AI Industry Chain Transmission Map — an open-source US equity alpha engine built on the thesis that AI capex flows through a predictable multi-layer physical supply chain. ProphetMap tracks where that capital flows, scores each ticker against five independent dimensions, and surfaces buy-signal candidates before consensus catches up.
Live: prophetmap.vercel.app
Universe: 87 tickers across 28 layers — universe v2.9.10 (2026-07-28), latest framework change v2.9.11 (2026-08-12). See CHANGELOG.md for the full history and GOVERNANCE.md for the rules that constrain it.
The screen is the visible part and the least interesting one. The engine's real claim is that a rating may not move quietly. Three mechanisms enforce it:
- Pre-declared falsifiers. Every ticker carries a
thesisFalsificationarray — observable events, written before the fact, that would invalidate it. A falsifier that is written but never tested is not a falsifier, so outcomes are logged infalsificationObservationswhether or not they fire; "tested and survived" and "nobody looked" must not be the same shape in the file. Two classes of defect are recorded rather than hidden: gameable thresholds (satisfiable without validating the thesis) and unmeasurable ones — a falsifier whose expiry passed with no published number to test it against can never trigger and never be refuted, which turns a position into faith while it still looks like discipline. - Methodology events are separated from signals. When the engine changes, names enter and leave the pass set for reasons that have nothing to do with the companies. Those commits state the full per-ticker delta up front, and a benchmark or peer change is not allowed to re-rate an existing holding as a side effect. The pricing gate has hysteresis (2.8 in / 3.2 out) precisely because a single-day threshold crossing is not an audit trail.
- Defects are published, not fixed quietly.
GOVERNANCE.mdcarries 12 numbered gaps, several marked deliberately not fixed with the reason — usually that the fix would silently re-rate names. One layer is recorded as permanently unanchorable: its external-peer count is 0 because every public comparable is either held or private, so its pricing output is a within-layer ranking and never an absolute valuation.
An engine allowed to revise its own history can prove anything. These are the constraints that make the record worth reading.
"Are these sources answering the same question?"
Most tooling asks whether independent sources agree. This one asks whether they are measuring the same thing at all — two sources can differ by 68 points and both be right.
Five failure families keep surfacing in five unrelated domains; a sixth has so far appeared in one domain, and a seventh was found the other way round — first in someone else's published work, and only then looked for here. Three of the five domains are public repositories, below; the fourth is a local experiment cited in the evidence table; the fifth is a third-party skill I did not write. This repo is the self-built-scoring instance: the one where the rule is turned on its author.
| repo | domain | the question it asks |
|---|---|---|
| decision-confidence | third-party risk vendors | do these vendors answer the same question? |
| assay | LLM-as-judge | does this metric measure what its name claims? |
| prophetmap ← you are here | self-built equity scoring | does my own score survive my own rule? |
Cross-domain evidence → failure-families.md
Since v2.9.18 the first row is no longer only a sibling — it is the instrument
this repository is measured with. decision-confidence states the rules;
this repository is the case where they were turned on, published as
a case study
there. The two stay separate on purpose: the freeze's integrity check asks
this repository's git history whether the frozen roster still has exactly
one commit, so the anchor is the history itself. Merge the repositories, or
retire this one, and the freeze ends — inside a lock its own author wrote and
has to sit out.
Not investment advice, and deliberately not a portfolio. ProphetMap emits screening signals, not position sizes (
GOVERNANCE.mdgap #5). The funnel is an entry gate: a held name dropping out of the pass set is not a sell signal.
AI demand doesn't teleport to software companies. It travels through a physical chain. Each layer has physical bottlenecks (fab lead times, power permits, optical fiber manufacturing, defense contract cycles) that create durable pricing power — the kind harder to compete away than software margins.
The universe is organized into 5 chains × 28 layers:
- L0 AI Foundation Model & Platform (MSFT, AMZN, META, ORCL)
- L1 AI Demand & Application (GOOG, PLTR)
- L2 AI Training Compute (NVDA primary; AMD, CBRS)
- L2_5 AI Inference & Edge Silicon (QCOM, ARM, MRVL, AVGO)
- L3 EDA & Chip Design Tools (SNPS, CDNS; ALAB)
- L3_5 Semiconductor Materials & Specialty Gases (ENTG, LIN; APD, TSEM)
- L4 Semiconductor Manufacturing Equipment (AMAT, LRCX, ASML, KLAC)
- L5 Advanced Packaging (AMKR, TSM)
- L5_5 PCB & Substrates (TTMI)
- L6 Memory & Storage / HBM (MU)
- L7 Server OEM & EMS (CLS, CRWV, SMCI; DELL)
- L8_NET Data Center Network Interconnect (ANET)
- L8_COOL Data Center Cooling & Power (VRT; ETN)
- L8_OPT Optical Fiber & Transceiver (GLW, COHR, LITE)
- L9 Data Center Construction & Real Estate (EQIX, DLR; FLR, CRH)
- L10 Cybersecurity (CRWD, PANW)
- L11 Clean Baseload & Nuclear Power (CEG; VST, OKLO, SMR)
- L11_FUEL Nuclear Fuel Feedstock (CCJ) (new 2026-05-19)
- L12 Grid Equipment & Electrical Infrastructure (GEV, PWR)
- L13 Natural Gas Production (EOG; EQT)
- L14 Critical Commodities (FCX, SCCO; AA)
- L_EMBI Embodied AI Infrastructure (ISRG, CGNX; PH, TSLA)
- L_SPACE Space Infrastructure (RKLB, ASTS; MOG)
- L_DEF Defense AI (AVAV; KTOS) — drones, autonomous combat systems
- L_DCOMP Decentralized AI Infrastructure (RNDR, TAO, FIL, LINK, ETH; CRCL)
- L_EXP_QC Quantum Computing (IONQ; RGTI) — graduation criteria in universe.json
_graduation
Every ticker is scored against five dimensions. Defensibility is an either/or — a physical chokepoint or a non-physical moat the company actually captures — and the remaining gates are conjunctive:
| Dimension | Pass Threshold | What It Measures |
|---|---|---|
physicalConstraint |
≥ 4 OR ↓ | Moat depth: how hard is it to replicate this position? (1=pure software, 5=hard physical monopoly with multi-year lead times) |
moatCapture |
≥ 4 (either suffices) | Who keeps the moat. A durable moat can exist and accrue entirely to someone else. 1 = pure supplier, the rent goes to the incumbent or the customer; 5 = the company itself holds the licence, the liability, the private ground truth or the user habit. ≤ 2 raises a supplier-trap warning — surfaced, never a veto. |
aiContribution |
≥ 0.30 | What % of forward revenue growth is directly attributable to the AI thesis? |
timeToRealize |
near or mid |
How soon does the thesis cash flow? (near = <12m, mid = 12-36m, far = 36m+) |
pricingScore |
hysteresis: ≤ 2.8 to enter, > 3.2 to lose | Is the market already pricing in the thesis? (1=deep value opportunity, 5=fully priced euphoria) |
Two details that are load-bearing rather than cosmetic:
- The pricing gate has a hysteresis band, not a line. One name crossed a hard 3.0 threshold six times in sixteen trading days; a pass that flickers daily cannot be used as an audit record. Names sitting inside 2.8–3.2 are flagged as carried over, not as fresh signals.
- Unpriceable fails closed. When the pricing inputs do not apply to a name, the
result is
pricingApplicable: falseand it does not pass. No way to price it is not the same as cheap — and a loss-making company being locked outside the gate byforward P/E < 0is recorded as a policy, not presented as a measurement.
A screen nobody scores is a hobby. The engine's own claim is under a pre-registered test with two gates that must both pass:
| test | why this one | |
|---|---|---|
| Gate A — outcome | the pass-set basket beats SMH, risk-adjusted | the claim is that picking inside the chain beats holding the chain. Beating a broad market index would only prove long beta — the most self-flattering benchmark available. |
| Gate B — process | ≥ 3 timestamped names that were genuinely non-consensus when the signal fired, and eventually right | an outcome gate alone can be passed by luck in one direction |
Window: 12 months, with a mid-point review that does not execute. Failing either gate is a result — "this screen produces no stock-selection alpha" — not a bug, and the honest response is to stop running it as if it did.
Gap #12 in GOVERNANCE.md records the sharpest known weakness of
that test: gate B currently stores whether a call was non-consensus but not which
kind. Since every input here is public and no sentiment is read, the engine can
structurally only produce analytical and technical edge — so if the cases turn out
to be mostly behavioural, gate B passing would be a false positive. That is a
falsifier gate B cannot generate about itself.
The pricingScore is computed live from Yahoo Finance data using a weighted composite:
30% × Forward P/E deviation from layer median
25% × EV/Revenue deviation from layer median
25% × Analyst consensus upside (inverted — low upside = high score)
20% × 6-month price momentum vs SPY
Crypto tickers (L_DCOMP) use a separate pricingScore formula (35% Market Cap / Protocol Revenue, 35% Market Cap / TVL, 30% 6m momentum vs ETH inverted) computed via update-crypto-valuations.js.
| View | URL | What You See |
|---|---|---|
| Chain Map | / |
All layers with tickers colored by pricing score. Green dot = funnel PASS. |
| Funnel | /funnel |
PASS table sorted by pricing score + near-miss watchlist |
| Signals | /signals |
Gemini 2.5 Flash thesis falsification proximity assessment per ticker |
Every ticker has an explicit thesisFalsification array — observable events that would invalidate the thesis. The signal analysis pipeline (Gemini 2.5 Flash) assesses proximity (0=intact, 1=watch, 2=approaching, 3=imminent) for each signal daily.
Discipline rule: When a pre-declared falsification signal fires at proximity ≥2, status must change (e.g., active → watchlist), otherwise the falsification field becomes cosmetic. Distinction maintained: thesis-itself breaks → universe removal; entry-timing thesis fails → status downgrade only.
git clone https://github.com/Beltran12138/prophetmap.git
cd prophetmap
npm install
# Required for signal analysis
cp .env.local.example .env.local
# Add GEMINI_API_KEY=AI... (https://aistudio.google.com/apikey)
# Update sector benchmarks (run once or weekly)
node scripts/update-benchmarks.js
# Update pricing scores (live Yahoo Finance data, ~10 min for the equity book)
node scripts/update-valuations.js
# Update crypto pricing scores (CoinGecko + DeFiLlama, no API key)
node scripts/update-crypto-valuations.js
# Analyze thesis falsification signals via Gemini 2.5 Flash
node scripts/analyze-signals.js
# Opt-in to also analyze watchlist tickers (informational, no critical alerts):
INCLUDE_WATCHLIST=true node scripts/analyze-signals.js
# Surface new candidates via Yahoo peer recommendations + Gemini industry verification
node scripts/discover-candidates.js
# Universe audit (promotion/demotion candidates, no writes)
node scripts/audit-universe.js
# Layer audit (correlation, merge flags, layer health)
node scripts/audit-layers.js
# Start the dashboard
npm run devOpen http://localhost:3000.
GitHub Actions runs three workflow groups:
daily— Weekdays 14:30 UTC (after US market close).update-valuations+update-crypto-valuations+analyze-signals. Commits todata/and triggers Vercel redeploy. CRITICAL/HIGH signals create GitHub issues.update-benchmarks— Every Monday 12:00 UTC. Recomputes layer-median P/E and EV/Revenue benchmarks. Without this, pricing scores drift as sector valuations shift.audit— Weekly Sunday 13:00 UTC (universe promotion/demotion) + monthly 1st 11:00 UTC (candidate discovery + layer audit). Surfaces candidates only; never auto-modifies universe.
Secrets required (GitHub repo → Settings → Secrets):
GEMINI_API_KEY— aistudio.google.comVERCEL_DEPLOY_HOOK— Vercel project → Settings → Git → Deploy Hooks
Migration note: signal analysis migrated from DeepSeek to Gemini 2.5 Flash on 2026-05-10 due to DeepSeek API geo-restriction blocking GitHub Actions US runners. Local .env DEEPSEEK_API_KEY reference deprecated.
87 tickers across 28 layers (universe v2.9.10). Defined in data/universe.json with hand-set static fields (thesis, physicalConstraint, constraintType, moatCapture, moatLocks, moatFalsification, aiContribution estimate, timeToRealize, thesisFalsification signals, falsificationObservations). Dynamic pricing scores written daily to data/scores/YYYY-MM-DD.json. Signal alerts written to data/alerts/YYYY-MM-DD.json.
Every hand-set field is a judgement, and the ones that gate anything carry a written
falsifier. constraintType, moatLocks and falsificationObservations carry zero
weight — no gate reads them. They exist so that a future disagreement has something
specific to attack.
Governance protocol in GOVERNANCE.md. Three-tier:
- Automated daily — pricing scores, signal proximity, Gemini falsification check
- Quarterly human review — benchmark recalibration, thesis validity
- Event-triggered — ADD/REMOVE decisions when new physical bottlenecks emerge or theses break (with thesisFalsification trigger response discipline)
Universe versioning follows semantic-style:
- Patch (
v2.1.x):_notehousekeeping, status changes, single-field updates - Minor (
v2.x.0): new tickers, layer migrations, thesisFalsification additions - Major (
vx.0.0): new layers, chain restructuring
Each universe change is recorded in the ticker's _changeLog array and aggregated in CHANGELOG.md.
- Next.js 15 App Router, server components,
force-dynamicrendering - Yahoo Finance 2 (v3) for live equity market data
- CoinGecko + DeFiLlama for crypto pricing (L_DCOMP layer)
- Gemini 2.5 Flash for thesis falsification assessment
- GitHub Actions for automated data pipeline
- Vercel for deployment
Institutional coverage of AI infrastructure is dense at L0 (NVDA, MSFT) and sparse at L2_5–L8_OPT (specialty silicon, packaging, optics). The most durable alpha historically comes from identifying physical bottlenecks before consensus: the company that builds what AI needs next, not what it needs now.
Recent expansion (chains B–E) reflects the thesis that AI capex doesn't stop at the rack — it propagates to power generation (L11/L11_FUEL), grid equipment (L12), commodities (L14), embodied execution (L_EMBI), space (L_SPACE), and defense AI (L_DEF). Each chain has its own funnel discipline; the framework rejects cross-narrative drift.
ProphetMap is the tool I built to systematize that search — keeping personal portfolio decisions separate from the objective signal engine.
MIT