The derivations

How every number on live.somen.ai is derived. All gauges re-compute from the

knowledge graph at every render; the brief is template-rendered from the same

derived numbers and cannot drift from the record.

THE SPREAD

  • Question: where do companies' secondary-market tapes trade relative to the

    • print their own last round set?

  • Derivation: for every company with a funded round (any source, dated by

    • evidence_date else observed_at) AND tape coverage at that date, compare the

    tape print within +/-5 days of the round date against the latest print

    (both per-share - no share-count arithmetic). Weekly-return outliers beyond

    5x are excluded as units glitches.

  • Reported: % of the comparable set trading below the round-date print; the

    • median; and n (the sample).

    BLIND MONEY (2026)

  • Question: how much money was raised in filings that disclose no valuation?
  • Derivation: SEC Form D funded edges with evidence dates in 2026, an amount,

    • and no post_money; per company the MAXIMUM amount (D/A amendments restate

    cumulative amounts, so summing would double-count).

    THE CHAIN

  • Question: how narrow is the dependency stack?
  • Derivation: distinct companies with any customer_of / depends_on_commercial

    • / depends_on_hardware edge pointing at NVIDIA, TSMC, ASML, and the two

    German optics sources; live counts at every render.

    THE CIRCULAR SET

  • Question: which companies buy from suppliers that hold their equity?
  • Derivation: strategic-vehicle alias map (NVentures/NVIDIA, Samsung

    • Catalyst/NEXT, Intel Capital, Google/Alphabet, M12/Microsoft, Qualcomm

    Ventures, ...) over invested_in portfolios, intersected with dependency

    edges; philanthropic vehicles (e.g. NVIDIA Foundation) are excluded.

    BREADTH

  • Question: the private market's advance/decline line.
  • Derivation: weekly tape moves (baseline = the median print 3-11 days back,

    • corroborated by neighbors - a lone thin print cannot manufacture a move);

    advancers > +1%, decliners < -1%.

    GROUND TRUTH

  • Question: the physical layer, measured from orbit.
  • Derivation: Sentinel-2 L2A (Copernicus) 10m window reads per site,

    • operator-confirmed before entering the graph (site_observation table);

    built-surface percentage deltas x window area, summed across campuses.

    Method badges: P pixel / L llm-draft / H human-counted.

    THE LEDGER

  • Question: the track record, forward-framed.
  • Derivation: the public scorecard - published claims with dated falsifiers

    • and graded resolutions (data.somen.ai/scorecard).

    The daily brief

    Template-rendered from the gauges above; zero LLM. If a number is in the

    brief, it is in the graph; the brief cannot state what the graph does not.

    Honesty rules

  • n/a renders as n/a, never a recomputed substitute.
  • Stale tapes render dimmed with their as-of date.
  • What-changed chips are signed deltas vs the previous hourly render.
  • z-scores ("a 3.2 sigma week for this name") compare a weekly move against

    • that name's OWN weekly-return history (>= 8 weeks, glitch-guarded; names

    with regime steps render no sigma - it would be meaningless).

    Rendered 2026-10-21T07:12:18Z from the live graph.