The Lab, End to End

This is the five-minute map from an idea to a tiny, kill-switch-guarded live order.

• Current truth: Two equity candidates have earned pre-registration; neither has opened the 2022+ lockbox or earned trading authority.
• Operating principle: The lab optimizes for trustworthy decisions, not attractive backtests.

What this is

This is a discovery-to-live pipeline for turning owned data into small, supervised real-money experiments.

• Discovery: Pre-registered strategy research runs on owned data before a result is visible.
• Validation: Survivors face frozen confirm-cards, a sealed lockbox, and a separately sealed risk policy.
• Execution: A passing candidate can reach Alpaca only through rehearsed plan → submit → reconcile chains and an always-on kill-switch daemon.
• North star: Build realized net returns from a portfolio of small, uncorrelated edges; the honest mature target is about 1–3%/month net if the survivors hold up.
• Not the thesis: There is no magic strategy, single heroic backtest, or automatic path from a promising screen to capital.

Why the numbers can be trusted

The machinery is designed to make self-deception visible before capital is exposed.

• Trial ledger: The campaign ledger has 57 physical JSONL rows; each strategy look was registered before evaluation with at most one declared variant.
• Numbering caveat: The ledger contains 49 numbered registrations across 1–42 and 44–50, plus an unnumbered recency-gate method amendment with trial:null; no source row exists for trial 43.
• Lockbox: Screening uses 1998–2021 only; 2022-01-01→2026-08-21 stays sealed for one opening, planned no later than 3 September.
• One-shot rule: Only the two frozen preregistrations may use that opening; a failed gate means refusal, not repair and retry.
• Recency gate: New ideas do not advance unless their 2017–2021 result is positive and not materially worse on its own.
• Costs: Every equity screen charges actual turnover at 10 bp per side for current-metadata Mid+ names and 25 bp otherwise, then stress-tests higher costs.
• Statistics: Overlap-aware Newey–West t-statistics, seeded permutation nulls, multiplicity control, negative controls, and tail-removal checks challenge each headline.
• Confirm-cards: Frozen execution re-derives every apparent survivor; they corrected gap fade from +96.60%/yr to −56.17%/yr and overnight premium from +25.52%/yr to −4.20%/yr.
• Independent challenge: Six read-only reviewers audited the field guide and live/statistics boundaries; live-lane acceptance also exercises adversarial mutation, recovery, and refusal paths.

What survived

Two complementary equity sleeves earned one-shot preregistrations, not permission to trade.

• Diversification: Fundable momentum and GP/A active-return correlation is about 0.20.
• Capacity: GP/A’s median binding sleeve at 1% of ADV is $27.3m; that is a modeled capacity estimate, not an observed fill guarantee.
• Shelf candidate: Cash operating profitability returned +5.00%/yr with NW t 3.11 in 2017–2021, but lost the identical-rule head-to-head to GP/A.
• Why shelved: COP returned +2.06%/yr with t 1.74 versus GP/A’s +2.77%/yr with t 2.92; correlation was 0.36, making COP a complement and the first post-live incubation candidate.

What was killed

We kill things in public — that is the product.

• Campaign balance: Roughly 46 numbered slots ended dead, weak, invalid, inconclusive, refuted, diagnostic-only, or conditional on data and venues the current lane does not have.
• Gap fade: Same-auction look-ahead ranked the official open and booked that open-to-close return; frozen next-open execution produced −56.17%/yr.
• Overnight premium: The screen entered at an observed close and churned about 39.9 round trips/month; corrected execution produced −4.20%/yr.
• Insider clusters: Promising raw event studies disappeared under pessimistic filing availability; the fundable confirm-card returned −3.30%/yr with NW t −1.09.
• 13F: Abnormal ownership returned −4.48%/yr net excess with NW t −1.22 and only 35 executable monthly observations.
• PEAD: The filing-date proxy was weak and cannot prove announcement-time availability; after-close filings can contaminate same-close decisions.
• Crypto momentum: The historical top-50 universe was selected with 2021 hindsight; the Alpaca subset stayed interesting but the confirm-card was inconclusive.
• FX carry: The proxy screen survived, but Alpaca has no executable venue and the estate has no matched-tenor forward points; it is conditional on an IBKR rerun.
• Classic fundamentals: Asset growth, Piotroski F-score, and EV/EBITDA failed the prospective 2017–2021 recency gate.
• Net share issuance: The anomaly was real at +6.18%/yr long/short with t 3.58, but the long-only capture was only +0.45%/yr with t 0.34; the alpha lives on the unusable short side.
• Grid bots: The viral $105k/week story decomposed into leveraged trend beta plus a grid overlay; 89.8% of bot-months broke range, the 5× insolvency proxy hit 29.8%, and full bots lost −371.1%/yr or −34.2%/yr at Alpaca-like fees.

The recency answer

The 2017–2021 lens asked whether the edge survived the market regime an outsider actually recognizes.

• Coverage: Trial 42 re-scored all 34 canonical saved strategy lineages on 2017–2021 alone with Holm family-wise α=0.10.
• Candidates: Volatility-scaled momentum improved from +6.75%/yr to +7.53%/yr; GP/A improved from +2.83%/yr to +5.05%/yr.
• No resurrection: Gap fade and overnight premium mechanically passed Holm only because the saved series retained the defects later caught by their confirm-cards; zero killed strategies became credible again.
• Decay: Eight equity anomaly rows were measurably weaker; FX carry was also flagged decayed, making nine decayed rows across the complete 34-lineage table.

What is built

The research estate and the thin live lane now form a usable machine, while every authority boundary still fails closed.

• Owned equity history: Sharadar SEP contains 45,324,177 verified rows spanning 1997-12-31→2026-08-21.
• Fundamentals estate: SF1, SF2, SF3, EVENTS, and DAILY sit with SEP evidence in a 2,741,514,240-byte, 2.742 GB decimal DuckDB store.
• Market data: The realtime Alpaca SIP ingestor supports an exact one-to-five-symbol set, with current-generation per-symbol progress and stale-stream monitoring as its required operating proof.
• Daily research: A content-hashed five-stock SIP schedule, source publisher, dispatcher, producer, verifier, and reviewer define the continuous production lane; receipts, not cron configuration, prove activity.
• LL-1/2/3: One isolated live lane now owns deterministic planning, Alpaca opg submission, exact reconciliation, and the sealed transition from fixtures to paper to live authority.
• Hard-action coordinator: Every hard breach uses one fenced cancel → resnapshot → flatten → two-zero-proof path; incomplete cleanup remains durable and blocks new strategy work.
• Kill-switch daemon: The default 30-second loop checks policy, account, positions, orders, marked equity, calendar/data freshness, and pending hard actions; a stale heartbeat blocks submission.
• Graduation gate: Live construction, planning, daemon start, and submit refuse unless the account-specific chain is sizing_approval → rehearsal_receipt → live_graduation → post-graduation arm.
• Risk policy v1: Hard kills cover day loss, high-water drawdown, any bounded reconciliation/data failure, a short, or insufficient buying power.
• Soft policy: Scheduled underperformance below the Monte Carlo p05 envelope cuts exposure to 25% of the approved target; a second consecutive breach retires the sleeve.

The timeline

The first possible live order is 15 September, and every date before it produces required evidence.

• By 3 September: Open the lockbox once for the two preregistered candidates, bind the owner-approved risk event, refresh and verify data/calendar manifests, and seal the paper arm.
• 4 and 8–11 September: Run five paper execution sessions, each with the exact plan → submit → matched reconcile chain; 7 September is the NYSE holiday.
• After 11 September: Seal the rehearsal receipt only after the final post-open reconciliation, the hard-kill drill, two-zero proof, zero unresolved incidents, and at least 95% daemon-heartbeat coverage.
• From 12 September: Seal live_graduation no earlier than this date; its effective date and first exchange session remain 15 September.
• 15 September: Begin at stage 10 bp of NAV; for a £100,000 NAV example that is £100, while the owner’s planned first tiny order is about $10 after whole-share and cap constraints.
• Sizing ladder: Only 10→25→50 bp exists; each larger stage needs new evidence, a prior-stage review, and a separate owner seal.

Doing this together

A partner expands capacity and the opportunity set without changing a single evidence gate.

• Capital: More capital scales only after the same review ladder; GP/A alone models about $27.3m of median binding capacity at 1% ADV.
• Short venue: A short-capable broker unlocks shelved edges such as net share issuance whose evidence lives on the short side.
• IBKR: Executable spot-FX financing or matched one-month forward bid/ask data unlocks the conditional FX-carry rerun.
• Intraday data: A roughly $99/month minute-bar feed unlocks the preregistered intraday battery and execution-quality work.
• New ideas: Every partner idea enters the same ledger → screening → confirm-card → lockbox → rehearsal pipeline; there are no exceptions, including ours.
• Replication: Independent reproduction is welcome; the campaign record below names the source artifacts and SHA-256 hashes behind its key claims.

Reading map

These records carry the detail behind every claim on this page.

• Method: Backtesting: The Complete Field Guide.
• Candidate #1: Momentum 12−1 long-only v1 preregistration.
• Candidate #2: Gross profitability long-only v1 preregistration.
• Risk: Live risk policy v1.
• Execution: LL-3 graduation and always-on safety runbook.
• Campaign: Discovery Campaign: 50 Trials, Two Survivors.
• Objective: Edge-discovery north star.

Provenance

The overview is a reading map over retained records, not a new performance calculation.

• Campaign ledger: [local source path] — 57 rows — SHA-256 c1bb3c7cfdf6b854bbee2b1d8156b1b09cfbe7b554629221b7c04b9fa65ad0f5`.
• Modern lens: [local source path] — SHA-256 5bf3fe390eadce6e0b29c7de6334b1345ad96e2011ccc26cb5af846e924229e2`.
• Grid decomposition: [local source path] — SHA-256 50e5270b2a43491ccb6d9afeb9b7794928e58d99cef663ab0ec2eae8e0784c42`.
• Repository record: The linked field guide, preregistrations, risk policy, LL-3 runbook, campaign record, and north-star record are the durable public interpretation of session-local scratch evidence.