Query proof

Chart a community's drift along a meaning axis with SQL

A Scry query pattern that mints two parallel pole sentences as vectors, takes the balanced axis between them, and averages every LessWrong chunk's projection quarter by quarter, centered on the era mean.

Answer shape

Two embed calls mint the poles; one SQL query joins chunk embeddings to real post timestamps, averages the projection per quarter, and a window function centers each quarter on the era mean so the chart shows movement rather than the axis's arbitrary zero.

Sources

  • LessWrong

Methods

  • compositional embeddings
  • contrast axis
  • source-native JOIN
  • time series
  • window function

SQL

Run this through POST /v1/scry/query after inspecting GET /v1/scry/schema for current live fields.

-- Mint the poles once, then query:
-- POST /v1/scry/embed {"name": "achievable", "text": "Robust alignment seems achievable by defense in depth: iterative training, scalable supervision, interpretability, red-teaming, corrigible tool use, formal checks where possible, and institutions that prevent single-point failure. No one technique must be perfect if several imperfect safeguards compose well."}
-- POST /v1/scry/embed {"name": "unreachable", "text": "Robust alignment seems unreachable by defense in depth: training optimizes proxies, supervision fails once systems surpass us, deception can hide until deployment, interpretability is incomplete, and institutions face racing pressure. Combining several imperfect safeguards still leaves correlated catastrophic failure modes."}
SELECT quarter, lean - avg(lean) OVER () AS drift, posts
FROM (
  SELECT toStartOfQuarter(p.original_timestamp) AS quarter,
         avg(scry_cosine_similarity(e.embedding,
             scry_contrast_axis_balanced(@achievable, @unreachable))) AS lean,
         uniq(e.target_key) AS posts
  FROM embeddings.chunks AS e
  JOIN forums.posts AS p ON p.post_key = e.target_key
  WHERE e.source = 'forum_posts' AND e.model_name = 'voyage-4-lite'
    AND p.source = 'lesswrong' AND p.original_timestamp >= '2009-01-01'
  GROUP BY quarter)
ORDER BY quarter
LIMIT 100;

Verification notes

  • NoteVectors are named SQL values, minted from any text via POST /v1/scry/embed.
  • Notescry_contrast_axis_balanced(@achievable, @unreachable) is the axis between the two poles; the poles are parallel sentences that differ only in stance, so shared topic words cancel.
  • NoteEach quarter keeps uniq(target_key), so the sample size behind every point stays visible.
  • Notelean - avg(lean) OVER () centers on the era mean: the zero of a contrast axis only means equidistant from the two sentences.