The money
One WHERE decision revives 94 discarded factors: Sharpe 0.85 → 1.30. Pure venue spread alpha +2.3%/yr (t = 5.4); GRS rejects spanning F(2,417) = 51.
load-bearing · survives FF6 + BABHuannian Jin
I build an AI agent factory that proposes return factors by the thousand— then prove that unconditional screens have a structural blind spot for factors that only pay inside a market state.
A significance hurdle in one market venue is a fixed-aperture instrument — like Abbe's microscope, it has a diffraction limit. I give the resolution law of factor discovery, q²(τ)=τ/(1+τ), matched to Monte Carlo to three decimals. 94 anomalies a |t|<1.96 screen rejects earn out-of-sample Sharpe 0.85 where tested, 1.30 reallocated as a pool (+0.45, FF6 α 3.8%/yr, t=8.3, GRS F=51); a China A-share panel of 5,441 stocks reproduces it out of sample (+0.95, p=0). The per-factor ceiling is a winner's curse — in a zero-advantage null the oracle manufactures 5.9× the real gain. The replication crisis is a measurement problem. Working paper, Oct 2026.
0bps/yr VW active
0information ratio
0of grave→oracle gap (coarse)
0Sharpe gain over practice
0machine factors
0falsification excess
0modules @ Zheshang
0LoC shipped
Two mirages
The fine-routing ceiling is a winner's curse — 5.9× of the gap is selection over noise. Risk-parity super-Sharpes turn negative vs BAB (t = −2 to −4).
falsified, not leaned onThe limit
Geometrically multi-rank (γ 0.38–0.78), tradable-rank one. The only direction that survives out of sample is the common small-cap tilt. m* ≈ 1.3.
impossibility theorem · 10 falsifications
UNCONDITIONAL SCREEN
OWN-REGIME TILT
Interactive sketch — hard gates discard legs; continuous tilt keeps breadth. Move mouse to perturb states.
00a · Why HAA · the case in sixty seconds
Why HAA, in three claims — all checkable on this page
No adjectives. Every number below is either printed elsewhere on this page or computed in this tab — and the agent in §00b will re-run any of it on request. If a claim looks too good, ask it to prove the claim; it is built to say no.
What I bring
Twenty-two months on one system, not one prompt-weekend. A personal A-share book at 172 bp/yr · IR 1.99; 64k lines shipped; 6,881 machine-mined factors; 56 strategy modules; a screen that covers 75% of the gross→oracle gap; evidence surplus 6.4× the median claim.
Most portfolios show a demo. This one shows the audit trail — longer, and considerably less flattering, which is rather the point.
What the agent is
Router first, model second. Twenty-seven instruments compute before any language model is allowed to speak; a seven-persona council adjudicates what the numbers mean; the paper's four propositions are callable instruments. Every figure is computed in this tab under one seed, with zero network calls — the provenance line under each answer says so.
A chatbot with a calculator is a toy; a calculator with a chatbot is a colleague. This is the second thing, mostly because the numbers refuse to move when asked nicely.
Why I am the one who can drive it
I wrote the grammar, the declared tape, the nine gates, the council's thresholds and the paper the whole thing serves. The agent's instruments are my research method, compiled — which is why it can be pushed hard without breaking: there is nothing in it I did not build and cannot explain line by line.
I did not fine-tune a model to sound like me. I wrote the instruments it is not allowed to lie about.
The paper, in four lines
P1 dilution t_g = t·√π · P2 the wall t ≥ √(2 ln N) · P3 the ceiling q²(τ) = τ/(1+τ) · P4 pooling mean(t)·√n. Ask the agent “论文的四个命题” and it prints where each one runs on this page; ask “the ceiling at tau 0.3” and it computes the bound live.
The paper's most useful sentence is the one admitting what it cannot do. Investors call that “risk disclosure”; I call it the only honest slide in the deck.
Sixty seconds, if you only have sixty:
1 · this page
2 · ask the agent to generate five factors
3 · run the terminal's
agent
the rest is arithmetic
00 · BOAT Terminal
A full-stack quant desk — Chanlun · physics · factors · a research agent that computes
Nearly one year of nights shipping a usable terminal: Rust for the compute core, Python for strategies and AI, Flutter for the desk UI. Walkthrough lens is 600519.SH (Moutai). Posted to learn in public — backtests labeled as backtests, personal A-share book labeled separately from the JF paper track.
SYMBOL600519.SH
LAST1413.64+1.55%
TFDaily
ISING T2.075HIGH
LPPLS35%
V12 SR2.595
MAX DD11.6%
BOOK¥50k → ~¥1.00M
Chanlun engine
Strokes, pivots, B1/S1/B2/S3 labels with pairing lines — structure as audit trail, not decoration.
Ising · LPPLS · TDA
Temperature / force-field, bubble probability, Betti persistence — microstructure read in statistical physics language.
V12 cross-section
CSI-300 book: equity + underwater, factor matrix, SHAP, Monte Carlo permutation, Walk-Forward folds.
Research agent
boat-1 judges factors, evaluates them on a declared market, trains real models on them and permutation-tests the result — every number computed in this tab. Try it below.
Why ship it public
Quant is lonely. I want critique from people who have built Chanlun tools, factor stacks, physics reads, or real multi-agent systems — not applause for screenshots.
Honest boundaries
Still rough. Backtest ≠ live book ≠ JF claim. Each surface is labeled. This terminal is the builder desk; the paper is the scientific claim.
What to pressure-test
Ask boat-1 hard questions below. It answers only from instruments that run in this tab — and says so when it can't.
The brain
Two ways to give this page a real language model. Either way the model cannot compute anything itself: the ten instruments stay its tools, and the system prompt forbids stating a number that did not come back from one.
Local endpoints (Ollama, LM Studio) work only when this page is opened over http — over https, Chrome's private-network policy refuses the loopback call, and no header can lift it. On the live https site, use a public endpoint.
Your own machine as the server (16 GB card): install LM Studio, pull Qwen3.5-9B GGUF from ModelScope — pip install modelscope then modelscope download --model Qwen/Qwen3.5-9B-Instruct-GGUF --local_dir ./qwen35-9b — load the Q4/Q6 file in LM Studio, start its local server (port 1234), open this page over http, and put http://localhost:1234/v1 below with model qwen3.5-9b. Native tool calling, all 24 instruments, no download from huggingface at all.
The model file itself (~1.1–2 GB) downloads from huggingface.co the first time — if that host is unreachable on your network, use your own key instead.
The key is written to this browser's localStorage and is sent to nobody except the provider you chose. The tape stays declared-synthetic either way: admission at the gate never means a program earns.
00b · research agent
Ask boat-1.
It judges factor expressions, evaluates them on a declared market, trains real models on the eight signals and permutation-tests the answer — all computed in this tab. No model call, no network, no API key. Attach a brain (top right) and a real language model answers instead — with the same instruments as its tools, and still no invented numbers.
Factor workbench — your factors, our machinery
Paste a library — one “name, t” per line, or “name :: formula”, or bare expressions, or sharpe= for your own strategy — or load a .txt/.csv file. It is measured here, then run through both ways: the universal screen against home-venue routing plus pooling.
computed in this tab · nothing leaves the page
Figures · Moutai desk walkthrough
Click any panel to expand. Same tape boat-1 reasons over.
Desk layer · Chart + Chanlun + Ising
Candles with strokes, pivots, labeled B1/S1/B2/S3 and pairing lines; RSI below. Bottom panes: Ising force-field (long/short spins) and thermometer (T≈2.075 · High) — microstructure as statistical physics.
Research layer · V12 + antifragile checks
CSI-300 constituents. This run: Sharpe ≈ 2.60 · ann. ≈ 54% · max DD ≈ 11.6%. Correlation matrix, importance ranks, Monte Carlo permutation, Walk-Forward — overfitting detectors on purpose.
Decision layer · six lanes, one verdict
Captain · Chanlun · Physics · Factor · Risk · Execution. Lane reports first; Captain tallies. DeepSeek on the real desk. Collaboration depth is still open — challenge it.
01 · The thesis
Global Weakness, Local Strength
Core claim: unconditional factor screens have a structural blind spot for state-dependent premia. Machines make the supply cheap; the filter still asks the wrong question. This is not a weekend demo — it is a multi-year research program I am on leave to pursue full-time.
Machines flood the zoo
LLMs + search propose predictors faster than humans can vet. Screens become the bottleneck.
Unconditional filter
Keep if average premium clears multiple-testing. State-confined premia look like zeros.
Structural blind spot
Global t dilutes with √π. Locally strong factors are discarded as globally weak.
Own-regime tilt
Per-factor states + hierarchical pooling. Continuous tilt—no hard gate that kills breadth.
Falsify, then trade
Block-shuffle nulls, placebos, cross-fit. Then show deployable bps on public panels.
Honest tiers
Load-bearing vs identification vs suggestive—said upfront. Science over storytelling.
Regime overlay adds deployable, net-of-cost alpha to every implementable baseline.
- 24/24 same-base active increments positive (RW + e-BH)
- VW: 172 bps/yr active · IR 1.99 · deployable t = 6.81
- OOS VW: 134 bps · IR 2.05
Discarded & machine factors carry genuine state-dependent structure.
- Falsification excess 6.4× (219 vs 34 under global null)
- Funding-scarcity interaction t = 3.62 / 2.81
- Vol / broad-stress placebos: no such gradient
Reported, not leaned on—honest science.
- Cross-asset travel is incomplete
- Reg SHO natural experiment is a null
- Shown for completeness only
01a · Interactive argument tour
Referee-speed walkthrough
Step through the paper the way an editor skims: claim → mechanism → machine pool → economics → gates. Use the buttons or keyboard ← →.
1 / 6
01b · Submission track
Where the paper is in flight
Not “published.” Not vapor. A multi-venue academic pipeline with honest status labels— the Academy cares about motion; journals care about priority and process.
Research arc: Global Weakness, Local Strength is the phenomenon and the deployable economics (at the Journal of Finance, external review). Discarded but Alive is the law that explains and bounds it — a working paper (Oct 2026 draft) with a fully public repro, companion to the JF submission. Acceptance is not claimed for either.
Journal of Finance
- Top finance journal · AFA
- Passed desk review · now in external / referee review
- Manuscript: Global Weakness, Local Strength
- Full PDF password-gated during review — acceptance is not claimed
Eastern Finance Association · US
- Conference submission of the same research program
- Signals: peer academic pipeline, not only a private notebook
- Status updates as decisions arrive — no overclaiming acceptance
Western Finance Association · US
- Top-tier US conference — same research, submitted as Discarded but Alive
- Paper + fee on file; decision pending — acceptance not claimed
- Peer-reviewed pipeline, not a private notebook
SFS Cavalcade · US
- Preparing submission of the same program
- Broadening the peer-review surface across top US conferences
- Status will update honestly as it is submitted
European Finance Association · EU
- Preparing submission for the European track
- Note: “EFA” here is European, not the US Eastern FA above
- No acceptance claimed — a submission pipeline across continents
SSRN
- The resolution-limit law behind the deployable economics
- Three acts: the money · two mirages · tradable-rank one
- Every number regenerates from the public repro pack
Receipts — click to enlarge:
All tracks are submissions. Acceptance is not claimed. The proof of work is the paper + repro pack + systems letter—not the journal logo.
01c · Interactive paper map
Walk the argument like a referee
Click a claim tier. The right pane opens the mechanism, the number that carries weight, and what a third party can recompute from the public pack.
MS2026-0738
StatusPassed desk · external review
EditorAntoinette Schoar
Public checkpython reproduce_headline.py
01d · Paper atlas
Argument graph from the manuscript
Click a node to open the load-bearing sentence from Global Weakness, Local Strength.
02 · Mechanism lab
Why the global t goes blind
A premium confined to a paying state of frequency π is diluted in the unconditional mean. Drag the controls — watch the screen discard a factor whose conditional edge never changed.
Unconditional mean
State-frequency-weighted average. Local strength is averaged away by construction.
Blind-spot identity
Proposition 2: global t vanishes as the paying state grows rare—even when conditional t stays large.
Breadth, not leverage
Hard gates raise per-bet skill but destroy effective breadth. Continuous tilt keeps legs and scales weights.
Live mechanism lab
Illustrative first-order lab for Proposition 2. Not a trading simulator — a screen-discard visualizer.
π0.25
tstate3.00
tg1.50
threshold2.00
verdictdiscarded
dilution√π = 0.50
Hard gate
CUT
Binary: keep if |tg| ≥ threshold, else discard the entire leg.
Continuous tilt
WEIGHTED
Keep the leg; scale exposure by state-dispersion / local strength. Rare states shrink weight — they do not delete breadth.
Curve: tg ≈ tstate√π. Same conditional edge — only π / threshold change. Move mouse on the hero canvas for the same intuition in particle form.
02b · Resolution limit
The resolution limit
In 1873 Ernst Abbe showed a microscope's sharpest detail is fixed by its aperture, not the grinding of the lens. A universal significance hurdle is the same kind of instrument, and factor discovery has the same kind of limit. How much of a unit's true home-tilt the data recovers obeys one clean attenuation law, q²(τ)=τ/(1+τ) with τ=m/m*, m*≈1.3 factors — negligible for a single factor, nearly complete for a pool. Move the whole discard pool to its home venue once (coarse) and out-of-sample Sharpe goes 0.85 → 1.30, three-quarters of the way to the 1.45 hindsight ceiling. The converse is what makes it a law: no router closes the rest.
Comparative advantage
A global component β_i plus a venue-local component a_iv. Factor-to-venue assignment is a self-selection problem: the screen keeps the global part and discards the local part as noise. On 94 discarded factors the coarse move earns FF6 alpha +3.8%/yr (t=8.3), HXZ q5 +3.4%/yr (t=7.5), increment alpha +2.3%/yr (t=5.1), GRS F(8,594)=5.7 (p<1e-6).
A diffraction limit, not a wall
The squared overlap between the estimated home-tilt and the truth converges to q²(τ)=τ/(1+τ): negligible as τ→0, exactly one-half at the half-recovery scale m*≈1.3 (τ=1), rising smoothly to one. There is no hard wall — but on the log scale of m the crossover is sharp. The pooled coarse move (m=94, τ≈72) lands at q²≈0.99; per-factor routing (m=1, τ≈0.77) lands at q²≈0.43, below half. Matched to Monte Carlo to three decimals (Appendix IA.2).
No router closes the gap
Coarse leaves only 0.14 to the ceiling. Five routers — per-factor argmax, economic-family prior, trailing-performance timing, an ML characteristic map, empirical-Bayes shrinkage — each tuned out of sample and frozen, recover at most 0.05 of it. Routing every factor (fine) loses 0.18. The empirical-Bayes rule, free to shrink, moves zero factors.
Coarse is the capacity
The claim is +0.45 Sharpe over current practice (discard = zero), not beating the oracle. Coarse captures 3/4 of the grave→oracle gap, and no method beats it by more than noise. Same wall holds internationally: JKP 47 countries, single-factor venue persistence median 0.23.
Diffraction limit — live
q²(τ)=τ/(1+τ) · τ=m/1.3 · half-recovery m*≈1.3
m=1
τ=0.77
q²=0.43
below resolution
incoherent (noise)
coherent (recovered tilt)
residual→oracle0.14
best router≤ 0.05
fine routing−0.18
clear τ_i>16 / 94
Curve: q²=τ/(1+τ) vs. pooled size m (log axis), half-recovery marked at m* ≈ 1.3. The coarse pool (m=94) sits far right at q²≈0.99; per-factor routing (m=1) sits left of m* at q²≈0.43. Illustrative, not a trading simulator.
Resolution law q²(τ)=τ/(1+τ), τ=m/1.3: a single factor (m=1) recovers q²≈0.43 of its home tilt; the pooled discard set (m=94) recovers ≈0.99. Half-recovery at m*≈1.3 factors.
Scale law: the discovery hurdle grows like √(2 log N); the allocation threshold like √(2 log log N). The gap diverges — stricter policing of the zoo discards MORE recoverable local strength as N grows.
03 · Deployable economics
Numbers that survive baselines
Overlay on Chen–Zimmermann OSAP panels. Against seven allocators × four panels, 24/24 same-base active increments stay positive. Click a panel row — the chart locks onto that book.
Two levels of one program: the 172 bps/yr value-weight overlay below is the deployable payoff; a single coarse reallocation moves Sharpe 0.85 → 1.30 (three-quarters of the way to the 1.45 ceiling), and the converse proves no router closes the rest. Same idea — route strength to its venue — read once as economics, once as a resolution limit.
The Sharpe ladder — and the mirage
discarded (practice)
coarse — tradable
oracle mirage → collapse
ML router (OOS)
Out-of-sample Sharpe. One coarse reallocation earns +0.45 over discard (95% CI excludes zero, p(>0)=1.000). The in-sample oracle (1.45, hollow) is a winner's curse: held out of sample it collapses to 1.12, and ~5.9× of its apparent edge is selection over noise. The China A-share panel (5,441 stocks) reproduces the gain out of sample (+0.95, bootstrap p=0). Hover a bar for t-stat, CI and p.
OOS Sharpe: discarded 0.85 → coarse pool 1.30 (+0.45, 95% CI excludes 0). The in-sample oracle 1.45 collapses to 1.12 out of sample; no router beats coarse by more than noise. China A-share: +0.95, bootstrap p=0.
| Panel | MinVar SR | Overlay SR | Active bps/yr | IR | Deployable t | OOS bps | OOS IR |
|---|---|---|---|---|---|---|---|
| Equal-weight | 2.79 | 3.01 | 128 | 1.73 | 1.49 | 101 | 1.72 |
| Value-weight ★ | 1.37 | 1.91 | 172 | 1.99 | 6.81 | 134 | 2.05 |
| NYSE breakpoint | 1.73 | 2.15 | 134 | 1.68 | 4.11 | 116 | 1.52 |
| Large-cap (ME20) | 1.77 | 2.05 | 109 | 1.66 | 2.79 | 75 | 1.60 |
Breadth content
EW: ~115 factors significant unconditionally → ~268 factor–state cells (~2× independent bets from the same raw material).
Credibility stack
Cross-fit (labels ⊥ alpha), own-regime beats placebo 3–5×, states persist ~8.6 months, Romano–Wolf / e-BH / Lo (2002), deflated Sharpe ≈ 1.
Economics
Local strength steepens when funding capital is scarce (HKM & Baa–Aaa). Volatility placebos fail—limits to arbitrage, not vol timing.
04 · Machine population
What screens throw away when agents propose
6,881 genuinely machine-generated factors. None clears the global screen. Click each funnel step — see the standard tighten and what survives.
6.4×
Flagged set’s falsification passes vs global-null allowance (219 vs 34)
Alpha191
13.8%
regime-local rate · human expert libraryAlpha101
11.6%
regime-local rate · human expert libraryReading
Same order of magnitude as the machine pool → structural property of unconditional screening, not one generator’s artifact.
The factor zoo — two hurdles diverge
kept: 6%
profitable, discarded: 94%
hurdle |t|=4.21
kept
profitable at home, discarded
null
N=6,960 candidate factors; effective dimension d_eff=17. The discovery hurdle grows like √(2 log N) and the allocation threshold like √(2 log d_eff) — the gap diverges as the zoo grows. Judged in one venue by the universal hurdle, ~94% of factors that are genuinely profitable in their home venue are wrongly discarded. Hover a point for both t-stats.
N=6,960 factors, d_eff=17. Universal hurdle √(2 log N)≈4.21 vs allocation hurdle √(2 log d_eff)≈2.38. Judged in one venue, ~94% of genuinely home-profitable factors are wrongly discarded.
05 · AI agent factory
Visible traces — not chatbot cosplay
Agents emit typed ASTs over nine microstructure signals.
Select a machine-emitted factor →
06 · Industry proof
Zheshang Securities · recommendation
Quantitative System Development Intern (from Aug 2025) at Zheshang Securities Co., Ltd. (Hangzhou). Letter by Liao Jingchi, Managing Director — highest recommendation for graduate study and research/engineering practice.
56modules from ground up
~64klines of code
>95%test coverage
1sole developer · task system
Quote
What he owned
- Fully involved in design & development of the firm’s quantitative trading system
- Sole developer of the task subsystem — complex architecture, end-to-end delivery
- Built mathematical foundation layer, data processing layer, and multiple core modules from scratch
- Shipped 56 modules · ~64,000 LoC · >95% test coverage under a rigorous QA framework
Engineering depth (letter)
- High-precision computation
- Distributed data storage
- Real-time data synchronization
- Dependency management for complex systems
- Root-cause debugging under production pressure
How he worked
- Integrated quantitative trading business logic into system architecture
- Clear technical communication; proactive alignment with business requirements
- Described as rigorous, efficient, meticulous, highly focused, and passionate about quant finance
- MD endorsement: highest recommendation for graduate programs
06b · Proof of work · different shape
Same Academy bar · different instrument
The Founding Class includes apps, hardware, nonprofits, ARR. Official language also counts original research as built. My instrument is inspectable judgment under AI-scale factor supply.
Common Stage-1 surfaces
- Product demos · audience · ARR
- Hardware / community / teaching
- “I used AI to ship faster”
- One metric, thin falsification
This surface (research × systems)
- JF MS 2026-0738: passed desk · external review (Schoar) + EFA
- AI×quant factory with gates — propose cheap, admit expensive
- Falsification excess 6.4× · claim tiers stated upfront
- Public repro · CI greens · walkthrough video
- Zheshang MD letter: sole task-system owner · 56 modules · 64k LoC · >95%
- On academic leave · still ≤1 year of college completed
- Killed own wrong regime layer when data said no
06e · AI × quant
Not “AI for trading tweets” — a gated research factory
The Academy asks for agency when something captures attention.
AI side
- LLM agents propose typed factor ASTs
- Sandbox + constitution + rationale gates before admission
- Same vibe-coding stack for the public site
- Policy as code: AGENTS.md · human gate · CI
Quant side
- Proposition 2 blind-spot identity · hierarchical own-regime tilt
- OSAP deployable economics · Romano–Wolf / e-BH / Lo (2002)
- 6,881 machine factors studied as a population — including discards
- Production systems internship: math layer → data layer → 56 modules
The join
AI expands the proposal set. Quant discipline decides what enters science.
Long game: keep tightening identification.
Live contrast · what each side owns
Toggle AI vs Quant emphasis.
06c · Engineering discipline
How work actually ships
I use agents every day. That is normal.
Loop
intent → agent draft → human gate → verify → CI → ship.
Policy as code
Machine-readable rules live in AGENTS.md…
Thin public surface
Static site/ on purpose — fast for reviewers on phones.
Fail closed
No invented metrics. No JF PDF in public git.
Same gates as research
LLM propose → type-check → sandbox → constitution → admit.
Receipts
Local scripts/verify.* · Actions repro-smoke + pages.
06d · Engineering mass
Large private body · thin public receipt
Competitors dump folders. I show partition.
12crates in private boatfx
~2.3kRust source files
~670kRust LoC (ex-build)
77llm_agent modules
56Zheshang modules shipped
>95%enterprise coverage
Public GitHub is deliberately small.
07 · Proof of work
Verify in one command
Headline arithmetic regenerates from shipped tilt output—no router source required.
- Public grammar + 60 sample factors · python repro.py → 60/60
- Tilt output shipped · python reproduce_headline.py → ~172 bps · IR ≈ 2
- Full pool & ranking policy withheld on purpose
- JF MS 2026-0738 · passed desk · external review · EFA submitted
09 · Field presence
Wiki Finance Expo Hong Kong 2026 · VIP
Invited as a VIP guest to Wiki Finance Expo Hong Kong 2026 — Hopewell Hotel, 15 Kennedy Road, Wan Chai · 23–24 July 2026.
VIP field dossier
Not a photo dump — a structured record.
Day 1 · Floor + rooms
Morning VIP badge…
Day 2 · Continuity + night
Returned to the floor…
What I took back to the notebook
Three frictions…
Paper ↔ field bridge
Proposition 2 bridge…
Why VIP matters (to me)
VIP was not a sticker for LinkedIn.
What I actually did
Walked the branded floor.
Bridge back to the paper
The expo’s public language maps onto my research tension.
09b · Exchange visit
Hong Kong Exchanges (HKEX) · invited learning visit
Invited to Hong Kong Exchanges and Clearing for learning and exchange — not a tourist photo. Visitor badge on, standing at the HKEX brand wall (香港交易所). A separate track from Wiki Finance Expo VIP: here the object is market infrastructure and how a major exchange presents itself to learners.
Why this visit
To see exchange infrastructure and institutional language up close — clearing, listing, market structure talk — and pressure-test whether my screening-science agenda still sounds real next to the institutions that define “the market.”
What I did
Entered on invitation with a visitor pass. Walked the branded lobby, listened, asked. Kept the receipt photo on purpose: body on site at 香港交易所, not a remote claim.
How it differs
Separate from Wiki Finance Expo VIP (industry expo rooms) and from the Zheshang summer camp (where I was an instructor). HKEX = invited learning & exchange at the exchange itself.
09c · Instruction
Zheshang Research · Houlang Project · Elite Class VII
Instructor (指导员) for 浙商证券研究所「后浪计划」投研精英班第七期暨暑期考察营 — the 7th Investment Research Elite Class & Summer Inspection Camp. Not a spectator seat: I guided participants through research / quant-engineering practice while the program ran under Zheshang Research Institute branding.
Instructor dossier
Structured record of role, rooms, and what “指导员” meant in practice.
Program identity
浙商证券研究所 · 后浪计划 · 投研精英班第七期暨暑期考察营. Corporate talent track — research elite class plus summer inspection camp — not a casual meetup.
My role
指导员 / instructor: stay with the cohort, unblock research and engineering questions, keep sessions moving from talk to runnable work. Distinct from being a trainee in the seats.
What the photos show
Signed red backdrop (cohort presence). Full classroom under the program LED. Laptop in the hall debugging quant / data issues during the camp window (timestamp visible July 2025) — instruction that includes code, not only slides.
Bridge to systems letter
Same Zheshang orbit as the MD recommendation letter (production systems internship). Camp = teaching / guiding face; letter = shipping face. Two proofs, one firm.
Why instructor matters
Agency is not only building alone. Guiding a research cohort under a securities-institute brand is proof I can hold a room, translate quant ideas, and stay accountable when others’ work depends on clarity.
Separate from HKEX & Wiki
Wiki VIP = industry expo guest. HKEX = invited learning at the exchange. This camp = instructor inside Zheshang’s Houlang talent program. Three receipts, three roles.
Academy signal
Already in motion across venues: research paper track, production systems, field VIP, exchange visit, and formal instruction. Different instruments, same bar — show up and leave a checkable trace.
10 · Private boatfx stack
The body behind the thin public receipt
Public GitHub is curated on purpose. The private monorepo is a multi-crate trading / research system.
Crates12
Appcaptain
UI packagescharts · tokens · ui · icons · shared
Rust mass~2.3k files · ~670k LoC
Public surface stays thin: boatfx-repro = site · repro · showcase · CI.
08 · The Horowitz Andreessen Academy
Already in motion
We’re looking for people who are already in motion…
I am on an academic leave of absence from my university to pursue this research full-time. I remain within Academy eligibility: I have not completed more than one year of full-time college after high school (through August 2027). International applicant. Willing to live in San Francisco full-time.
Essay and video are mine. AI helps me code and edit — not invent the story.
Artifacts
For reviewers
- Live site (public)nemo02070118.github.io/boatfx-repro →
- Reproducibility packGitHub boatfx-repro →
- AI factory source excerptssystem_showcase →
- JF manuscript + Zheshang letterPassword-gated for admissions
- Walkthrough videoyoutu.be/tVHLUQy93rg →
- Emailhuanian.jin@outlook.com