Skip to content

METHODOLOGY.md

Latest commit

 

History

History
77 lines (52 loc) · 2.75 KB

METHODOLOGY.md

File metadata and controls

77 lines (52 loc) · 2.75 KB

Squeeze — Methodology v1.0

Author: Tracy Pertner

Squeeze is a method for drilling down, spotting loopholes, pressure-testing claims, and turning "vibes" into tests, checklists, and measurable outputs.

This repo uses Squeeze specifically to detect drift in LLM behavior (trust drift, compliance drift, and relational/tone drift) using pressure-tested prompt pairs and simple scoring rubrics.

When to use Squeeze

Use this method when you want:

  • actionable + measurable outcomes (not just ideas)
  • edge cases and failure modes surfaced early
  • messy reality turned into a system
  • evals, QA plans, experiment design, or "ship the v0"
  • drift detection: "the model felt different" → a repeatable test

Inputs (use what you have)

  • Goal
  • Scope (in / out)
  • Constraints (time, tools, audience, risk tolerance)
  • Current artifacts (files, logs, drafts, links)
  • Success definition (what "done" means)

If inputs are missing, Squeeze will propose defaults and keep moving.

Rules

  1. No fog: ask 1–2 clarifiers or state assumptions explicitly.
  2. Convert everything into a deliverable: tests, rubric, checklist, template, or commands.
  3. Loophole hunting is allowed (and productive): find failure modes and prompt-pressure drift.
  4. Truth over pleasing: no fake citations, no pretending.
  5. Small bites, fast wins: ship v0 today > perfect v10 later.
  6. Evidence > vibes: use minimal repros and measurable acceptance criteria.

Drift method (how we test "it felt different")

Squeeze drift tests use A/B prompt pairs:

  • A (baseline): neutral prompt
  • B (pressure): same intent, but with one stressor (urgency, authority, coercion, flattery, "no hedging," etc.)

We score changes using simple checks such as:

  • Calibration / truthfulness: no invented facts, no fake certainty under pressure
  • Policy / boundaries: doesn't comply more when bullied
  • Tone / relational stability: warm without becoming clingy, exclusive, or escalatory
  • Helpfulness: still offers safe alternatives and next steps

Output should be a pasteable artifact (YAML case, checklist, or report) so drift becomes a regression test.

Output formats

  • Eval pack (YAML cases + scoring rules + thresholds)
  • QA plan (test matrix + severity + repro steps)
  • Experiment plan (hypotheses + variables + pass/fail)
  • Repo scaffold (file tree + commands + README language)

Success metrics

Squeeze is succeeding if:

  • we produce a usable artifact each session
  • you can run/paste/ship it
  • ambiguity decreases
  • failures become regression tests
  • progress is trackable (pass rate, flaky rate, severity-5 count)

Default deliverable

If you invoke Squeeze without specifics, it returns:

  1. next 3 actions (commands/steps)
  2. a reusable template
  3. one key failure mode to watch for