A Constraint‑Driven Lens on Evolution Across Physical, Biological, Cognitive, Social, and AI Systems
Zenodo (all versions): https://doi.org/10.5281/zenodo.18012401 | Latest release (v2.4.1): https://doi.org/10.5281/zenodo.18112020
Papers (Zenodo):
- Start here — FIT (Force-Information-Time) Dynamics: Origin and Design Goals: https://doi.org/10.5281/zenodo.18142211
- AI safety — Irreversible Operations and Tempo Mismatch in AI Learning Systems: https://doi.org/10.5281/zenodo.18142151
- AI safety (grokking) — Controlled Nirvana: Emptiness Windows as a Structural Safety Mechanism for Post-Grokking AI Systems: https://doi.org/10.5281/zenodo.18155425
Case studies:
- Li² Grokking scaling-law replication (Tian, 2025; independent)
- Smartphones and the Attention System
- Content Platforms and Involution
- Learning: From Memorization to Understanding
Note: names are cited as paper authors; no affiliation or endorsement is implied.
Structural archetypes (patterns):
Current spec (v2.4.1): docs/v2.4.md
Framework established: Dec 10th, 2025 (original)
Author: Qien Huang (Independent Researcher)
E-mail: qienhuang@hotmail.com
License: CC BY 4.0
Repository: https://github.com/qienhuang/F-I-T
ORCID: https://orcid.org/0009-0003-7731-4294
Recommended order for first-time readers: Core Card → Current Spec (v2.4) → Entry points
- Fastest entry (Core Card, v2.4.1+): docs/core/fit_core_card.md
- Current spec (v2.4.1, EST + Tier‑1 validation): docs/v2.4.md (EN), docs/zh_cn/v2.4.zh_cn.md (中文/Chinese)
- Previous edition (v2.3, Tier‑1 validation): docs/v2.3.md
- Legacy discussion edition (v2.1): docs/v2.1.md
- Changelog: CHANGELOG.md
- Versioning policy: docs/core/Versioning_Policy.md
Stability note: 2.x core is stabilized; revisions are counterexample-driven. See Versioning Policy.
From quantum and molecules to cells, individuals, organizations, nations, and civilizations—why do clearly defined hierarchical structures emerge? Why does evolution often manifest as a repeating rhythm of "oscillation—stability—aggregation—re-stability"? Why do many systems fail not because of insufficient power or lack of information, but because the "pace of doing things" is wrong?
I attempt to answer these questions by compressing "evolution" into three minimal variables:
- Force (F): The action that drives or constrains system change (interactions, selection pressures, institutional constraints, objective function gradients).
- Information (I): Structures that can persist in time and produce causal effects (codes, forms, patterns, models).
- Time (T): Not a background scale, but a spectrum of characteristic time scales (rhythms) emergent from the interaction of F and I.
F‑I‑T is a meta-framework, not a theory of a specific domain.
Its purpose: first reduce any problem of "evolution, development, origin, collapse, innovation" to (F, I, T), then discuss levels, critical points, and transition paths.
What FIT does NOT claim:
- ❌ A "theory of everything" for complex systems
- ❌ Replacement for existing frameworks (Free Energy Principle, Constructor Theory, etc.)
- ❌ Ability to predict exact trajectories of complex systems
- ❌ That all propositions have been validated across all domains
What FIT DOES claim:
- ✅ A minimal meta-language for discussing evolution across domains
- ✅ Falsifiable through computational and empirical experiments
- ✅ Initial Tier-1 validation shows promising results in controlled systems
- ✅ Applications to AI safety and complexity science are tractable
The problem: Modern science approaches evolution through fragmented lenses (thermodynamics, information theory, complexity science, ML). They succeed in isolation but lack shared axioms for cross-domain synthesis.
FIT's response: Compress "evolution" into five primitives and six principles. Generate 18 falsifiable propositions bound to explicit estimator tuples.
Core insight: Many systems fail not from lack of power or information, but because high-impact changes become irreversible faster than correction can occur. FIT treats tempo (correction timescales) as a first-class variable.
The five primitives:
| Primitive | Definition | Interpretation |
|---|---|---|
| State (S) | System configuration at time |
|
| Force (F) | Generalized drift / directed influence | |
| Information (I) | Entropy reduction / knowledge gain | |
| Constraint (C) | Reachable state space reduction | |
| Time (T) | Ordered index |
Emergent from F–I interaction |
v2.4 key features:
- Estimator Selection Theory (EST): 8 admissibility axioms (A1–A8) preventing "estimator-hacking" critiques
- 18 falsifiable propositions with explicit success/failure criteria
- Tier-1 validation: 97.5% theory–observation match (Langton's Ant), P7 bounds 0% violations (Conway's GoL)
- AI safety track: tempo mismatch + Irreversible Operations as distinct failure mode
Read the full spec: docs/v2.4.md
- FIT for AI Safety (start here): docs/ai_safety/fit_ai_safety_mapping.md — 5-min overview + 2-hour self-assessment checklist
- Two-week pilot (teams): proposals/tempo-io-pilot.md + proposals/tempo-io-pilot-pack/
- Self-referential IO standard: docs/ai_safety/self_referential_io.md + docs/ai_safety/io_sr_mapping.md
- Case note (R1-style RL + risk control): docs/ai_safety/deepseek_r1_fit_case_note.md — defense-in-depth: content gating + action gating
- Runnable demo: examples/self_referential_io_demo.ipynb + examples/run_demo.py
- Tier-2.5 demonstration (preregistered): experiments/real_world/nyc_311_tier2p5/ — NYC 311 service requests; applying FIT metrics to real-world data (not a validation claim)
- arXiv anchor draft (IO × tempo mismatch): papers/irreversible-operations-tempo-mismatch.arxiv.compact.md
This is a preregistered demonstration (not a "real-world validation" claim). The vertical marker indicates the created-date boundary: arrivals are filtered to 2024 by construction, while closures may continue into 2025 (a closure tail that is out of scope for H1 under this boundary).
Current reading (HPD 2024; W=14, H=14): coherence gate PASS, but H1 is INCONCLUSIVE because there are 0 in-scope tempo-mismatch events (rho stays < 1 within the created-date boundary).
Reproducibility + guardrails: prereg_v3.yaml and experiment README.
- Langton's Ant (open boundary): 97.5% theory–observation match for net displacement; supports key phase-transition / nirvana predictions.
- Conway's Game of Life: P7 information bounds (0% violations), P10 estimator coherence (rho = 0.775); P2 constraint monotonicity challenged under current estimator.
Figure: Conway's Game of Life Tier‑1 validation snapshot (details in docs/v2.4.md).
| Milestone | Goal | Horizon |
|---|---|---|
| M0 | Stabilize 2.x spec; publish Tier-1 validation scripts (GoL, Langton's Ant) | 0–3 mo |
| M1 | Reference Python implementations; 5–8 propositions with reproducible status | 3–9 mo |
| M2 | Continuous-time FIT (SDE layer); prove constraint-accumulation theorem | 6–18 mo |
| M3 | Quantum FIT (Lindbladian layer); demonstrate quantum analogues of P2/P3 | 9–24 mo |
| M4 | Merge discrete / continuous / quantum into unified v3.0 | 18–36 mo |
| M5 | Applications: AI safety, complexity science, institutional design | ongoing |
Full roadmap: docs/roadmap.v2.4.md
docs/- specifications and notesproposals/- practitioner pilots and templatesdocs/ai_safety/- self-referential IO and governance docsexamples/andexperiments/- runnable demos and validation artifactspapers/- drafts and venue-specific writeupsCITATION.cff- citation metadata for this repository
Use CITATION.cff for copy/paste formats, or cite via Zenodo:
- Zenodo (all versions): https://doi.org/10.5281/zenodo.18012401
- Latest release (v2.4.1): https://doi.org/10.5281/zenodo.18112020
Text and documentation in this repository are licensed under CC BY 4.0.
Portions of drafting and editing were assisted by large language models. The author takes full responsibility for all content, claims, and errors.
