lottery reversal: the high minus low MAX21 tercile over a 5-session hold, plus the low-tercile tradable cell) and H3 (the intraday leg, the overnight leg, and the paired overnight-minus-intraday statistic). That is 5 items in one population with Holm over m = 5. It also fixes:

  • S1: every stage runs from one frozen study/ tree plus a runtime lock, and the evaluator refuses any other tree;
  • S2: a terminal rule for symbols that stop quoting;
  • S3: asof is the decision session, and the universe is enumerated from the asset master plus corporate actions.

• Deferred. H2, H4, H5, H6 and population I move to the README appendix "Deferred to v3.1+", each with the open findings that apply to it.

• Core review, 2 rounds. 2 high findings remain, both implementation-level:

  • R8-1: a transport deviation under study/fetch/ could change stamp derivation;
  • E1: the pinned coverage.dedupe_identity cannot run under the protocol's import rule.

  Conclusion: the protocol text will not converge further until the study code exists. The next step is code-first preregistration: implement study/ with synthetic tests, then freeze the code and protocol together before any data read, as Mover studies: preregistered v1 and v2 (no pass in any family), data corrections D1-D6 #162 did.

• Bot threads. All 9 Codex threads were handled: fixed, already addressed, or made obsolete by the deferral. Each has a reply and is resolved.

This PR lands the research plan and drafts as guidance only. Nothing here is frozen, and no data was read.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-24T07:14:42.554434Z	0b924e7	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0b924e761f

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".