00 / The glass box

Kept in the open.

Including the strategies that failed, with their real numbers.

Glass box Every number read from a real artifact

Every figure here is read straight from a real engine artifact, never typed by hand, down to the strategies we killed and the test we did not pass.

  • Zero fabrication. If a value is not available, the element is omitted, never invented.
  • No fake history. The live track record accrues from go-live. It claims no history it did not live.
  • Failures published. Killed strategies, and the overfitting test crypto alone failed, in full.
  • Reproducible. The backtest is content-hashed and reproduces byte-for-byte: re-run the engine, get the same numbers.

9 killed 3 survived Failed deflation test Day 47 accruing

Scroll

01 / Live Track Record

We publish no return until it is earned in the open.

The live paper record begins at go-live and is shown only as it accrues. We publish no return until it is earned in the open.

Status LIVE
Day 47 days accruing
Go-live
2026-09-24
Live NAV
$98,117
Live return
-1.88%
Baseline
$100,000

paper-state.json broker-derived composite marks (three Alpaca paper accounts + local AlphaForge PaperBroker crypto)

02 / The Kill Log

Most ideas die. We publish ours, with their real numbers.

Most ideas die. We publish ours with their real net-of-cost numbers. 12 killed at the deployed-path gauntlet, 2 at a pre-registered research gauntlet, 32 more at the cheap prototype screen, 3 survived.

// Gauntlet numbers are read from the walk-forward summary.txt of a real engine backtest, net of the conservative cost model. The screen-stage kills (labelled 'screen_prototype') are standalone screens - free-data prototypes AND screens on our Sharadar/Deribit research lakes - each run through a four-test robustness gauntlet but BEFORE full deployed-path engine integration; their numbers come from those screens, NOT the deployed engine, and we say so. The 2026-06-27 batch is a 17-agent adversarial sleeve-discovery campaign that killed all 10 candidates it tested. The 2026-07-09 batch is three probes from a sweep of the modern-sleeve frontier (futures carry, mechanical rebalancing flows, prediction-market odds as a signal) - all three killed at screen. The 2026-07-11 batch is the AlphaMax improvement campaign: six documented momentum enhancements tested against the live construction, all null or worse - the frozen 12-1 build stays. The 2026-07-12 batch closes that campaign (the 52-week-high variant, run to a full walk-forward to honor a fired pre-registration) and adds three modern-frontier probes: two free intraday-flow studies (leveraged-ETF end-of-day rebalance flow, intraday flow-momentum - both killed at screen) and the economic-trend sleeve (macro-fundamental trend, one pre-registered walk-forward after a 3-auditor leakage panel signed off - a real crisis-alpha personality but too little edge, DSR 0.00). The 2026-08-02 batch is the AlphaMax construction-axis campaign, opened after the July 2026 momentum junk-squeeze: three ways to change how the equity book is BUILT, each leaving the 12-1 signal byte-identical (beta-neutral construction, a 4x4 weighting-by-breadth grid, nine short-leg tail controls). All three null, nothing adopted, zero trial slots burned - each was a read-only screen with its gate written down first, and in the grid's case that pre-registered family-wise gate is what stopped the best-looking cell from being harvested (White Reality Check p=0.315). The beta-neutral candidate was killed by measuring its own premise: the hidden negative market beta it was built to remove does not exist (research -0.014, live +0.319). A few entries are labelled 'deployed_gauntlet': those went through the FULL deployed-path walk-forward, so their numbers ARE the real engine's; they sit in this list only because their artifacts live in probe directories. A 'research_gauntlet' is a locked, costed OOS probe with complete return artifacts but not the production engine path. Nothing is re-tuned.

Select a row to see why it lived or died. Killed sleeves leave the book and are not re-tuned.

03 / Pre-Registration

We committed the trial budget before we touched the data.

The trial budget was committed to git before any sleeve touched the data lake. Every config that could have been measured counts toward the deflation denominator, so a tiny budget measured once is the only honest path to a clean grade.

Trial budget, hard ceiling
9
Document
docs/design/PRE_REGISTRATION.md

Measure-once protocol Budget and test design were fixed before the data was touched. The final count cannot be cherry-picked. No re-runs, no re-windowing, no cadence search, no sign-flips, no subset search over sleeve combinations, no MVO, no static scheme in the deployable path. KILL = exclude, never re-tune.

Expected B+ to A-, honest combined Sharpe ~0.75-0.85 net (NOT 1.46)

Actual C+ honest null. The free-breadth path is exhausted: value and quality do not replicate on the narrow universe; crypto-perp alone fails honest deflation. The deployed book is the two survivors, awaiting forward live proof.

  1. 1
    Momentum (12-1)Deployed as the frozen 2023+ k30_dn_63 sleeve at Sharpe 0.9071. The deep-history pre-reg variant scored -0.0493 and failed the DSR gate; the clean sleeve is used.Deployed Sharpe 0.9071
    PASSED
  2. 2
    Value (composite)Net Sharpe -0.5994 below the 0.30 gate. Killed.Measured Sharpe -0.5994
    FAILED
  3. 3
    Quality (GP/A + ROE)Net Sharpe -0.8316 below the 0.30 gate. Killed.Measured Sharpe -0.8316
    FAILED
  4. 4
    Betting-Against-Beta (low-risk)Net Sharpe -0.0679, failed the DSR >= 0.95 gate. Killed.Measured Sharpe -0.0679
    FAILED
  5. 5
    Short-term reversal (1-month)Dropped pre-registration: 20x turnover capacity liability.
    DROPPED
  6. 6
    Combined book (one fixed combiner)Structure survives (PIT-clean, decorrelated) but the in-sample magnitude fails multiple-testing deflation. Recorded as the honest null; the path is forward live evidence, not a re-combine.
    HONEST-NULL
  7. 7
    Crypto-carry satellite (re-measured once)Net Sharpe 0.6766, kept as a capacity-capped satellite (finite ~$100k to $1M, decays above ~$100M AUM).Measured Sharpe 0.6766
    PASSED
  8. 8-9
    Contingency reservesSpendable only against a structural trigger logged before its result is seen.
    UNSPENT

04 / The Deflation Verdict

Crypto alone did not clear the bar. We publish the null in full.

We graded crypto-perps alone against an honest multiple-testing benchmark. It is noise. We publish the failure as a feature, not a footnote.

Deflated Sharpe (shared) BELOW GATE
0.2112 gate 0.95
Prob. of backtest overfitting OVER GATE
0.8818 gate < 0.20

Capacity decay, annualised Sharpe by deployed capital

  • $100kSR 0.401
  • $1MSR 0.042
  • $10MSR -0.372
Failed Deflation test

NO-DEPLOY (honest null)

Best config DSR 0.2112 < 0.95 gate AND PBO 0.8818 > 0.20 gate. No variant beats baseline and clears deflation simultaneously. Crypto-perps alone is a thin cross-sectional signal that does not survive honest deflation net of cost; the equity sleeve is required as ballast.

05 / Capacity & Scalability

Size has a price. The crypto carry caps near $10M.

How much capital this book can run before the edge decays, told straight. The crypto carry sleeve is MEASURED finite-capacity alpha: its realized Sharpe goes negative by $10M as market impact consumes the thin signal. The equity momentum core scales much further (MODELED), so above ~$100M AUM the book IS the equity core. Forward Sharpe is the deflated 0.3-0.9 expectation (lowered from 0.7-1.0 on 2026-07-29; this artifact carried the superseded band until 2026-08-06), never the in-sample headline.

Forward book Sharpe by deployed capital AUM (log) to deflated forward Sharpe
Crypto carry capacity $10M MEASURED + STRUCTURAL

Crypto-carry capacity is approximately $10M. The MEASURED sweep shows sr_ann decaying 0.4009 ($100k) -> 0.0424 ($1M) -> -0.3720 ($10M) as notional/ADV cost scaling lifts realized transaction cost until the thin edge is consumed. Above the cap the satellite's DEPLOYED weight decays to zero (capacity-proportional, PRE_REGISTRATION.md l.82), scale-honest, never extrapolated as positive.

Equity-core threshold $100M STRUCTURAL

Above ~$100M AUM the book IS the equity core (CROSS_ASSET_BOOK.md l.85). The crypto satellite weight is zero past its cliff; what remains is the equity momentum portfolio scaled for ~14% vol within the Reg-T 2.0x gross cap.

  • $1M0.70 to 0.80FORWARD5.6 to 6.4%-2.5%crypto 50% / equity 50%Cost negligible on the top-200 anchor; both sleeves live. Crypto carry is 100% of the proven LIVE edge; equity is the MODELED ballast (not yet live).
  • $10M0.50 to 0.55FORWARD4.0 to 4.4%-3.5%equity coreCrypto capacity cliff: at $10M the MEASURED crypto sr_ann is -0.3720 (impact-dead); its book weight decays to zero by the capacity-proportional satellite rule. Equity impact still well under the 1% ADV pre-trade cap.
  • $100M0.45 to 0.50FORWARD3.6 to 4.0%-5.0%equity coreAbove ~$100M AUM the book IS the equity core (CROSS_ASSET_BOOK.md l.85); crypto satellite weight = 0. Per-name participation still <0.5% ADV on the top-200 anchor; Reg-T 2.0x equity gross cap is active.
  • $1B0.30 to 0.40FORWARD2.4 to 3.2%-10.0%equity corePure equity momentum core. Per-name participation rises toward the 1% ADV pre-trade cap on many names; the portfolio must tilt to the most-liquid anchors to stay inside it. Reg-T 2.0x gross cap is the hard limit. 10-15% MODELED impact drag if forced into mid-caps; the liquid-anchor cost-model drag alone is far smaller (see equity_sleeve).

// Every figure is labelled MEASURED (read from a real artifact), MODELED (derived from the committed transaction-cost model), or FORWARD (the deflated honest expectation). Crypto capacity is MEASURED at three AUM points (verdict.md). Equity capacity is MODELED: the pre-registered equity sleeves are not yet live, so its decay is the cost model's sqrt-impact law on a stated ADV anchor, not a measurement. Nothing between/above measured points is dressed up as measured.

Deflated forward Sharpe
0.40 to 1.20 · central 0.70
Net return, 14% vol-target
13-16%

// Our commitment about all of this, signed so we cannot quietly walk it back:

  • We will never advertise a capacity figure above our own published capacity model (glassbox/capacity.json). The model is the ceiling, not marketing.
  • Crypto funding carry is MEASURED finite-capacity alpha: realized Sharpe decays from 0.40 at $100k to 0.04 at $1M and turns negative by $10M. We commit to not scaling the crypto sleeve past the point our own measured sweep shows the edge largely consumed, and never past its ~$10M structural cliff. We would rather cap the sleeve than collect a fee on a dead edge.
  • At every capital milestone we will publish the REALIZED capacity decay beside the modeled curve, so the gap between model and reality is visible, never hidden.
  • Above ~$100M the book IS the equity core (the crypto satellite weight is zero past its cliff). We will say so plainly rather than imply the crypto edge scales.
  • If we ever close to new capital, we will publish the level and the reason.

Ed25519-signed, key 0c6606a6b0c5b66f… Verify against the published key.

06 / Reproducibility

Content-hashed. Byte-reproducible. Reproduce our numbers.

The backtest is content-hashed and byte-reproducible. A golden-master test asserts every fill price, fee, funding payment and the final equity from hand-written arithmetic to 1e-9 precision. The backtest itself is not timestamp-anchored; the separate live track-record chain has selected OpenTimestamps checkpoints.

Verify us in one command

// Download our published files and our verifier, and check every number yourself. No install, no data, no taking our word: it recomputes each artifact's content hash, re-checks the Ed25519 signatures, and (in the repo) reruns the golden master. 

curl -sO https://canlicapital.com/glassbox/reproduce.py
for f in capacity deflation track_record kill_log pre_registration \
         red_team reproducibility research capacity_commitment; do
  curl -sO https://canlicapital.com/glassbox/$f.json
done
python3 reproduce.py --dir .   # -> ALL CHECKS REPRODUCED

We claim. content-hashed + byte-reproducible backtest

We do not claim. No blockchain timestamp or chain-of-custody claim for the backtest itself

tests/integration/test_golden_master.py 636 lines

Phase-5 FULL acceptance gate (execDesign.md section 4.5)

  1. Part 1: a 3-instrument 30-bar scripted run (opens, an add, a partial close, a long-to-short flip, a gap-drop, mixed 8h/4h funding). EVERY fill price, fee, funding payment and the final equity asserted to abs=1e-9.
  2. Part 2: the engine's local ADV/sigma (LakeCostInputs) asserted numerically EQUAL to the registered feature-library adv_quote_30d / sigma_daily, so the backtester and the feature library never fork the formula.
  3. Part 3: real Binance BTCUSDT 2024-03-01..2024-03-08. The stored funding table must hold exactly 21 settlements (the 8h schedule) and a week-long long must settle exactly those events (rel=1e-12 on each payment).

Verifiability

Fill-price tolerance
abs=1e-9 (pytest.approx)
Funding settlements, real BTC week
21
Registered factors
52
  • kill_log.jsonsha256:b4bd493e4647...08f82a
  • pre_registration.jsonsha256:145d3d07ee56...74d202
  • deflation.jsonsha256:b0a1a527990e...64d410
  • track_record.jsonsha256:2a74c3874600...f7afa5
  • reproducibility.jsonsha256:7e51062fd83b...ee7b0a
  • research.jsonsha256:cd0708826cc0...b4ee50
  • capacity.jsonsha256:f6729ac04883...58317e

Tamper-evident track record

// Every published state is hashed, chained to the one before, and Ed25519-signed. Change any past entry and every signature after it breaks. The chain head is also stamped into the Bitcoin blockchain, a clock we don't control, so we can't backdate it either. That proves when the record existed, never whether the numbers are good. Don't trust us. Verify us.

Signed chain
129 days · 1854 entries · head sha256:36490969bbaa...fdeb96
Public key (Ed25519)
0c6606a6b0c5b66f…
Verify it yourself
uv run python scripts/verify_transparency.py (re-checks each chain hash + sig)
External anchor (Bitcoin)
130 heads · pending Bitcoin (calendar-attested)
Founder skin-in-the-game
$1,000 at first live deploy, signed

// Every weight and leverage applied to a forward day is a pure function of returns strictly before that day, enforced by the PIT reader and a truncation-invariance test.

How to verify this yourself Every measurement, rendered in full The full methodology Join the waitlist