Skip to content

shakemyspeak/first-principle-protocol

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

2 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

FPP v1.3 — First Principle Protocol

Created by Dimitri Wolf

A constraint-to-foundation path mapping system. Decomposes problems to their irreducible verified requirements, locates where the actual blockage exists, and maps possible paths forward. The human decides.


What changed in v1.3

Three targeted fixes from v1.2 cross-platform testing.

Area v1.2 v1.3 Reason
LOOKING FOR validation Term list only Term list + general principle: flag any LOOKING FOR that cannot produce a non-arbitrary path Prevents over- and under-application of the flag list
Multiple classifications "Map each separately" — no mode distinction Conversational: state found classifications, ask which to pursue first. Autonomous: map each separately. Protocol was silent on conversational behavior with multiple classifications
Problem dissolved Reachable from Gate 4 flag alone Requires WHY chain completion. Gate 4 flags produce a halt and a question — not a result. Grok dissolved a CNC spindle problem without first asking if the user had a specific variant — premature dissolution

What changed in v1.2

Eight changes from external technical review and novelty research.

Area Change
Opening statement "...final answers...before the frame is verified" — removes contradiction with RESULT section
TYPE 1 — Physical law Removed "self-verifying if named correctly" shortcut. Both steps must pass.
Stop hierarchy labels Hard/Soft/Open → Fixed/Conditional/Open
Theater detection Reframed: "restates original frame without adding verification" not "confirms prior belief"
Preference conflict reclassification Added: "restart Gate 3 priority scan from item 1"
Gate 4 — GIVEN reintegration Added: update GIVEN when user confirms a specific variant
TYPE 1 — fallback Added UNVERIFIED ASSUMPTION fallback (was missing; all other types had it)
Result Type 6 Added: Problem deferred — for TYPE 7 stops where data is unavailable

What changed in v1.1

Three targeted fixes confirmed by 5-platform comparison testing.

v1.0 v1.1 Reason
Framing gate fires only when no question is present Framing gate also fires when a stated question contains undefined key terms 3/4 platforms answered Q7 directly — "What books?" passed as valid LOOKING FOR despite "best impact" being undefined
Preference conflict → flag as non-FPP-stable, stop Preference conflict → attempt ONE WHY translation, then flag if still unstable Q6 showed "coding without AI" is preference on the surface, capability gap underneath
No definition of conversational vs. autonomous mode Both modes defined explicitly All 4 non-Claude platforms treated the protocol as single-shot

Quick Start

  1. Copy the full protocol text from protocol/fpp-v1.3.md.
  2. Paste it as a System Prompt in Claude, Grok, ChatGPT, DeepSeek, Gemini, or any LLM that supports system instructions.
  3. Present your problem. The protocol runs the gates before responding.

Use FPP when the cost of solving the wrong problem is higher than the cost of questioning the frame.

Do not use FPP for routine decisions, pre-verified A vs. B selections, or problems where the frame is already confirmed correct.


Cross-Platform Test Results

v1.0 baseline — 10 real forum questions, 5 platforms, 50 runs

Q Topic Claude DeepSeek OpenAI Gemini Grok
1 Marketing co-founder Reframed Reframed (deep) Framing gate only Reframed (skipped gate) Dissolved
2 Career + AI anxiety Unstable Framing gate Framing gate Gate 1-4, stopped Unstable
3 Career change 40s Reframed Framing gate Framing gate Framing gate Framing gate
4 Sell software to employer Confirmed + blocked Framing gate Framing gate Full analysis + paths Framing gate
5 Robotics startup Out of scope Framing gate Framing gate "What is the question?" Framing gate
6 Coding without AI Reduced Framing gate GIVEN + stopped GIVEN + questions Dissolved (non-FPP)
7 Upskilling after layoff Reframed Framing gate Answered directly Answered directly Answered directly
8 ML vs SE degree Dissolved Near-dissolved A vs. B detected Partial "Why these options?"
9 Don't want software job Unstable Framing gate A vs. B detected Asked why "Why these options?"
10 VPN market positioning Reframed Full WHY chain Framing gate Full analysis Dissolved

Key finding — Q7: 3/4 platforms answered directly when LOOKING FOR was vague but syntactically complete. Fixed in v1.1.


v1.2 targeted test — 10 gate-specific questions, 5 platforms, 50 runs

Q Target Claude DeepSeek OpenAI Gemini Grok
Q1 GATE 1 — "best" undefined PASS PASS PASS PASS PASS
Q2 GATE 4 + TYPE 1 domain PASS Gate 1 first PASS Gate 1 first PASS
Q3 GATE 1 — A vs. B PASS PASS PASS PASS PASS
Q4 GATE 3 — preference PASS Partial Gate 1 first PASS Partial
Q5 GATE 4 — false premise PASS PASS Partial PASS Partial
Q6 Theater detection PASS PASS PASS PASS Partial
Q7 Result Type 6 — deferred Gate 1 all Gate 1 all Gate 1 all Gate 1 all Gate 1 all
Q8 TYPE 1 step 2 domain PASS Partial PASS Partial Partial
Q9 GATE 3 — multi-class Partial PASS PASS Gate 1 first PASS
Q10 TYPE 2 — citation check PASS Partial PASS PASS PASS

Key findings:

  • Q1: 5/5 PASS — the v1.1 LOOKING FOR fix holds on all platforms
  • Q10: 4/5 PASS — TYPE 2 regulatory citation check is the strongest new gate behavior
  • Q7: 5/5 Gate 1 — all platforms correctly caught "should" as undefined before the WHY chain; test question redesign needed to probe TYPE 7 stop and Result Type 6
  • Result Type 6 appeared unprompted — OpenAI produced "Problem deferred" organically on Q9
  • Theater detection never fired — because Gate 1 caught every problem first (correct design: Gate 1 prevents garbage from entering the WHY chain)

Best single response (v1.2 test): Grok Q2 — "Problem dissolved. The stated modeling approach rests on an inapplicable physical law." One sentence. Correct result type. No excess.


Platform Profiles (v1.2 test)

Platform Gate compliance WHY chain depth Result type usage
Claude High High Full taxonomy
OpenAI High Medium-High Good — produced Result Type 6 unprompted
DeepSeek High High (when reaches chain) Good — Result Type 6 on Q9
Gemini High Low-Medium Stops at gate trigger
Grok High Low Correct, brief

Folder Structure

FPP/
├── README.md                              ← this file
├── LICENSE
├── progress.md
├── examples.md                            ← real-world examples of each result type
├── protocol/
│   ├── fpp-v1.0.md                        ← original
│   ├── fpp-v1.1.md
│   ├── fpp-v1.2.md
│   └── fpp-v1.3.md                        ← current — use this
├── reports/
│   ├── test_run_01.md                     ← Claude single-platform test (v1.0)
│   ├── test_run_01_comparison.md         ← 5-platform comparison (v1.0)
│   └── test_run_02_comparison.md         ← 5-platform targeted test (v1.2)
├── tests/
│   ├── test_design_v1.2.md               ← question design + expected outputs
│   └── runs/
│       ├── v1.2/                          ← raw JSON responses, all 5 platforms
│       ├── deepseek.json                  ← v1.0 baseline results
│       ├── openai.json
│       ├── gemini.json
│       └── grok.json
├── scripts/
│   └── run_test.py                        ← test runner (requires .env with API keys)
└── archive/
    └── concept.md                         ← original research and design notes

Novelty

FPP draws conceptually from: 5 Whys (Toyota), NASA Root Cause Analysis, TRIZ, first principles reasoning, and Ackoff's problem dissolution taxonomy. None of these contain the full combination.

Elements confirmed novel after research across TRIZ, Kepner-Tregoe, INCOSE, NASA SE Handbook, and 2025–2026 LLM reasoning literature:

  • Framing gate as a pre-reasoning halt protocol keyed on undefined LOOKING FOR
  • Typed WHY chain stop conditions with explicit verification requirements (not just iterative asking)
  • Theater detection as a restart trigger keyed on zero verification work
  • Regulatory consequence test (cannot sell / cannot operate / legal liability) as a formal gate
  • Function validation before solving consequences, as a named mandatory protocol step
  • Problem dissolution as a first-class protocol output (not just a theoretical category)

The integrated combination — all elements as a unified pre-reasoning sequence — has no found equivalent.


Sister Protocols

Protocol Purpose
LJP — Lateral Jump Protocol Forces LLMs to make lateral conceptual connections before answering linearly
AMP — Anti-Mirroring Protocol Frame-evaluation layer sycophancy fix

FPP runs first — it validates the frame before any response is constructed. LJP governs how the response is built. AMP governs whether the response is honest. They address different failure modes at different stages.

About

A constraint-to-foundation path mapping protocol for LLMs. Decomposes problems to irreducible verified requirements before any answer is constructed.

Resources

License

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors