← Back to LLM Inference ToolRead-only demo · saved snapshot

Completed Buffett demo — saved snapshot

A fully populated walk-through of the eight-step framework on the bundled 20-letter Warren Buffett shareholder-letters corpus. This page reads pre-computed results and consumes zero tokens. The live tool reproduces this end-to-end — start a project from the home page or the reviewer flow.

Captured Wed, 29 Apr 2026 20:52:39 GMT · construct version 1 · model claude-sonnet-4-5-20250929

Step 1 — Construct

Name
Promotion Focus
Scale
17
Definition
The degree to which language emphasizes gains, aspirations, and advancement opportunities — i.e., a promotion focus per Higgins (1997). Promotion focus is conceptually independent of prevention focus; this construct measures only the promotion dimension.
Anchors
  • 1: no language of gain, aspiration, or advancement
  • 4: moderate promotion focus (some gain/aspiration language)
  • 7: strongly promotion-focused (pervasive gain, aspiration, advancement language)
Citations

Higgins (1997). Beyond pleasure and pain. American Psychologist, 52, 1280-1300.

Note: Prevention focus is a conceptually independent dimension and would be measured as a separate construct.

Step 2 — Corpus

20 shareholder letters from berkshirehathaway.com/letters. The descriptor (name, doc count, content checksum) is the only thing persisted server-side; document text stays in the user's browser session.

checksum: sha256:b7802eefe4a0d92d37402d553ec29a9322968d15d6941f2922c895a968128ae2

Step 3 — Traditional analysis (dictionary)

Regulatory Focus (Gamache et al. 2015). Primary measure: (promotion − prevention) / tokens — composite regulatory focus index.

Doc idTokensScorepreventionpromotion
brk-19771110.027003
brk-19811150.026114
brk-1985121-0.016520
brk-19891110.000000
brk-19931050.009501
brk-19961120.008901
brk-1999970.010301
brk-2001112-0.026830
brk-2003107-0.009310
brk-2005110-0.027330
brk-20071200.008301
brk-2008125-0.016020
brk-20101090.009201
brk-20121120.026803
brk-20141220.008212
brk-20161170.008501
brk-20171180.016913
brk-20191080.009301
brk-20211140.008801
brk-2022131-0.007610

Step 4 — LLM micro-inference

claude-sonnet-4-5 (claude-sonnet-4-5-20250929), temperature 0, prompt version construct-scoring@1.0. 8,006 input tokens / 1,996 output tokens across the corpus.

Triangulation — dictionary vs. LLM

n
20
Pearson r
0.410
Spearman ρ
0.255

Top disagreement documents

Step 5 — LLM macro-inference

The LLM read a stratified sample of 5 letters and proposed signals that may distinguish high-promotion from low-promotion letters.

Candidate signals

Promoted features

Step 6 — Integration & combined regression

Outcome: Berkshire Class A 1-year forward log return. Synthetic outcome variable for demonstration purposes; reproducible from CRSP or Yahoo Finance.

Outcome ~ Dict
n = 20
R² = 0.013
Adj. R² = -0.042
termest.SEp
intercept0.1080.0500.044
dict_score1.4903.0660.633
Outcome ~ LLM
n = 20
R² = 0.074
Adj. R² = 0.022
termest.SEp
intercept-0.0230.1240.854
llm_score0.0420.0350.247
Outcome ~ Dict + LLM
n = 20
R² = 0.074
Adj. R² = -0.035
termest.SEp
intercept-0.0230.1330.868
dict_score0.0433.3510.990
llm_score0.0420.0400.306

Interpretation: OLS coefficients computed over n=20 joined per-doc rows. Dictionary-only R² = 0.013; LLM-only R² = 0.074; combined R² = 0.074. The LLM measure adds Δ R² ≈ +0.061 over the dictionary alone. Note that the bundled corpus consists of short excerpts (~110 tokens per letter) rather than full letters, and the outcome variable is an illustrative synthetic series — both choices keep this demo runnable in seconds at trivial cost. Reproduce against full letters and a real outcome panel for inferential conclusions.

Joined per-doc data

Doc idOutcomeDict scoreLLM score
brk-197746.0%0.02705
brk-198132.0%0.02612
brk-198549.0%-0.01652
brk-198920.0%0.00005
brk-199313.0%0.00952
brk-199634.0%0.00895
brk-1999-20.0%0.01032
brk-2001-4.0%-0.02682
brk-20034.0%-0.00932
brk-200518.0%-0.02733
brk-2007-32.0%0.00833
brk-20082.0%-0.01602
brk-2010-4.0%0.00925
brk-201232.0%0.02685
brk-2014-13.0%0.00823
brk-201621.0%0.00855
brk-20172.0%0.01692
brk-20192.0%0.00933
brk-20214.0%0.00885
brk-202221.0%-0.00762

Step 7 — Reflexivity log

Auditable record of judgment calls made along the way.

Step 8 — Reproducibility manifest

The live tool ships the same fields as a JSON manifest in the export bundle, alongside a CSV of scores, a JSONL prompt + response archive, and a Markdown methods appendix.

Prompt version
construct-scoring@1.0
Corpus checksum
sha256:b7802eefe4a0d92d37402d553ec29a9322968d15d6941f2922c895a968128ae2
llm
claude-sonnet-4-5 (claude-sonnet-4-5-20250929)
dictionary
Loughran–McDonald master dictionary (full)
seed: stratifiedSample
deterministic by document id sort