title	TAF Agent
emoji	🔬
colorFrom	blue
colorTo	green
sdk	static
pinned	true
license	apache-2.0
short_description	29 anti-bullshit LLM diagnostics in browser. 4 langs.
tags	transformer llm evaluation diagnostic anti-bullshit long-context sliding-window yarn chat-template arena-elo contamination quantization drift ruler niah lean-mathlib formal-verification rope kv-cache gguf llama-cpp ollama viability thermodynamics free browser webgpu mamba ssm rwkv linear-attention memory
language	en es fr zh

🔬 TAF Agent

Diagnose any transformer LLM in 30 seconds. Free. No GPU. No signup. 28 browser-only modes · 4 languages · 37 Lean+Mathlib-verified theorems · 0 telemetry.

🌐 Live: https://karlesmarin.github.io/tafagent · HF Space: https://huggingface.co/spaces/karlexmarin/taf-agent 📦 Source: https://github.com/karlesmarin/tafagent · Lean repo: https://github.com/karlesmarin/lean-taf 📄 Papers: Predicting How Transformers Attend (Marín 2026) — a continuing I → II → III series: Part I · Part II · Part III · Lean formalization (each explained below) 🗂️ Dataset: taf-attention-decay (58 measurements, 32 models)

---

🆕 v0.13 — Depth map (Part III) + the papers as a series

The third paper — Predicting How Transformers Attend, Part III — opens a second, orthogonal axis, and this release brings it into the tool.

🪜 Depth map — everything before this measured a model along distance (γ — how far attention reaches). Depth map reads the depth axis straight from config.json: the two independent geometric axes side by side (distance γ_geom ⟂ depth L_crit/N), the layer where the attention regime changes (L_crit/N = log_θ(T/2π)·(d_head/64)^(1/√12)), the three RoPE scales (λ₀ · T_cross · T_max), and — for models in the shipped γ atlas — the δ band-position lead (γ_obs − γ_geom) as an exploratory hint of where the J-lens advantage band sits. Built with Part III's honesty caveats in front: L_crit is not the commitment knee, and the three depth observables (transport / writing / commitment) are measured, not derived — so the mode says so and doesn't fake them. Browser-only, no inference, full 4-language support (mode + guided demo + autocomplete + manual). Try it: ?demo=depth.

📚 The papers, now as a continuing series — the docs no longer point at one paper. Manual, inventory and README now present Predicting How Transformers Attend as a I → II → III sequence and explain what each part contributes: Part I (the distance axis + γ atlas), Part II (the six-axis decomposition + learned imprint ν = −1/2π), Part III (the depth axis) and the Lean 4 / Mathlib formalization backbone (52 theorems, 0 sorrys). New dedicated v0.13 manual card, matching inventory entry, and a fixed/expanded BibTeX.

🆕 v0.12 — Fit Check + everything in your language

Two things, both born from what the community actually asked for:

💾 Fit Check ("will it fit?") — the #1 recurring question on the HF forum, answered before you download anything. Model + precision (fp16/bf16/int8/nf4 or any GGUF quant) + GPU + target context → the full VRAM budget (weights + KV cache + scratch), a verdict that tells you which side is the problem (weights-bound vs KV-bound — at 128K the KV cache of an 8B model can equal its weights), the max context that does fit, and the cheapest rescue (quantized cache → lower precision → partial offload). Same budget math as Launch Flags — the two can never disagree. Try it: ?demo=fitcheck.

🌍 Demos and results in your language — until v0.11 the menus were translated but the guided demos and the recipe results (reasons, actions, step interpretations) came out in English. Fixed at the root: demos follow the browser language automatically, and the Python recipe engine now emits message codes the UI localizes into EN/ES/FR/ZH (all 8 recipes migrated; English kept as fallback). Guarded by a real-browser regression test (test_lang.mjs: Spanish locale, clean storage, full demo + full Profile, zero English residue) and a new CI job that sweeps every mode in every language on each push.

🆕 v0.11 — Guided demos for every mode (self-teaching · 4 languages)

Every one of the 28 modes now has a 🎬 Demo button that runs a short guided walkthrough — watch the tool work, understand the result, then repeat it with your own data.

What's new	Detail
🎬 Demo button on every mode	Or open ?demo=<mode> (e.g. ?demo=profile). Step banners fill + highlight the inputs, click the real buttons, produce the real result, then append a plain-language "What this tells you" panel — section by section, with honesty caveats.
🌍 Fully localized	EN / ES / FR / ZH (230 demo strings, idiom-reviewed; standard terms like KV-cache, γ, θ kept in English).
🔗 Self-teaching & shareable	The companion field guide deep-links its Try it boxes straight into the matching demo.

Browser-only, no inference, no new dependencies — index.html untouched (buttons inject via JS). Full notes: docs/RELEASE_NOTES_v0.11.md.

▶ Try it now

---

🆕 v0.10 — Architecture-aware + Reality-check (2 new modes + confidence)

Three additions, all browser-only — no inference, no server.

Mode	What it answers
🧠 Memory Reality Check	"My model says 128k / 1M context — what does that actually mean?" Detects the architecture (full-attention / SWA / SSM-Mamba / RWKV / linear / TTT / hybrid) from config.json and tells you how its memory really works and how it fails — e.g. why a Mamba/RWKV model misses an exact needle. Detection validated against 13 live HF configs.
📊 Prediction vs Reality	"Don't take the numbers on faith." Compares TAF's predictions against MEASURED values (the shipped dataset or your Diagnose-CLI JSON), with a confidence score — and lets you contribute your measurement back to the public dataset, server-less.
✅ Confidence score	Every viability verdict now carries a 0–100% confidence with a ✓/⚠ evidence checklist (γ measured vs closed-form, validated regime, benchmark available). Predictions are never shown as absolute truth.

Plus: HF model-id autocomplete on every input, the manual reorganised into scannable cards, and an honest fix to the long-context verdict (the d_horizon ≡ T_eval tautology, see docs/LIMITATIONS.md). Full notes: docs/RELEASE_NOTES_v0.10.md.

▶ Try it now

---

🆕 v0.9 — The "fits ≠ works" pack (3 new modes)

The dozen GGUF/VRAM calculators on HF answer one question: does it fit in my GPU? None answer the harder one: does it fit and still work? These three close that gap — all browser-only, no inference.

Mode	What it answers	Pain it kills
🧵 YaRN Planner	"How do I set rope_scaling to extend context — and will quality hold?"	The most-asked HF question; nobody gives the config and a γ/d_horizon verdict
🧊 GGUF Bridge	"This Q4 fits 8GB — but is it brain-dead past 30K?"	Reads the .gguf header (HTTP Range, no full download) → quality verdict + per-quant comparison the VRAM calcs skip
🚀 Launch Flags	"What -ngl for Llama-70B-Q4 on my 4090?"	Exact llama.cpp / Ollama command (-ngl, -c, --no-mmap, KV-type) + warns when your context is past the usable horizon

▶ Try it now

---

🆕 v0.7 — Anti-bullshit pack (7 new modes)

Each mode solves a documented HuggingFace community pain. All run in your browser with zero inference — pure metadata + math. 3 of these have no comparable tool publicly available (validated against existing literature):

Mode	What it answers	Pain it kills
🪟 Unmask	"Is max_position_embeddings honest?"	Mistral-7B-v0.1 says 32k, attends ~8k via SWA
📜 Chat-template	"Which CLI flag for lm-eval / vLLM / transformers?"	lm-eval-harness #1841 silently halves multi-turn accuracy
🎯 Arena CI	"Is GPT-4 actually better than Claude — or tied?"	Chatbot Arena strips CIs from public leaderboard
🧪 Contamination	"Should I trust this MMLU score?"	Open LLM Leaderboard v1 killed by MMLU/HellaSwag contamination
⚖️ Quant	"Will NF4 break my model? AWQ better?"	Generic "AWQ ~95%" claims hide model-specific cliffs
🔀 Drift	"lm-eval gives 67.2, vLLM gives 65.1. Bug or noise?"	arxiv 2506.09501 — eval irreproducibility
🔍 NIAH→Reason	"Does my 128k-context model actually reason there?"	RULER paper finding — needle pass ≠ multi-hop reasoning

Plus a search-as-you-type HF Hub autocomplete on every model-id input (5 minute cache, privacy-noted, gated-license link).

▶ Try it now · v0.7 release notes

---

A note before you read on

This tool was built by one independent researcher, with no funding, no team, no GPUs beyond a single consumer card, and the full collaborative help of large language models as research instruments. It exists because the paper it complements (Predicting How Transformers Attend — Marin 2026) needed a way for any reader to check the framework's predictions on their own model in seconds, without installing anything, without paying anyone, and without trusting a server they don't control.

If it is useful to you — even once — that is enough. If it is wrong about your model, please tell us so we can fix the framework. The point is the common ground, not the artefact.

The closed-form predictions used in TAF Agent have explicit assumptions and a regime of validity. When the in-tool Validity Gate fires (γ panel banner), or when you want the full discussion of what the closed-forms assume and when they break, see docs/LIMITATIONS.md.

---

What it does

Drop in a model id (or paste any HuggingFace public model), get a falsifiable answer to "will this work?" — backed by the Thermodynamic Attention Framework (TAF) formulas:

Decision recipes

• Will Llama-3-8B serve 32K context with NIAH retrieval? → X-2
• Should I train a custom 7B model or pay for API access? → X-1
• I have $5,000 — what model can I afford to train? → X-3
• Cheapest GPU to serve Llama-70B at 100M tokens/day? → X-5
• Soft KV decay or hard cutoff for compression? → X-19

Diagnostic recipes (NEW v0.4 — sesión 29 findings 2026-04-28)

• How much positional bias did training imprint on this model? → X-21 (✗ retired — ν is not identified by the data; measure γ directly instead)
• Does this model fit the empirical compute-context invariant band? → X-22
• Is this checkpoint pre- or post-induction-head? → X-23

Each as a chain of TAF formulas (paper §17, §19, §20, §24, §26, §28-§30) rendered with full audit trail. Every number is deterministic Python; nothing is hallucinated.

Four ways to use it

• 📇 Profile a model — paste id, get all 5 recipes scored as a unified TAF Card (best starting point)
• 🆚 Compare models — 2-3 candidates side-by-side on the same recipe
• 💬 Ask plain English — free-form question, in-browser LLM picks the right recipe
• 📋 Pick recipe — manual selection with full form control

How it stays free + unlimited

• Static HTML/JS hosted on GitHub Pages (truly unlimited bandwidth)
• Python TAF computation runs in your browser via Pyodide (no server-side compute)
• Plain-English synthesis runs Qwen2.5-0.5B-Instruct in your browser via WebLLM (your GPU/CPU, your electricity, ~350MB cached after first load)
• Model config.json files fetched directly from HuggingFace Hub (free, public, no auth for non-gated models)
• Your data never leaves your browser

If 1 user or 1 million users hit it, our cost stays the same: $0.

Why static HTML+JS+Pyodide instead of Gradio/Streamlit?

A reasonable question. Three TAF Agent USPs are only possible with browser-only architecture:

1. Your inputs never leave the tab. No server = no privacy compromise. The "anti-bullshit" framing depends on this.
2. $0 forever, even at infinite scale. Static Spaces have unlimited HF bandwidth; there is no cold-start, no queue, no rate limit. Going viral can't bankrupt the project.
3. Lean+Mathlib formal verification ships as a static manifest. The 37 theorem badges link to source lines that anyone can lake build themselves — no hidden server logic.

Bonus: in-browser LLM (WebLLM running Qwen2.5-0.5B in your GPU/CPU) for the 💬 Ask mode is only viable in static. Pyodide running deterministic Python in your browser means you can audit every number — no opaque server.

The cost: HuggingFace's "Trending Spaces" algorithm favours Gradio/Streamlit Spaces. We compensate with detailed tags + forum presence + this README. If you'd prefer a Python-API client, that's a planned gradio_client companion (v0.9).

---

Architecture coverage

Supports any model whose config.json is parseable:

Family	Examples	Status
RoPE-MHA	pythia, gpt-j, original LLaMA	✓ supported
RoPE-GQA	Llama-3, Mistral, Qwen2.5, gemma-2	✓ supported
ALiBi	BLOOM, Falcon	✓ supported
AbsPE	gpt2 family	✓ supported
SWA (sliding window)	Mistral, gemma-2, phi-3	✓ supported
SSM	Mamba, Mamba-2	✓ partial (γ doesn't apply, KV does)
Any HF Hub public model	(any)	✓ via 📥 Fetch button

Languages

Interface available in:

• 🇬🇧 English
• 🇪🇸 Español
• 🇫🇷 Français
• 🇨🇳 中文

Click flags top-right to switch.

Local development

Browser application

git clone https://github.com/karlesmarin/tafagent
cd tafagent
python -m http.server 8000
# open http://localhost:8000

CLI diagnostic (for the paper)

The directory cli/diagnose_model.py is the command-line companion described in the paper Predicting How Transformers Attend (Marin 2026). It characterises any causal language model from HuggingFace in minutes on CPU and produces the raw gamma_obs, R², and thermodynamic profile used in the manuscript.

pip install torch transformers numpy
python cli/diagnose_model.py --model EleutherAI/pythia-2.8b --fast --cpu

Reproducibility data

The directory data/ ships every measurement referenced in the paper (343 JSON files, ~5.5 MB). See data/README.md for the layout.

Browser requirements

• Chrome / Edge / Firefox 113+ for WebGPU acceleration (recommended)
• Older browsers fall back to CPU inference (slower but works)
• ~2 GB free RAM for the synthesis LLM
• ~350 MB disk for model cache (one-time)

What's new in v0.4 (2026-04-28)

Three new diagnostic recipes derived from cross-model panel analysis (n=22 LLMs):

X-21 — Imprint Purity Diagnostic · ✗ RETIRED (2026-08)

Predicted γ on RANDOM-token input via the learned-imprint formula:

γ_random = γ_pade(θ, T) + ν · log_10(P / 14M)
   ν = −1/(2π) ≈ −0.1592

✗ Not identified by the data. Part II says so upfront — "ν remains unconstrained by the current data": bootstrap 95% CI [−0.260, −0.008], only an 11.1% chance a resampled slope lands within 1% of −1/(2π), and pythia-31m drives much of the panel slope. Part II also withdrew the "0.3% match" framing by name. What does hold: the Lean algebra, and the AbsPE falsification control (GPT-2 family, no 2π period → ν ≈ 0, mean |ν| = 0.028 vs RoPE range [−0.5, +0.3]). So the slope is consistent with −1/(2π) but not strongly identified — which is too weak a footing for a shipped diagnostic, hence the retirement. The recipe stays callable so published runs remain reproducible; its verdict is prefixed ✗ RETIRED and must not be acted on. See docs/LIMITATIONS.md §2.

Instead: measure γ directly (Recipe 1 / cli/diagnose_model.py) rather than predicting it from ν.

X-22 — Compute-Context Invariant

Computes the empirical Chinchilla×attention invariant:

K = γ × log(N² · D)   where D = 20·N (Chinchilla compute-optimal)
Empirical band: K ∈ [34, 68]   (51.2 ± 16.8, CV=0.329, n=22)

K-outliers indicate scaling/training anomalies. Llama-3-8B with γ=1.045 gives K=74.6 (z=1.39, high-K OUTLIER) — flags supra-Padé attention.

X-23 — IH-Phase Detector

Uses the Δγ probe (cheaper than ICL benchmark):

sign(γ_text − γ_random) > 0   ⟺   post-induction-head formation

Pre-IH (P<400M, n=7): ⟨Δγ⟩=−0.19±0.26 Post-IH (P≥400M, n=15): ⟨Δγ⟩=+0.03±0.26

Use case: monitor training trajectories without running ICL benchmarks; detect anomalous checkpoints.

Other v0.4 additions

• gamma_decompose_v2(...) — 6-axis decomposition with the new imprint axis
• famous_constant_proximity(...) — detects γ-cluster on famous constants (e.g. CodeLlama-13b γ=0.382 ≈ 1−1/φ golden conjugate)

---

What's new in v0.5.3 (2026-05-02) — 🔧 Audit-driven bug fixes

The TAF Agent was applied to its own author's paper (recursive Sócrates audit) and to the agent's own formula implementations. Several real bugs were detected and corrected. All v0.5.0–v0.5.2 users running diagnostics on Phase B models (γ > 1: LLaMA-2/3, Mistral, Gemma, Qwen2.5-7B near-Hagedorn) received incorrect KV-compression recommendations. This release fixes all known issues.

Critical fixes

• D_f_closed (KV compression window): replaced asymptotic / Hagedorn-buffer branches with discrete cumulative sum. Old code clamped Phase B (γ>1) to N when truth was ~3 % of N (LLaMA-3-8B at γ=1.046 with N=2000 should compress to ~750 tokens; old code returned 2000). Boundary γ ∈ [0.99, 1.01] was off by factor ~2×. Now exact for any γ.

• partition_Z(γ=1, N): was log(N + 0.5), missing Euler-Mascheroni constant γ_E ≈ 0.577 (~7 % underestimate of H_N). Now log(N) + γ_E.

• free_energy_F: returned −log(Z) (β·F convention). Now −log(Z)/γ, consistent with the Helmholtz definition F = −T·log(Z) and the thermodynamic identity S = γ·(U − F).

• γ_pred: replaced obsolete C/lnθ heuristic with γ_Padé(θ, T_eval) (paper §3.3).

Calibration audit (cross-panel re-check, n=22)

Re-running the empirical δ corrections of gamma_decompose against the panel revealed:

Constant	Hardcoded	Panel re-audit	Verdict
δ_GQA	+0.11	+0.115	✓ replicates
δ_SWA	−0.21	originally fit on n=1 model	✗ disabled (insufficient data)
δ_post_IH	−0.15	group-mean ≈ 0 (n=16 yes / 6 no)	⚠ flagged exploratory
δ_instruct (v2)	−0.10	n=3, p=0.06 (already noted)	⚠ flagged exploratory

gamma_decompose and gamma_decompose_v2 now return per-axis status fields (delta_SWA_status, delta_post_IH_status, etc.) and a top-level calibration_warning so consumers can detect which corrections are reliable.

The TAF Card UI now displays a collapsible "v0.5.3 — Calibration audit" banner in all four supported languages (EN/ES/FR/ZH) explaining this.

Paper §5.2 erratum

The framework's own self-audit found that paper §5.2 Theorem 5.2 claims C_V(γ=1, N) = (log N)²/4. Sócrates triangulation (numerical Python + Sage exact rational + SymPy symbolic integral) confirms the correct asymptotic is (log N)²/12 — a factor-3 error in the paper's truncated Z-expansion proof. The agent's heat_capacity_Cv already computes the correct value via numerical derivative of U; only the paper's analytic formula is wrong, not the tool. A formal erratum will be published as a separate document.

Tests

22/22 unit tests pass (tests/test_taf_formulas.py), including regression tests for D_f Phase B, partition_Z γ_E, free_energy_F convention, and δ_SWA disabled.

Why this happened

These bugs survived prior reviews because the affected code paths were exercised mainly on Phase A models (γ < 0.95) where the asymptotic approximation is close enough. Phase B (γ > 1) and the boundary near Hagedorn (|γ−1| < 0.05) were under-tested. The agent now uses direct discrete computation, so accuracy is uniform across all γ.

---

What's new in v0.5 (2026-05-01) — 🔬 Machine-verified consistency

First transformer-attention framework with formal machine-proof backing.

Sage Groebner basis (algebraic decision procedure) + Lean Mathlib4 (dependent type theory) dual-tool verification of 15 algebraic identities of TAF critical exponents.

verify_algebraic_consistency(γ) — new function

Given measured γ ∈ Phase A (0,1), checks 12 D-SAGE identities derived from TAF exponents (β=γ−1, ν=1/(1−γ), η=γ−1, etc.):

• D-SAGE-1 (★★ core): 2η² + η·γ_χ + 1 = 0 (quadratic identity)
• D-SAGE-2: β·χ = −1 (Phase A)
• D-SAGE-4: α + χ = 2
• D-SAGE-5: α + γ_χ = 2(2 − γ)
• D-SAGE-6: β·γ_χ = −2γ² + 4γ − 3 (factored)
• Rushbrooke + Josephson tautologies (d=1)
• Fisher residual = γ(2γ−3)/(1−γ) (NOT zero generally; corrects "triple closure")
• η=2γ refutation (Phase A residual > 0; paper 1's claim was wrong)
• D-SAGE-7: c · |ν_imprint| · 2π = 3 (dimensional closure)

Pass = framework intact. Fail = bf16 outlier, quantization artifact, or γ measurement noise.

Paper 1 erratum

Paper 1 originally claimed η = 2γ. Sage Groebner + Lean Mathlib4 detected this is algebraically wrong (residual (−4γ³+5γ+1)/(1−γ) > 0 ∀γ ∈ Phase A). Correct value: η = γ − 1, satisfying D-SAGE-1.

Reproducibility

# Sage verification
docker run --rm -v "$(pwd)/analysis:/work" sagemath/sagemath:latest \
    sage /work/sage_recursive_sweep_2026-04-30.sage

# Lean verification
docker run --rm -v "$(pwd)/lean_taf:/work" \
    leanprovercommunity/lean:latest \
    -c "cd /work/taf && lake build"

Build success: 1973/1973 jobs (Mathlib4 + 15 TAF theorems), DONE_EXIT=0.

Lean code: lean_taf/taf/Taf/Identities.lean Sage script: analysis/sage_recursive_sweep_2026-04-30.sage

---

How you can help

This tool is at v0.11. There's a long way to go.

• 🐛 Report bugs: https://github.com/karlesmarin/tafagent/issues
• 🌐 Translate: add a language to js/i18n.js, send a PR
• 🧪 Falsify a prediction: run the tool on a model where you have ground-truth measurements; if our verdict disagrees with reality, open an issue. We take refutations as seriously as confirmations.
• ➕ New recipe: implement an X-N recipe in python/taf_browser.py following the pattern of X-1...X-19
• ➕ New preset: add a popular model to the PRESETS dict
• 📝 Improve docs / examples: anything that helps the next person

📚 The paper series (I → II → III)

TAF Agent is the practical, browser-side tool that closes the loop of a continuing paper series — Predicting How Transformers Attend (Marín 2026). Each part opens a new axis, and the modes here compute from all three:

• Part I — Analytic Power-Law Theory. The distance axis. Attention decays as a power law A(d) ∝ d^−γ; a closed-form predictor γ_Padé(θ,T) fixes γ from the RoPE geometry alone; the γ atlas places dozens of models on one axis, and the phase structure (Hagedorn point γ=1) separates looks far from concentrates. This is the foundation nearly every mode in the app reads from.
• Part II — A Six-Axis Decomposition. Splits the measured γ into six interpretable axes (RoPE geometry + training + architecture …), pins the learned imprint ν ≈ −1/2π, adds an NF4 precision-direction rule, and machine-verifies the algebraic backbone in Sage + Lean.
• Part III — From Attention to Residual Computation. Opens the orthogonal depth axis: three separately-measurable observables (transport / writing / commitment), an exact direct-logit-attribution identity, and the honest reframe that the J-lens "advantage band" is a property of the transport operator, not of stored content. This powers the new 🪜 Depth map mode.
• Formalization — A Machine-Checked Backbone. The load-bearing identities across I–III, verified in Lean 4 / Mathlib (52 theorems, zero sorrys).

Citation

If this tool helps you — paper or code:

@misc{marin2026taf1,
  author = {Mar\'in, Carles},
  title  = {Predicting How Transformers Attend: Analytic Power-Law Theory, Phase Transitions, and Practical Compression Tools},
  year   = {2026},
  note   = {Part I},
  url    = {https://doi.org/10.5281/zenodo.19826342},
  doi    = {10.5281/zenodo.19826342},
}

@misc{marin2026taf2,
  author = {Mar\'in, Carles},
  title  = {Predicting How Transformers Attend, Part II: A Six-Axis Decomposition with the Learned Imprint},
  year   = {2026},
  url    = {https://doi.org/10.5281/zenodo.19960572},
  doi    = {10.5281/zenodo.19960572},
}

@misc{marin2026taf3,
  author = {Mar\'in, Carles},
  title  = {Predicting How Transformers Attend, Part III: From Attention to Residual Computation},
  year   = {2026},
  url    = {https://doi.org/10.5281/zenodo.21288263},
  doi    = {10.5281/zenodo.21288263},
}

@misc{marin2026tafagent,
  author = {Marin, Carles},
  title  = {{TAF Agent}: Browser-Based Transformer Diagnostic Tool},
  year   = {2026},
  url    = {https://karlesmarin.github.io/tafagent},
}

License

Apache-2.0 (this code).

Synthesis model: Qwen2.5-0.5B-Instruct distributed under Apache-2.0.

---

Acknowledgements

This tool would not exist without:

• The model commons: EleutherAI, Meta AI, Alibaba Qwen team, Mistral AI, Google DeepMind, Microsoft Research, AI2, BigScience, TII, DeepSeek-AI, HuggingFace SmolLM team, the Mamba authors, the RWKV community, and OpenAI for releasing weights and configs publicly.
• The infrastructure commons: Pyodide, WebLLM, HuggingFace Hub, GitHub Pages, jsdelivr CDN.
• The maintainers of transformers, numpy, scipy, sympy, tokenizers, accelerate, and the dozens of small libraries that make modern ML possible.
• The wider ML community — bloggers, reproducibility checkers, Discord moderators, Stack Overflow answerers, blog post writers (Lilian Weng, Andrej Karpathy, Sebastian Raschka, Jay Alammar, Sasha Rush, Phil Wang, the EleutherAI team, and many more) whose explanations carried the author through every concept this tool uses.
• Large language models as research instruments — Claude (Anthropic), GPT (OpenAI), Gemini (Google DeepMind), Mistral, Llama, DeepSeek, Grok, Qwen-Chat, and Microsoft phi — for the symbolic derivations, sage cross-checks, prose revision, audit work, and long-form co-writing that underlie both this tool and the underlying paper.

The author was the hand that typed; the work itself belongs to the commons that made it possible.