PHYSICAL ALPHA

A Choke-Point Layer Rotation Framework for AI & Robotics Value Chains

Working Paper — Methodology v1.0 (DRAFT) Author: DDD · Aither Labs Version: 1.0.0-draft · Date: March 2026 · License: Proprietary public excerpt. Implementation, calibration, live signals, maintained maps, and registry feeds are proprietary and separately licensed — see Appendix B.

Living-paper convention: statistics marked proprietary are auto-filled from
the AlphaOS evaluation database (signal_eval, book_nav) at render time
when released. If a token has no public backing export, it renders
PROPRIETARY - pending public release — this paper never states an
unmeasured number.

Abstract

The AI and robotics build-out is a sequence of physical bottlenecks — power generation and interconnection, advanced memory, advanced packaging, optics, grid equipment, actuators — that migrate through the value chain faster than equity markets re-rate the layers that own them. We propose a framework, Choke-Point Layer Rotation (CLR), for measuring this migration directly. We define a layer-level tightness state τ from physical and contractual indicators (capacity auctions, lead times, backlog, contracted revenue), a re-rating stage ρ from price-derived crowding, and a tradable Tightness Gap, G = z(Δτ) − z(ρ): rising real-world tightness not yet reflected in relative strength. Three hypotheses follow: (H1) tightness leads layer-level equity re-rating; (H2) signal provenance quality — fresh causal fundamentals versus stale or price-proxy fallbacks — predicts realized edge; (H3) high re-rating stage predicts forward underperformance. The framework is implemented in an open evaluation architecture with point-in-time discipline (EDGAR-dated fundamentals, survivorship labeling), cross-sectional testing (rank IC, Newey-West inference, cost-adjusted quantile spreads, deflated Sharpe ratios), and — unusually for a practitioner methodology — a pre-registered falsification registry: dated public events specified in advance that would disconfirm each hypothesis. Current out-of-sample tightness-gap ICIR: PROPRIETARY - pending public release. Anticipation-vs-Crowded spread since inception: PROPRIETARY - pending public release.

1. Motivation

• The installation phase of a general-purpose technology is capital-led and bottleneck-sequenced: railway iron before rolling stock, fiber before routers, physical capacity before the market recognizes the layer that owns it.
• The 2023–2026 AI sequence so far: GPUs → HBM/CoWoS → power → grid equipment → proprietary next-layer watchlist. Each hand-off re-rated a different layer.
• Claim to establish: the binding constraint is observable in physical data before it is priced, because physical data (auction results, lead times, filings) publishes on regulatory and contractual calendars, not market ones.
• Position against alternatives: this is not momentum (H3 is anti-momentum at the layer level), not thematic indexing (layers, not themes), not fundamental stock-picking alone (the unit of analysis is the layer).

2. Definitions

2.1 Layer ontology. The AI/robotics value chain is partitioned into layers L = {compute, memory, packaging, optics, power generation, grid & electrical equipment, cooling, robotics actuation, …} via a maintained ticker→layer map. The ontology is versioned; changes are logged. (Public: ontology structure. Licensed: the maintained map and its history.)

2.2 Tightness τ. Per layer, a composite of causal physical/contractual indicators — capacity auction clearing prices, lead times, backlog-to-revenue, contracted-revenue share, utilization — each timestamped at public availability. τ is a state; the signal uses its momentum Δτ.

2.3 Re-rating stage ρ. Price-derived crowding of the layer: trailing relative strength and valuation richness percentile. ρ measures how much of the story equities already carry.

2.4 Tightness Gap.

G_ℓ(t) = z( Δτ_ℓ(t) ) − z( ρ_ℓ(t) )

Cross-sectionally standardized across layers at t. High G: physical tightness building faster than the market has re-rated. G is the framework's ranking variable for layer allocation.

2.5 Provenance tiers. Every τ input carries a tier at read time: FRESH (causal, current), STALE (>45d), MISSING, FALLBACK (price proxy only). Composite scores are multiplicatively haircut by tier. This is a testable design choice, not a convention (see H2).

2.6 SPREAD_AC. Two rule-maintained books — ANTICIPATION (high-G exposure) and CROWDED (high-ρ exposure) — whose NAV ratio is the framework's public proof-of-life series.

3. Hypotheses

H1 (Lead–lag). Layer tightness momentum positively predicts layer-relative forward equity returns at 21–63 trading-day horizons. Test: rank IC of Δτ (and of G) vs forward layer-proxy returns; Newey-West t at lags = horizon. Current: IC(21d) = PROPRIETARY - pending public release, NW-t = PROPRIETARY - pending public release — proprietary until Signal Lab public release.

H2 (Provenance). Decisions and scores backed by FRESH provenance realize higher forward relative returns than STALE/FALLBACK-backed equivalents; i.e. the haircut ordering is empirically correct. Test: decision-outcome stamping grouped by provenance tier at decision time; minimum n=10 per cell. Current: PROPRIETARY - pending public release.

H3 (Re-rating exhaustion). High ρ predicts negative forward layer-relative returns; chasing a re-rated layer is the framework's canonical error. Test: IC of −ρ vs forward returns; Q5–Q1 spread of ρ-sorted layers, net of costs. Current: PROPRIETARY - pending public release.

Economic rationale: physical data is unglamorous, layer ontologies are maintenance-heavy, and institutional mandates are sector-based, not layer-based.

4. Data & Provenance Architecture

• Point-in-time discipline: fundamentals dated by EDGAR filing timestamps (report_date_source = EDGAR, fallback share disclosed: PROPRIETARY - pending public release); universe membership is point-in-time with survivorship flags carried into every statistic.
• Snapshot and analyst-judgment inputs are structurally banned from backtests (enforced in code, not by policy).
• Public architecture is disclosed at the level of the data spine, provenance tiers, evaluation engine, books, decision log, and outcome stamping. Implementation details remain proprietary.

5. Empirical Protocol

AlphaOS follows a proprietary research protocol for all statistics in this paper: Spearman rank IC and ICIR; Newey-West inference for overlapping horizons; quintile spread portfolios reported gross AND net of a disclosed cost model; in-sample/out-of-sample split anchored to the March 2026 methodology release and reported separately; deflated Sharpe ratios against the count of specifications tried; minimum-sample masking. Public results cite spec hashes inline when released.

6. Portfolio Implementation (CLR in operation)

The seven-step operating protocol: P1 Provenance check → P2 Macro regime gate → P3 Layer rotation (rank by G) → P4 Reallocation discipline (cut/delever/wait taxonomy — sell-side rules precede buy-side sourcing) → P5 Implementation (constraint-aware advisory sizing from freed capital) → P6 Falsification watch → P7 Review (decision efficacy & calibration). The detailed operational mapping is proprietary.

7. Pre-Registered Falsification Registry

The registry lists, in advance and with dates, observable events that would disconfirm each hypothesis (e.g., capacity auction outcomes, HBM qualification timelines, interconnection-queue reforms), each mapped to H1/H2/H3 and to the positions it would invalidate. Registry is append-only; resolutions are logged with outcome and action taken. Current registry: PROPRIETARY - pending public release.

Design statement: a framework that cannot name what would prove it wrong is marketing. This section is the framework's core credibility asset.

8. Limitations

Limitations: ~4-year fundamental depth (installation-phase regime only; no deleveraging trough in sample); approximate universe backfill; layer ontology involves judgment; several τ inputs are manual until adapters land; H2 requires decision-sample accumulation; nothing herein is investment advice.

9. Version History

Appendix A — Landing Page Copy Block

Hero:

Bottlenecks migrate. Prices lag. We measure the gap.
Physical Alpha is a research framework for the AI & robotics build-out:
layer-level tightness from physical data, tested point-in-time, published
with its own falsification criteria.

Three-column strip:

• Measured, not narrated. Every signal carries IC, sample, and cost-adjusted spread — or says "pending evaluation."
• Point-in-time or it doesn't count. EDGAR-dated fundamentals, survivorship labels, leak checks enforced in code.
• Falsifiable in public. Kill criteria pre-registered with dates. View the public registry.

Proof-of-life module: live SPREAD_AC chart + next three falsification dates. Primary CTA: Read the methodology (v1.0) · Secondary: Open the workbench Specimen: proprietary research memo with provenance lineage, available by request.

Appendix B — License Structure

Appendix C — Regulatory Note

Framework publication (Ring 1) is educational/research content. Ring 2 commercial distribution of live signals, model portfolios, or an investable index may constitute regulated investment advice or index administration depending on jurisdiction and format — obtain counsel before commercial launch. This document is not investment advice.