 █████  ███████  █████  ██   ██  █████  ███████  █████
██   ██ ██      ██   ██ ██  ██  ██   ██ ██      ██   ██
███████ ███████ ███████ █████   ███████ ███████ ███████
██   ██      ██ ██   ██ ██  ██  ██   ██      ██ ██   ██
██   ██ ███████ ██   ██ ██   ██ ██   ██ ███████ ██   ██

Asakasa · Applied Intelligence Lab · Bucharest

An applied intelligence lab building high-frequency market-research software.

Asakasa SRL develops research software and publishes market research for asymmetric technology regimes and the venues that price them. Raven, the market-intelligence brief, is live; three research-software engines are in active development — Translight for CLOB microstructure, the Decentralized Exchange AMM for delta-neutral perpetuals research, and Yutani for sub-100ms engineered signal-to-submit execution research.

The lab is also available for bespoke engineering — signal infrastructure, execution stacks, regime classifiers, methodology audits. It does not provide investment advice, manage third-party capital, or execute trades on behalf of clients.

Read the latest brief Inspect the engines Read the thesis

Research engines

Four

Raven · Translight · DEX AMM · Yutani

Live data streams

Filings · microstructure · macro · political · on-chain

Execution research

Python + Rust

Hybrid brain/executor over IPC · EIP-712 signing

Deployment posture

Phased & gated

Read-only → testnet → canary → bounded ramp

RAVEN · RESEARCH TELEMETRY LIVE TICK · representative

RAVEN CORE · SOURCE FUSION 6 CLUSTERS · 27 STREAMS

Source-cluster fusion

Six source clusters orbit a fusion core. Dot color and label color reflect the current health regime of that cluster. The lab will not act on a candidate read while the cluster supporting it is amber or red.

Regime: rotation · breadth=watch · vol_term=normal
Source health (snapshot): 22 green · 5 yellow · 0 red
Median ingest latency: 1.2s · p95 6.4s

EDGAR 2/4 green

OPTIONS 2/2 green

FINRA 0/2 green · 2 yellow

BREADTH 6/6 green

GEOPOL 4/4 green

POL-CAP 1/1 green · slow

CORE · regime=rotation · 22g / 5y / 0r · static example

Lab cycle telemetry · representative

Operational shape across the lab's engines.

The tail at right is a representative replay of what the lab's services emit when running — Raven scheduler cycles, decentralized-exchange observability heartbeats with perp/spot basis and funding, Yutani microsecond brain ticks, Translight quote-loop state, prediction-market flow-scanner restraint events. Line shapes are real; identifiers, paths, credentials, and quantitative outputs intentionally synthesised.

Research operations · every action human-approved

ssh ubuntu@vps-eu-1 $ journalctl -f -u raven,translight,hl-intel,yutani --output=cat

LIVE lines 0 last — warn 0 err 0 uptime 21d 9h

operator@vps-eu-1:~/asakasa$

Publishing cadence · US market sessions

Plan the day

Pre-market

before 09:30 ET

Overnight tape, premarket movers, event risk for the day ahead.

Trade the tape

Intraday

09:30 – 16:00 ET

Session reads through the day, plus a one-hour-before-close interrogation.

After-Action Review

Closing session

after 16:00 ET

After-hours close: regime composite, bottom-watch, source-health roll-up.

24/7 prediction-market flow scanner — a separate signal stream that monitors public on-chain prediction-market flow for high-conviction, wallet-concentrated, informed positioning, and emails a transition alert only when it clears a high bar. Most cycles publish nothing.

Raven Intelligence Brief · invite-only

A market-intelligence brief with restraint built in.

Raven is the lab's market-intelligence brief: a high-cadence scan of regime, flow, insiders, event risk, prediction-market pressure, and civilizational-technology bottlenecks. It is designed to say "wait" more often than it says "act."

Every input is source-tagged — aggregated mainstream reporting, official filings, exchange/SEC/FINRA/FRED/Cboe data, and Tier-A primaries only; no social-media or pundit feeds, and descriptive language throughout, never a trade instruction. It is invite-only and not commercially sold — a working research programme, by request.

Request research access

Sample close brief surface

Illustrative · static sample

Regime

WAIT / PATIENCE

0/5 core bottom triggers. Breadth washout active but vol curve remains contained. Cash retention.

Event risk

NVDA earnings inside 24h

AI beta and semis gamma catalyst. Options flow on NVDA is read as earnings-vol positioning before directional conviction.

PMKT flow

Probability move + wallet concentration

Large-trade pressure and top-wallet share are surfaced with the actual market question, not just an opaque score.

Included lenses

Macro

Oil → inflation → rates and duration → equities and semis, synthesised into one read with a per-theme stance: energy, defence, duration, semis.

Regime

Breadth, vol term, rates/credit, cross-asset stress — stress-gated, so the headline cannot read risk-on while the rates and oil tape contradict it.

Flow

Options V/OI, strike distance, DTE pressure, PMKT wallet concentration, large-trade pressure.

Earnings

Mega-cap beats and misses graded noise-vs-signal against the expected-move band; prints too fresh to judge are flagged reading-pending.

Insiders

EDGAR Form 4 materiality, owner role, timing, cluster corroboration, political-capital overlays.

Constraints

AI compute, power, memory, semicap, defence procurement, space and energy bottlenecks.

Overnight & event risk

Asia and Europe cash sessions, each dated, plus policy, geopolitics, sanctions, supply routes, and forward event-risk boards.

Latest brief · published

The 19 May close, as the lab actually rendered it.

A real Raven close brief, published verbatim from the lab's working repository. Specific options strikes, position sizing, scoring methodology, and edge-revealing per-name analytics are redacted. The framing, regime read, source-health roll-up, event-risk callout, and methodology note are exactly what the engine produced.

SESSIONS: pre-market · intraday
closing (US market)
FORMAT: terminal-rendered markdown
DELIVERY: email + web archive
ACCESS: invite-only · by request
not commercially sold

$ ls -t /var/reports/raven/

2026-05-19-close.md after-hours close · regime mixed open ↓
2026-05-17-weekend.md weekend holistic recap invite
2026-05-15-intraday.md intraday session read invite
2026-05-13-session.md eight intraday runs invite

Most recent shown below. Prior briefs go to the invite list; the dates are the published record of cadence.

Request access

ssh ubuntu@vps-eu-1 $ less /var/reports/raven/2026-05-19-close.md

-rw-r--r-- ubuntu ubuntu 2.4K May 19 20:32 2026-05-19-close.md

# Raven Close Brief — 2026-05-19

Asakasa Applied Intelligence Lab · Bucharest
Generated 2026-05-19T20:30 UTC · After-hours session
Engine: raven · template raven-intel v0.8

> Disclaimer. General market research and commentary, published for
> information purposes only. Asakasa SRL does not provide personalised
> investment advice, does not manage third-party capital, and does
> not solicit transactions in specific securities.
> Readers act on their own analysis.

## Read

  Regime: mixed   ·   Action:  HOLD / WAIT    ·   Cash retained.

  Not a clean buy, not a panic sell. Composite regime is mixed across
  the six axes the lab tracks. Bottom-watch panel reads 0/5 triggers
  firing. The framework prescribes patience on regimes that look like
  this, not entry.

## Tape — regular-session close

  SPY    −0.67%   5-day −0.60%
  QQQ    −0.62%
  XLK    −0.64%
  VIX    18.06   (+1.35%) — contained

  Leaders:   XLE +1.17%   XLV +1.10%   (defensives, energy bid)
  Laggards:  XLF −1.24%   XLB −2.35%   (financials, materials soft)

  Crash-anatomy classifier: intra-market rotation (sell tech, buy
                            defensives); volatility contained.
  McClellan oscillator:     −156.8 — breadth washout, no recovery
                            confirmation.
  Insider buys actionable (Form-4, ≥$1M, 7d):  0

## Regime axes

  ┌───────────────┬────────────┬────────────────────────────┐
  │ Axis          │ State      │ Evidence                   │
  ├───────────────┼────────────┼────────────────────────────┤
  │ Volatility    │ supportive │ normal contango            │
  │ Rates/credit  │ adverse    │ tightening + curve steep   │
  │ Breadth       │ neutral    │ mixed internals            │
  │ Cross-asset   │ neutral    │ —                          │
  │ Geopolitics   │ neutral    │ no HIGH-tier event in 24h  │
  │ Tape stress   │ supportive │ 0/5 bottom triggers        │
  └───────────────┴────────────┴────────────────────────────┘

## Event risk — inside 24 hours

  NVIDIA Q1 FY2027 results — Wed 20 May 2026.
  Release ~13:20 PT, conference call 14:00 PT. AI-beta options
  surface across the watchlist (multiple names with bullish call
  accumulation flags) is best read as event-vol positioning into
  the print, not a directional signal in either direction.
  Source: NVIDIA Investor Relations.

  Geopolitical scheduled-event window (next 7d): 1 LOW-tier event.
  No HIGH-tier event in the 24-hour window.

## Methodology change — this cycle

  Earlier in the day the engine surfaced a far-OTM short-dated call
  as its "top conviction" line. That output was a category-mixing
  artefact — high V/OI on short-dated far-OTM strikes during an
  earnings window can be event-vol or lottery-gamma flow rather
  than informed positioning, but the ranker did not separate the
  categories.

  Raven now distinguishes actionable conviction from exploratory /
  event-vol / lottery-gamma flow. Earnings-adjacent OTM gamma
  will no longer hijack the headline.

## Geopolitical context — top filings, last 24h

  3/5  un_news       Ukraine war 'becoming deadlier by the day'
  2/5  ft_world      Trump claims Xi will not arm Iran
  2/5  ft_world      Putin and Xi to discuss huge gas project
  2/5  ft_world_asia Japan and Korea deepen ties (energy)
  2/5  ft_world_uk   EU braces for fertiliser squeeze (Iran war)
  2/5  ft_world_us   Iran stockpiles oil on ageing tankers
  …four more, filtered against operator themes:
  defence · energy · semi · AI · civilizational_tech · nuclear

## Source health

  22 GREEN · 5 YELLOW · 0 RED     median 1.2s · p95 6.4s

  Yellow this cycle: edgar_form4_backfill, finra_otc_transparency,
  finra_short_volume, geopol_daemon, trendforce_dram.
  (Mostly off-hours empty-window valid-empty renders, not failures.)

## Sources used

  yfinance EOD: macro / sector / watchlist / breadth / vol term / xa
  Cboe-published options chains via yfinance
  SEC EDGAR Form-4 atom + 7-day backfill
  FINRA CNMS daily short-volume · FINRA OTC Transparency weekly ATS
  TrendForce DRAM pricing · FRED DGS2 + Treasury yields
  BLS / Treasury / FOMC / ECB / BOJ release schedules
  NASDAQ / SEC EDGAR S-1 IPO calendars
  Tier-A primary: state.gov · whitehouse.gov · un.org · nato.int · csis
  Tier-B mainstream: FT · Reuters · AP

--
Asakasa SRL · Bucharest, Romania · research-only artefact.
No personalised advice. No third-party capital. No client execution.
No solicitation. Readers act on their own analysis.

Lab engines · one live, three in development

Raven is live. The other three are research surfaces under active development.

These are not products available for purchase. They are the lab's working research surfaces — internal artifacts that exist so the lab can study venues, microstructure, and methodology under realistic constraints. The status label on each card is its true engineering state, not marketing.

OPERATIONAL · INTRADAY CADENCE

Raven

Market intelligence layer

Twenty-seven live sources, synthesised into decision-ready briefs across three US-market sessions a day.

Connects the tape into one read — oil, inflation, rates and duration through to equities and semis — with a per-theme stance across energy, defence, duration, and semis. Bayesian regime classification runs behind a stress gate: the headline regime cannot read risk-on while rates and oil contradict it. Earnings reactions are graded noise-vs-signal against an expected-move band, and the overnight Asia and Europe tape is carried, each session dated. Source-health tracked per cycle, per feed.

Bayesian inference · regime taxonomy · source-health discipline

Research portfolio · in active development

Three pre-production research surfaces. Each carries an honest engineering state — engineering-validation, read-only ingestion, or methodology pivot. None is trading live capital today.

ENGINEERING VALIDATION · DRY MODE

Translight

Polymarket V2 CLOB market-making research

Continuous quoting research engine on the Polymarket Central Limit Order Book.

Market-making research engine for the Polymarket V2 CLOB. Studies passive liquidity provision, queue dynamics, adverse selection, and inventory management under venue-specific microstructure. Costs (fees, gas, slippage) modeled as first-class constraints. Risk envelopes and drawdown halts wired upstream of every quote.

Research software. The SRL does not execute trades for clients, take third-party capital, or operate as an investment-services provider.

Almgren-Yamamoto inventory control · queue modeling · adverse-selection screening

PHASED DEPLOYMENT · READ-ONLY INGESTION

Decentralized Exchange Automated Market Maker

Delta-neutral perpetuals market-making research

Passive liquidity research on decentralized-exchange perpetuals — earning from microstructure, not direction.

Delta-neutral / near-delta-neutral market-making research engine for decentralized-exchange perpetual futures. Studies maker-rebate capture, funding-rate carry on hedged inventory, perp-spot basis, and on-chain liquidity-provider vault comparables. Explicitly not a directional speculator: the lab pivoted to passive liquidity provision after the directional hypothesis underlying Yutani was falsified by prospective evidence.

Research software. The SRL does not execute trades for clients, take third-party capital, or operate as an investment-services provider.

delta-neutral hedging · funding carry · basis research · on-chain protocol mechanics

METHODOLOGY PIVOT · LESSONS PUBLISHED INTERNALLY

Yutani

Sub-100ms execution research

Python brain + Rust executor over loopback IPC. Engineered for venue-floor signal-to-submit latency.

Hybrid execution research engine targeting the physical floor of signal-to-eligible-to-match latency on Polymarket binary markets. Python ingests multi-venue spot, depth, trades, and oracle data, builds a multi-timeframe feature vector, fuses signals (Bayesian), and emits decisions. Rust executor signs EIP-712 with secp256k1 and submits via keep-alive HTTP. Pre-registered hypothesis; directional bucket falsified at R12, methodology pivoted at R14 to theory-driven gates under sequential probability testing.

Research software. The SRL does not execute trades for clients, take third-party capital, or operate as an investment-services provider.

Where Yutani is run against live venues, it is operated by the principal on personal account, outside the SRL — the lab's role is to develop the software, not to trade it.

Python+Rust IPC · EIP-712 signing · sub-100ms target · SPRT discipline

Executor data-flow · mathematical basis

p̂ = Σ wᵢ · pᵢ

BRAIN python

fuse · sub-500µs

HMAC · TCP

IPC localhost

≤ 2 ms

Decision{side, p̂, sz}

EXECUTOR rust

no JSON re-parse

sign(domain, msg)

EIP-712 secp256k1

≤ 5 ms

POST /order

CLOB http/1.1

≤ 50 ms ack

Each stage carries a pre-declared time budget. The aggregate signal-to-submit floor is bounded by the stages the engine controls; venue-side latency (a 250 ms taker hold) is treated as an invariant, not an engine cost.

Architecture · execution research

A Python brain and a Rust executor, separated by a single IPC boundary.

Yutani is the lab's execution-research benchmark. A Python signal pipeline ingests venue data, builds a multi-timeframe feature vector, fuses signals under Bayesian weights, and emits a typed Decision{side, p_hat, size}. A Rust executor receives the decision over loopback IPC, signs EIP-712 with secp256k1, and submits with escalating aggressiveness against an explicit per-stage time budget. Any live-venue operation — including every latency and fill figure measured here — is conducted by the principal on personal account, outside Asakasa SRL; the lab develops the software and does not itself trade. The architecture is the same one the lab offers as a reference for bespoke engineering work.

Measured live · not backtested

Every latency, fill-probability, and slippage figure the lab reports is measured against live venue state — never historical replay. Backtests on synthetic order books systematically overstate fill probability and understate adverse selection; in microstructure and HFT research they are not indicative of real execution. The lab treats live-market measurement as the only credible budget, and publishes targets, not realised performance.

Stage	Target	Notes
Tick → decode	< 200 µs	venue WebSocket frame decoded in-process; numpy-vectorised, single-process, no Python-side allocation on the hot path
Feature-vector assembly	< 120 µs	twenty-indicator multi-timeframe vector built in a pre-allocated ring buffer, zero-copy from the tick decoder; no garbage collection on the hot path
Indicator → decision	< 500 µs	Bayesian fusion across the indicator set with fractional-Kelly sizing; emits a typed Decision{side, p̂, size} only when the score clears the gate
Regime-gate evaluation	< 80 µs	the candidate read is checked against the regime envelope it was pre-registered for; signals outside their declared regime are dropped before any sizing or order work
Decision IPC (Python → Rust)	< 2 ms	localhost TCP with HMAC authentication; compact binary protocol, no JSON serialisation on either side of the boundary
Order construction + nonce	< 1 ms	typed-data assembly with monotonic nonce sequencing and a replay guard; deterministic and idempotent per decision id
EIP-712 sign (secp256k1)	< 5 ms	secp256k1 signing in Rust with the EIP-712 typed-data hash precomputed; the executor never re-parses JSON it has already validated
Submit → venue acknowledgment	< 50 ms	HTTP/1.1 keep-alive over a warm connection; the acknowledgment returns order id and queue position, both fed back into markout
Signal → submit (engineered)	< 60 ms	aggregate floor across the engine stages above — the slice the lab controls and measures end-to-end against live venue state, never a backtest
Markout capture (post-fill)	async	t+1s / t+5s / t+30s markout logged per fill for adverse-selection accounting; runs off the hot path and never blocks the next decision
Signal → eligible (incl. venue)	~310 ms	adds the venue-side 250 ms taker hold — a venue invariant, the physical floor for a fill-eligible signal, not an engine cost the lab can optimise away

Targets, not realised performance · measured against live venue state, never a synthetic backtest · stages above the engineered floor are what the lab controls; the venue hold is an invariant, not an engine cost.

Engineering · available for engagement

The lab builds for itself first — and is open to building for others.

The infrastructure below is what the lab builds for its own engines; it is available for bespoke engagement with research teams who want the same. Deliverables are code, documentation, and methodology — scope and terms discussed per inquiry. The lab does not execute trades on behalf of clients, custody client assets, or provide investment advice.

Inquire

Market-signal infrastructure

High-cadence data ingestion pipelines with per-source freshness, latency, and failure-mode tracking. Adapters for regulatory filings, options surfaces, FINRA microstructure, executive disclosures, geopolitical feeds, and on-chain venues.

Execution research stacks

Hybrid Python brain / Rust executor architectures, IPC-bounded latency, EIP-712 signing pipelines, and venue-specific submission state machines, built to the same latency-budget discipline as the lab's own research engines. Asakasa develops and delivers the software; the client operates it under their own regulatory and operational responsibility.

Regime classification frameworks

Bayesian and rule-based classifiers for prevailing market regimes — breadth, vol term-structure, cross-asset correlation, liquidity stress, geopolitical tier — wired to the client's own data sources and decision surfaces.

Methodology audits

Independent review of an existing strategy or research pipeline: hypothesis pre-registration, sample-boundary discipline, multiple-comparison budget, HAC standard errors, walk-forward integrity, cost realism, and invalidation criteria.

Data coverage · 27 sources

Named streams. Tracked freshness. Documented gaps.

Every lab cycle opens with a source-health roll-up across the 27 live sources — each feed's last successful pull, age, median latency, and last failure mode. Streams flagged RED render downstream as valid-empty rather than zero-pretending; gaps are surfaced, not hidden. The clusters below are representative, not the full enumerated list.

GREEN · YELLOW · RED · per source · per cycle

Regulatory disclosure

EDGAR Form 4 (insider)
EDGAR 8-K (material events)
EDGAR 13D/G (block ownership)
NASDAQ IPO calendar
USASpending procurement

Market microstructure

Options surface & flow
Vol term structure
FINRA short-volume
FINRA OTC transparency
Alternative dark-pool feeds
Pre-market movers

Breadth & cross-asset

NYSE/NASDAQ breadth
McClellan oscillator
Sector ETF flows
Sector internals
Cross-asset correlation
Yields & credit spreads

Political risk

OGE Form 278-T executive trades
House committee disclosures
Smart-money convergence (§8c)

On-chain & venues

Prediction-market CLOB depth & trades
Decentralized-exchange perpetuals
On-chain LP-vault comparables
Major spot venues: depth + liquidations

Macro & geopolitical

Economic calendar
Geopolitical calendar (Tier-A)
Geopolitical news (Tier-A/B)
Macro state index
Crypto-macro overlays
Semiconductor capacity (TrendForce)

Methodology · mathematical basis

Rigor is not optional. It is the product.

Each hypothesis across the lab's engines is registered with its statistical apparatus declared up front: estimators, error structure, multiple-comparison budget, sample boundaries, and invalidation criteria. The lab publishes the apparatus, not just the result — and demotes hypotheses that do not survive prospective evidence.

Bayesian regime classification

P(R | D) = P(D | R) · P(R) / P(D) Bayes' theorem

Posterior probabilities over regime states (e.g. trend-persistence vs. rotation, low vs. high vol-of-vol). A candidate read is conditional on the regime it was specified for.

Pre-registered hypotheses

H₀, H₁, α fixed ⟂ D specification precedes data

Hypotheses, the rejection threshold α, and sample boundaries are registered before the data are observed. Promotion requires out-of-sample evidence; demotion is irreversible without re-registration.

Walk-forward validation

Θ̂ₖ = argmin_Θ Σ_{t∈Wₖ} L(yₜ, f(xₜ; Θ)) rolling in-sample fit, out-of-sample test

Parameters are fit on window Wₖ and evaluated on the disjoint next window Wₖ₊₁. In-sample tuning never sees out-of-sample data; boundaries are pre-declared and immutable per registration.

Multiple-comparison correction

α* = α / m ⇒ FWER ≤ α Bonferroni / family-wise error

When m hypotheses are tested, the per-test threshold is divided by m so the probability of any false rejection across the family stays bounded by α. More candidates considered ⇒ a higher bar each.

HAC standard errors

σ²_NW = γ₀ + 2 Σ_{j=1}^{L} (1 − j/(L+1)) γⱼ Newey–West estimator

Heteroskedasticity-and-autocorrelation-consistent variance with Bartlett weights on lag-j autocovariances γⱼ, used wherever time-series autocorrelation is plausible. Clustered errors where group structure is meaningful.

Sequential probability testing

Λₙ = ∏ᵢ p(xᵢ|H₁)/p(xᵢ|H₀); accept H₁ if Λₙ ≥ A Wald's SPRT (1945)

The likelihood ratio accumulates as prospective fills arrive; the hypothesis advances when Λₙ crosses the upper bound A, is demoted below the lower bound B, and otherwise flags HUMAN REVIEW. No fixed-n commitment.

Worked example · Yutani R12-Y

The apparatus turned on the lab's own hypothesis — and killed it.

Most operations show you what worked. This is the lab falsifying its own directional thesis on prospective evidence, then pivoting the methodology rather than the data. It is the clearest single demonstration that the discipline is real and not decorative.

Pre-registered · prospectively tested · demoted on evidence

01 · Registered

Directional bucket R12-Y

A pre-registered hypothesis: a fused multi-indicator signal predicts 5-minute directional outcomes on binary prediction markets, above cost, inside a declared regime. Invalidation criteria fixed in advance.

02 · Falsified

Prospective evidence said no

Under sequential probability testing on live fills (n≈99), the directional edge did not clear the pre-declared threshold. The likelihood ratio crossed toward the null, not the alternative. The bucket was demoted — no quiet re-fit to rescue it.

03 · Pivoted

R14 — methodology, not data

Rather than torture the data into a positive result, the lab pivoted the method: from chasing directional prediction to theory-driven gates and, at the engine level, toward passive-liquidity microstructure where the edge is structural, not predictive.

04 · Gated

SPRT advisory in production

The new gates run under an SPRT advisory that fires HUMAN REVIEW REQUESTED rather than auto-acting on thin evidence. The negative result is published internally as doctrine: a falsified hypothesis is an output, not a failure.

Operating cadence

A research lab proves itself by what it ships, and on what rhythm.

Raven publishes on the US-market clock: a pre-market brief before the open, intraday reads through the session, and a closing-session brief after the bell — with a 24/7 prediction-market flow scanner running alongside. Engineering across the lab's engines carries roughly fifteen commits a day; every working session leaves a dated handoff memo on disk.

Brief generation

Pre-market · intraday · closing

Three US-market sessions per trading day · ET schedule, UTC timestamps

Source health

22 / 27 GREEN

Reported per source, per cycle, with median latency 1.2s

Engineering log

~15 commits / day

Typical working cadence · every session leaves a dated handoff memo

Session discipline

Dated handoff memos

Every working session leaves an on-disk record

Lab notebook · recent entries

A dated record of work across the lab's engines and research surfaces.

Entries are internal lab artifacts — dated research notes, engineering handoff memos, and methodology audits from the lab's working repositories. The list shown is a recent slice. Selected entries may be excerpted publicly on a periodic basis.

Engine label · which research surface the entry belongs to · titles preserved verbatim

Commit activity · trailing 5 weeks

~15 commits/day · lighter on weekends

2026-05-19 engineering raven Prediction-market flow scanner — handoff and instrumentation
2026-05-18 engineering raven Intraday alert plumbing — gating and rate-limit hardening
2026-05-17 publication raven Raven Market Intel — weekend holistic recap
2026-05-16 research raven Spacecraft-sector dispersion — beta proxy and basket tape
2026-05-16 research raven 13F-HR delta engine — Smart-Money Convergence (§8c) activation
2026-05-15 research raven OGE Form 278-T live ingest — executive trades
2026-05-07 engineering dex DEX AMM — thesis and risk envelope, phased rollout plan
2026-05-07 engineering yutani Yutani R14 — pivot to theory-driven gates after R12-Y falsification
2026-05-04 engineering dex DEX AMM — committee synthesis, protocol-mechanics review
2026-05-04 engineering translight Translight V2 — RTDS hotfix, CLOB quoting recovery
2026-04-28 engineering translight Translight engine M — invariants harness and queue model
2026-05-12 research lab Strategy discovery sprint — Tier-1 verdict, infra cleanup

Operating doctrine

A small number of standing rules. Each one is load-bearing.

Evidence before action

A candidate read must clear pre-declared invalidation criteria.

Regime before signal

Context determines whether a candidate read is even interpretable.

Costs before P&L

Transaction costs, slippage, capacity are constraints, not footnotes.

Pre-registration before backtest

Hypotheses are registered before they are tested, not after.

Restraint as output

A high-quality publication is rare by design; silence is a valid result.

The lab knows when the regime is outside its declared envelope, when costs dominate the candidate read, when a backtest is decorative, and when the correct outcome is silence. Each engine observes, classifies, sizes, validates, and gates. A high-quality publication is rare by design.

Source discipline

The source matters before the story.

Trust, after corroboration

EDGAR primary filings (Form 4, 8-K, 13D/G) at filer-level granularity
FINRA short-volume, OTC transparency, and dark-pool alternates
OGE Form 278-T executive trade disclosures and House committee filings
Options surface and flow; vol term-structure and skew
Tier-A geopolitical sources (state.gov, whitehouse.gov, un.org, nato.int, csis)
Official macro releases (BLS, BEA, FOMC, central-bank calendars)

Distrust, unless proven otherwise

Single-account social-media calls without corroboration
Anonymous channels and unsourced urgency loops
Backtests without modeled fees, slippage, latency, or capacity
Strategies that worked only in one cherry-picked regime
Reads that cannot name their own invalidation criteria
Signals that fire after the market has already repriced

Lab thesis

Markets are over-instrumented in obvious places and under-instrumented in interesting ones.

The edge is not the claim that a model can see the future. The edge is the discipline of forcing every candidate read through regime context, source quality, transaction costs, capacity, and pre-registered invalidation criteria — and accepting silence as a valid result more often than not.

01

Asymmetry hides in friction, not in the obvious tape

The crowded trades are well-instrumented. The under-instrumented constraints sit in regulatory filings, procurement calendars, supply chains, options surfaces, executive disclosures, committee-level timing, prediction-market microstructure, and the gap between policy ambition and physical capacity. Asakasa builds where the instrumentation is thin.
02

Discipline is the edge — and it is falsifiable

Every candidate read passes through regime context, source quality, costs, capacity, and pre-registered invalidation criteria. Yutani is the worked example: a directional hypothesis prospectively falsified at R12, pivoted at R14 to theory-driven gates under sequential probability testing. Publishing the negative result is the methodology, not an embarrassment to it.
03

Four research surfaces, one standard of evidence

Raven compresses regime context across twenty-seven named sources. Translight studies passive liquidity on a central-limit order book. The Decentralized Exchange AMM studies delta-neutral perpetuals microstructure. Yutani benchmarks the physical floor of signal-to-submit latency. Each is measured against live venue state, never synthetic replay.
04

AI is leverage, not the alpha

Machine-learning and agentic tooling build scanners, summarise messy evidence, draft methodology audits, and generate adversarial checks. They are engineering and research leverage — never presented as a source of edge. The discipline is the edge; the tooling makes the discipline cheaper to enforce.

The lab builds for an audience that cares about asymmetric technology regimes — semiconductors, AI infrastructure, defence, energy, space, cyber, on-chain venues — and the political and physical constraints around them. These regimes move through bottlenecks before they move through consensus.

Asakasa SRL is incorporated and operates in Bucharest, Romania. All research and engineering work originates from the Bucharest office. The lab is a registered legal entity; jurisdiction and counterparty identity are not abstractions here.

Principal · attribution

Bảo Việt Hà

FRSA · FRAS

Principal of the lab. Bucharest, Romania.

Discipline constellation · five domains, one operator

Raised in Canada. Background spans defence, intelligence, and humanitarian operations across multiple regions, with subsequent work founding and running ventures in the United Kingdom, transition into quantitative systems work, and UK commercial law training. Now based in Bucharest. Fellow of the Royal Society of Arts and the Royal Asiatic Society.

Separately from the lab and on personal account only, an investor and quantitative operator running personal-capital strategies — covered-call income, energy, financials, and broad-market indices — with discretionary focus on asymmetric exposure in defence and civilizational-technology regimes: space and orbital infrastructure, semiconductors, energy and nuclear, and quantum computing. This personal activity is conducted outside Asakasa SRL.

Designs and builds the lab's market-intelligence and execution-research software with frontier agentic coding models. The lab's engines — Raven, Translight, the Decentralized Exchange AMM, and Yutani — are governed by the lab's standing doctrine: evidence before action, regime before signal, costs before P&L, pre-registration before backtest, and restraint as output.

The operating posture comes from prior work more than from the credentials. Defence and intelligence experience disciplines the lab's source taxonomy: Tier-A primary versus mainstream versus single-account; corroboration before action; the difference between a signal and an artifact. Operational experience under real consequence disciplines the risk posture: costs are first-class, the wrong action is worse than no action, and a quiet system is doing its job.

The lab is built so judgment stays with the operator. Automation observes, classifies, validates, and gates. The human signs off. Public outputs are research and infrastructure, not advice.

Conflict-of-interest disclosure

The principal holds personal positions, on own account, in the asymmetric-technology themes named in the bio above — space and orbital infrastructure, semiconductors, energy and nuclear, quantum — and in a broader personal portfolio across energy, financials, and broad-market indices. The lab researches several of these same themes. Raven and the lab's other outputs may reference these same sectors. The lab does not maintain per-instrument position disclosure and does not represent its research as free of positional bias. Readers should assume the principal may hold, or may transact in, instruments related to any theme discussed, and should weigh the research accordingly. Nothing the lab publishes is personalised investment advice.

Direct correspondence: ops [at] asakasa.org · subject line "Principal inquiry".

Contact

One inbox. Three routes.

All inquiries route to ops [at] asakasa.org.

Research access

Request access to the Raven research programme: regime context, event risk, asymmetric technology themes, and methodology notes. Invite-only and not commercially sold. Clicking opens a draft email; access is by reply.

Request access

Bespoke engineering

Engineering availability for research teams: market-signal infrastructure, execution research stacks, regime classifiers, and methodology audits — built to the same standard as the lab's own engines. Available for engagement; scope discussed per inquiry.

Inquire

Research collaboration

Academic or institutional conversations about methodology, regime classification, validation pipelines, and HFT-grade execution architecture.

Get in touch

An applied intelligence lab building high-frequency market-research software.

Operational shape across the lab's engines.

A market-intelligence brief with restraint built in.

The 19 May close, as the lab actually rendered it.

Raven is live. The other three are research surfaces under active development.

Raven

Translight

Decentralized Exchange Automated Market Maker

Yutani

A Python brain and a Rust executor, separated by a single IPC boundary.

The lab builds for itself first — and is open to building for others.

Market-signal infrastructure

Execution research stacks

Regime classification frameworks

Methodology audits

Named streams. Tracked freshness. Documented gaps.

Rigor is not optional. It is the product.

Bayesian regime classification

Pre-registered hypotheses

Walk-forward validation

Multiple-comparison correction

HAC standard errors

Sequential probability testing

The apparatus turned on the lab's own hypothesis — and killed it.

Directional bucket R12-Y

Prospective evidence said no

R14 — methodology, not data

SPRT advisory in production

A research lab proves itself by what it ships, and on what rhythm.

A dated record of work across the lab's engines and research surfaces.

A small number of standing rules. Each one is load-bearing.

Evidence before action

Regime before signal

Costs before P&L

Pre-registration before backtest

Restraint as output

The source matters before the story.

Markets are over-instrumented in obvious places and under-instrumented in interesting ones.

Asymmetry hides in friction, not in the obvious tape

Discipline is the edge — and it is falsifiable

Four research surfaces, one standard of evidence

AI is leverage, not the alpha

Bảo Việt Hà

One inbox. Three routes.

Research access

Bespoke engineering

Research collaboration