Skip to content
Now on model 0.7.0

What’s new.

Every version of the model, dated — so you can see how it’s grown, and hold us to it.

0.6.02026-07-17

The living graph — propagation, substrate fusion & self-measuring scores

  • Made the constraint graph live: a shock on one node now cascades to everything downstream of it (bounded by relation elasticity and a conservative coupling), so a stored node score reflects the whole chain, not just its own news — and the same living scores now feed the conviction verdicts, the Map, the Globe and the bottleneck radar from one computation.
  • Wired the previously-inert substrate into scoring: recent incidents, enforced policies, critically concentrated materials and single-geography concentration now add bounded node tension, and moves in the sourced input-price series (FRED / EIA / EDGAR) nudge their mapped node directly — each a dialable gain that reverts to zero.
  • Added four cross-checks between data types — a news move the market corroborates earns confidence, names exposed to a forecast-tightening constraint lean before it binds, the system-tightness regime tilts the weights (bottleneck leads when tight, momentum when easing), and elite accumulation confirms a standing chain-vs-price divergence.
  • Made the scoring transparent and pilotable: an admin Scoring Console shows every weight beside its default and its measured predictive skill (per data source), editable live with no redeploy; a nightly loop grades the forecast against what actually bound and measures which data type predicts, feeding a shadow self-tuning governor that will bend the weights toward what works once the record matures.
0.7.02026-07-17

Sharper signals — live macro coverage, a guarded top call, and a graph that can loosen

  • Lit up the live macro layer across the chain: real Federal Reserve, energy, and SEC-filing time-series now feed roughly forty of the sixty-seven constraint nodes (up from a handful), so far more of the model reacts to the actual economy — chip and equipment prices, networking and cooling costs, power and financing — rather than only to news; proprietary prices (memory, NAND, optical modules, indium, batteries) are filled from cited reporting as it appears.
  • Guarded the most aggressive verdict: a name can only earn the top "STRONG BUY" rating if it actually owns a binding bottleneck or matches the early-winner profile — a strong chain position alone is no longer enough — so the boldest call is always backed by a structural reason.
  • Made the constraint graph able to LOOSEN, not only tighten: a genuine upstream easing (a capacity expansion, a policy relief) now relaxes the nodes downstream of it, so the model can call the end of a squeeze, not just the start.
  • Sharpened three cross-checks: a news move is "confirmed" when the companies that benefit rise while those that are hurt fall (not a blurred average); the system-tightness regime only re-tilts the weights once it has held for two days (no single-day whipsaw); and the reliability watchdog now catches a collapsed verdict book, not just an empty one. All new behavior is bounded, reviewed against the full universe, and reversible from the admin console.
0.5.02026-07-09

The AI Value-Chain Index + exposure & valuation reads

  • Added the AI Value-Chain Index: a structurally-weighted read on the whole chain, each constituent weighted by chain centrality × √market-cap, capped per name with the excess redistributed, plus an HHI concentration read — published as a live, dated public index.
  • Added three per-company exposure lenses — material concentration, regulatory intensity (weighted by policy status) and supply-vs-demand — each shown only where the underlying data supports it, and left blank otherwise.
  • Added moat-erosion and priced-in valuation reads (structural value as a band, never a point target) plus a narrative-vs-structure divergence read that flags when a thesis’s story runs ahead of the model.
  • Surfaced the chain’s leading input-price series (FRED producer-price indices, EIA energy) as a sourced cost tape that self-hides when no sourced series are available.
0.4.02026-06-20

Provable bottlenecks + calibration

  • Added a max-flow / min-cut engine: the binding-constraint set is now the proven minimum cut of the Energy→Applications flow network, not a heuristic rank.
  • Added a backtest harness that replays authored domino scenarios through the propagation engine and reports recall, precision and a confidence-vs-reproduction reliability diagram.
  • Introduced model versioning + this changelog; forecast and nowcast outputs are now stamped with the model version.
0.3.02026-05-15

Quantified relations

  • Seeded elasticity and lag (in days) on the flagship causal relations, so propagation can reason about effect size and timing — not just direction.
  • Added critical-materials, geography and policy entities with exposure tables.
0.2.02026-04-02

Live overlay

  • News is ingested from public feeds, conservatively mapped to nodes, and decays by half-life into the live score — with a seed fallback when a source is unavailable.
  • Market quotes are shown as display-only signals and never move scores.
0.1.02026-03-01

Hybrid scoring

  • Seven sub-scores (structural, tension, news, cross-layer, investability, fragility, timing) combined by fixed weights into a global score, with the news/tension/timing components recomputed from context.