taf-agent / python /taf_browser.py
karlexmarin's picture
fix: demote δ_SWA to disabled; flag δ_post_IH and δ_instruct exploratory
c76c38e
"""
TAF Browser — Pyodide-compatible TAF formulas + recipes.
Pure-Python deterministic computations of TAF (Thermodynamic Attention Framework)
formulas, plus 5 cross-section recipes for the most common viability questions.
Author: Carles Marin <transformerkmarin@gmail.com>
License: Apache-2.0
"""
from __future__ import annotations
import math
import json
# ════════════════════════════════════════════════════════════════════════════
# §26 — γ-Thermodynamics (OUR contribution)
# ════════════════════════════════════════════════════════════════════════════
def gamma_pade(theta: float, T_eval: int) -> float:
"""§26.1 — γ = (2θ - T√2)/(2θ + T√2)"""
z_sqrt2 = T_eval * math.sqrt(2)
return (2 * theta - z_sqrt2) / (2 * theta + z_sqrt2)
def gamma_decompose(gamma_pade_val, has_GQA=False, has_SWA=False, n_params=0.0) -> dict:
"""§26.10 — 5-axis decomposition (n=23 OLS, paper sesión 28).
Calibration audit (2026-05-02 panel re-check):
δ_GQA = +0.11 ✓ replicates (group-mean +0.115 on n=9/13)
δ_SWA = -0.21 ⚠ ORIGINALLY FIT ON n=1; demoted to None here
(insufficient data; group-mean +0.355 with single yes-case).
Returning δ_SWA = 0 with `delta_SWA_status: 'exploratory_n1'`
instead of applying an unreliable correction.
δ_post_IH = -0.15 ⚠ does NOT replicate (group-mean ≈ 0 on n=16/6);
kept but flagged 'exploratory'.
δ_instruct not in this v1; v2 has -0.10 with n=3, p=0.06 caveat.
"""
delta_GQA = +0.11 if has_GQA else 0.0
# SWA: demoted — original constant rested on n=1.
delta_SWA = 0.0
delta_SWA_status = "exploratory_n1_disabled" if has_SWA else "not_applicable"
delta_post_IH = -0.15 if n_params >= 4e8 else 0.0
return {
"pade_centroid": gamma_pade_val,
"delta_GQA": delta_GQA,
"delta_SWA": delta_SWA,
"delta_SWA_status": delta_SWA_status,
"delta_post_IH": delta_post_IH,
"delta_post_IH_status": "exploratory_no_replication" if delta_post_IH != 0 else "not_applicable",
"gamma_corrected": gamma_pade_val + delta_GQA + delta_SWA + delta_post_IH,
"calibration_warning": ("SWA correction disabled (originally fit on n=1). "
"post_IH correction marked exploratory (group-mean ≈ 0 in re-audit). "
"GQA correction replicates."),
}
def d_horizon(theta: float, gamma: float):
"""§26.2 — d_h = θ(1-γ)√2/(1+γ). None if γ outside (0,1)."""
if gamma <= 0 or gamma >= 1:
return None
return theta * (1 - gamma) * math.sqrt(2) / (1 + gamma)
def l_niah_c(d_horizon_val):
"""§26.5 — L_NIAH^c = 2·d_horizon."""
return None if d_horizon_val is None else 2 * d_horizon_val
def chi_susceptibility(gamma: float) -> float:
"""§26.16 — χ = 1/|γ-1|."""
return float('inf') if gamma == 1.0 else 1.0 / abs(gamma - 1.0)
def p_hallucinate(L: int, theta: float, gamma: float):
"""§26.9 — Horizon-overshoot probability."""
dh = d_horizon(theta, gamma)
if dh is None or L <= 0:
return None
chi = chi_susceptibility(gamma)
if chi == float('inf'):
return None
geom = max(0.0, 1.0 - (dh / L) ** (1 - gamma))
return geom * (math.sqrt(chi) / (1 + math.sqrt(chi)))
def theta_design(gamma_target: float, T_eval: int) -> float:
"""§26.3 — θ to land at γ_target at T_eval (Padé inverse)."""
if gamma_target >= 1 or gamma_target <= -1:
raise ValueError("gamma_target must be in (-1, 1)")
return T_eval * math.sqrt(2) * (1 + gamma_target) / (2 * (1 - gamma_target))
def alpha_opt(gamma_target: float, T_eval: int, theta_nominal: float) -> float:
"""§26.4 — α = θ_design / θ_nominal."""
return theta_design(gamma_target, T_eval) / theta_nominal
def df_window(gamma: float, N: int, f: float = 0.90):
"""§26.7 — KV compression window via DISCRETE cumulative sum.
Returns None outside calibrated zone γ ∈ [0.65, 0.85]. Inside, computes
the smallest D such that ∑_{d=1}^D d^{-γ} / ∑_{d=1}^N d^{-γ} ≥ f.
This is exact for the discrete attention distribution and avoids the
continuum-approximation error that the paper's closed form has at γ→1.
"""
if not (0.65 <= gamma <= 0.85):
return None
if N <= 0:
return 1
weights = [d ** (-gamma) for d in range(1, N + 1)]
total = sum(weights)
target = f * total
cum = 0.0
for d, w in enumerate(weights, start=1):
cum += w
if cum >= target:
return d
return N
def kv_soft_decay_regime(theta: float, gamma: float, T_train: int) -> str:
"""§26.8 — Soft decay régimen-bound. d_h ≳ T_train/2 ⇒ applies."""
dh = d_horizon(theta, gamma)
if dh is None:
return "use-hard-cutoff"
ratio = dh / max(1, T_train / 2)
if ratio >= 1.2:
return "applies"
if ratio >= 0.8:
return "borderline"
return "use-hard-cutoff"
# ════════════════════════════════════════════════════════════════════════════
# §28 — Sesión 29 (2026-04-28): learned-imprint, F2 Chinchilla, Δγ-IH probe
# ════════════════════════════════════════════════════════════════════════════
NU_IMPRINT = -1.0 / (2 * math.pi) # §28 — learned-imprint slope (DERIVED, n=22 err 0.3%)
P_0_IMPRINT_M = 14.0 # baseline pythia-14m (smallest panel reference)
def gamma_random_predict(theta: float, T_eval: int, n_params_M: float) -> float:
"""§28.1 — Predicted γ on RANDOM-token input.
γ_random = γ_pade(θ,T) + ν · log_10(P / P_0), ν = -1/(2π) ≈ -0.1592.
Empirical n=22 LLMs (sesión 29). Random-input γ scales with model size
despite RoPE-Padé predicting only (θ,T) dependence — weights imprint
a learned positional bias proportional to log(N_params).
Predicted CI ≈ ±0.18 (95%).
"""
g_pade = gamma_pade(theta, T_eval)
return g_pade + NU_IMPRINT * math.log10(max(n_params_M, 1e-3) / P_0_IMPRINT_M)
def imprint_purity(gamma_random_obs: float, theta: float, T_eval: int,
n_params_M: float) -> dict:
"""§28.2 — Diagnostic: how clean is the model's RoPE-Padé prediction?
Compares observed γ_random to predicted (γ_pade + ν·log_10(P/P_0)).
Negative residual ⇒ extra-strong training imprint (less clean).
Positive ⇒ weaker than expected imprint (cleaner / less trained).
"""
g_pred = gamma_random_predict(theta, T_eval, n_params_M)
g_pade_only = gamma_pade(theta, T_eval)
residual = gamma_random_obs - g_pred
return {
"gamma_random_obs": gamma_random_obs,
"gamma_random_pred": g_pred,
"gamma_pade_only": g_pade_only,
"imprint_predicted": g_pred - g_pade_only,
"imprint_residual": residual,
"purity": "clean (within CI)" if abs(residual) < 0.18 else
("over-imprinted" if residual < 0 else "under-imprinted"),
"ci_95_half_width": 0.18,
}
def compute_invariant_K(gamma: float, n_params_M: float,
D_tokens: float = None) -> dict:
"""§29 — F2 Chinchilla compute-context invariant.
K = γ × log(N²·D), D = 20·N (Chinchilla compute-optimal) if not given.
Empirical: K ≈ 51.2 ± 16.8 (CV=0.329, n=22). In-distribution if K∈[34, 68].
"""
N = n_params_M * 1e6
if D_tokens is None:
D_tokens = 20 * N
K = gamma * math.log(N * N * D_tokens)
panel_mean, panel_std = 51.2, 16.8
z = (K - panel_mean) / panel_std
return {
"K": K,
"panel_mean": panel_mean,
"panel_std": panel_std,
"z_score": z,
"in_distribution": abs(z) <= 1.0,
"interpretation": "in-band" if abs(z) <= 1.0 else
("high-K outlier" if z > 0 else "low-K outlier"),
}
def ih_phase_check(gamma_text: float, gamma_random: float,
n_params_M: float = None) -> dict:
"""§30 — IH-formation phase discriminator.
sign(γ_text − γ_random) > 0 ⟺ post-IH (text concentrates more than random).
Pre-IH (P<400M, n=7): ⟨Δγ⟩ = -0.19 ± 0.26
Post-IH (P≥400M, n=15): ⟨Δγ⟩ = +0.03 ± 0.26
"""
delta = gamma_text - gamma_random
phase_observed = "post-IH" if delta > 0 else ("pre-IH" if delta < 0 else "ambiguous")
phase_expected = None
if n_params_M is not None:
phase_expected = "post-IH" if n_params_M * 1e6 >= 4e8 else "pre-IH"
consistent = (phase_expected is None) or (phase_observed == phase_expected)
return {
"delta_gamma": delta,
"phase_observed": phase_observed,
"phase_expected_by_size": phase_expected,
"consistent": consistent,
"panel_pre_IH_mean": -0.19,
"panel_post_IH_mean": +0.03,
"panel_std": 0.26,
}
def gamma_decompose_v2(gamma_pade_val: float, n_params_M: float,
has_GQA: bool = False, has_SWA: bool = False,
corpus: str = "text", is_instruct: bool = False) -> dict:
"""§28.3 — 6-axis decomposition (sesión 29 update with imprint axis).
γ_obs = γ_pade
+ ν·log_10(P/P_0)·𝟙[corpus=random] ← NEW imprint axis (DERIVED, n=22, err 0.3%)
+ Δ_corpus(text-rand)
+ δ_arch(GQA, SWA)
+ δ_circuit(IH phase)
+ δ_train(steps, RLHF, instruct)
+ ε
Calibration audit 2026-05-02:
δ_GQA solid; δ_SWA demoted (n=1); δ_post_IH exploratory; δ_instruct exploratory (n=3).
"""
delta_imprint = NU_IMPRINT * math.log10(max(n_params_M, 1e-3) / P_0_IMPRINT_M) \
if corpus == "random" else 0.0
delta_GQA = +0.11 if has_GQA else 0.0
# SWA disabled: originally fit on n=1.
delta_SWA = 0.0
delta_post_IH = -0.15 if n_params_M >= 400 else 0.0
delta_instruct = -0.10 if is_instruct else 0.0 # F9 tentative (n=3, p=0.06)
return {
"pade_centroid": gamma_pade_val,
"delta_imprint": delta_imprint,
"delta_GQA": delta_GQA,
"delta_SWA": delta_SWA,
"delta_SWA_status": "exploratory_n1_disabled" if has_SWA else "not_applicable",
"delta_post_IH": delta_post_IH,
"delta_post_IH_status": "exploratory_no_replication" if delta_post_IH != 0 else "not_applicable",
"delta_instruct": delta_instruct,
"delta_instruct_status": "tentative_n3_p0.06" if delta_instruct != 0 else "not_applicable",
"gamma_corrected": gamma_pade_val + delta_imprint + delta_GQA
+ delta_SWA + delta_post_IH + delta_instruct,
"corpus": corpus,
"axes": ["pade", "imprint", "GQA", "SWA", "IH", "instruct"],
"calibration_warning": ("SWA disabled (n=1). post_IH/instruct marked exploratory. "
"GQA + imprint axes are the most reliable."),
}
def famous_constant_proximity(gamma: float, tolerance: float = 0.01) -> dict:
"""§31 — Detect proximity to famous constants in γ-cluster (sesión 29).
Empirical hits (n=4 in panel):
CodeLlama-13b γ=0.3823 ≈ 1−1/φ = 0.3820 (golden conjugate)
pythia-1.4b γ=0.7051 ≈ 1/√2 = 0.7071
Llama-2-7b γ=0.2871 ≈ 1−1/√2 = 0.2929
Mistral-Nemo γ=0.4284 ≈ log_10(e) = 0.4343
Returns nearest constant within tolerance, or None.
"""
phi = (1 + math.sqrt(5)) / 2
constants = {
"1−1/φ (golden conjugate)": 1 - 1/phi,
"1/√2": 1 / math.sqrt(2),
"1−1/√2": 1 - 1/math.sqrt(2),
"log_10(e)": math.log10(math.e),
"1/π": 1 / math.pi,
"2/π": 2 / math.pi,
"1/φ": 1 / phi,
"ln(2)": math.log(2),
"z*_Cayley = (√17−3)/2": (math.sqrt(17) - 3) / 2,
}
hits = []
for name, val in constants.items():
err = abs(gamma - val)
if err <= tolerance:
hits.append({"constant": name, "value": val, "error": err})
hits.sort(key=lambda h: h["error"])
return {
"gamma": gamma,
"tolerance": tolerance,
"n_hits": len(hits),
"hits": hits[:3],
"caveat": "n=4 hits in panel; could be coincidence (continuous distribution)",
}
# ════════════════════════════════════════════════════════════════════════════
# §17 — Pre-training viability formulas
# ════════════════════════════════════════════════════════════════════════════
def chinchilla_optimal_tokens(N_params: float, ratio: float = 20.0) -> float:
"""§17.30 — Chinchilla 20:1 token budget. D = ratio · N."""
return ratio * N_params
def chinchilla_optimal_N(D_tokens: float, ratio: float = 20.0) -> float:
"""§17.30 inverse — given D tokens, optimal N = D/20."""
return D_tokens / ratio
def training_flops(N_params: float, D_tokens: float) -> float:
"""§17.10 — C ≈ 6·N·D total training FLOPs."""
return 6 * N_params * D_tokens
def training_memory_16N(N_params: float) -> dict:
"""§17.20 — total memory ≈ 16·N bytes (model + grads + Adam moments)."""
bytes_total = 16 * N_params
return {
"bytes": bytes_total,
"GB": bytes_total / 1e9,
}
def emergent_threshold(N_params: float) -> str:
"""§17.60 — capability threshold heuristic (Wei 2022)."""
if N_params >= 1e11:
return "above 100B — strong reasoning capabilities expected"
if N_params >= 1e10:
return "above 10B — most emergent capabilities present"
if N_params >= 1e9:
return "above 1B — basic instruction-following, not strong reasoning"
if N_params >= 1e8:
return "above 100M — useful for narrow tasks, no emergence"
return "below 100M — domain-specific tasks only"
# ════════════════════════════════════════════════════════════════════════════
# §19 — Inference economics
# ════════════════════════════════════════════════════════════════════════════
def kv_cache_memory(n_layers, n_kv_heads, d_head, seq_len, bytes_per_element=2.0) -> dict:
"""§19.1 — bytes = 2·L·n_kv·d_h·seq·B."""
bytes_total = 2 * n_layers * n_kv_heads * d_head * seq_len * bytes_per_element
return {"bytes": bytes_total, "MB": bytes_total / 1e6, "GB": bytes_total / 1e9}
def model_weights_memory(N_params, bytes_per_element=2.0) -> dict:
"""Inference memory for model weights only (BF16=2, INT8=1, INT4=0.5)."""
return {"GB": N_params * bytes_per_element / 1e9}
def inference_decode_throughput(N_params, hbm_GB_per_s, bytes_per_element=2.0) -> float:
"""§19.7 — memory-bound decode: tokens/sec = HBM_BW / model_size."""
model_GB = N_params * bytes_per_element / 1e9
return hbm_GB_per_s / model_GB
# ════════════════════════════════════════════════════════════════════════════
# §20 — Hardware catalog (curated from vendor docs 2026)
# ════════════════════════════════════════════════════════════════════════════
GPU_CATALOG = {
# name: {bf16_TFLOPs, hbm_GB, hbm_GB_s, cloud_USD_per_h_spot, tdp_W}
"H100 SXM": {"flops": 989, "vram_GB": 80, "bw_GB_s": 3350, "usd_h": 2.5, "tdp": 700},
"H100 PCIe": {"flops": 756, "vram_GB": 80, "bw_GB_s": 2000, "usd_h": 2.0, "tdp": 350},
"H200": {"flops": 989, "vram_GB": 141, "bw_GB_s": 4800, "usd_h": 3.5, "tdp": 700},
"B200": {"flops": 2250, "vram_GB": 192, "bw_GB_s": 8000, "usd_h": 5.0, "tdp": 1000},
"A100 80GB": {"flops": 312, "vram_GB": 80, "bw_GB_s": 2000, "usd_h": 1.2, "tdp": 400},
"A100 40GB": {"flops": 312, "vram_GB": 40, "bw_GB_s": 1555, "usd_h": 1.0, "tdp": 400},
"L40S": {"flops": 362, "vram_GB": 48, "bw_GB_s": 864, "usd_h": 0.7, "tdp": 350},
"MI300X": {"flops": 1307, "vram_GB": 192, "bw_GB_s": 5300, "usd_h": 2.1, "tdp": 750},
"RTX 4090": {"flops": 165, "vram_GB": 24, "bw_GB_s": 1008, "usd_h": 0.4, "tdp": 450},
"RTX 5090": {"flops": 419, "vram_GB": 32, "bw_GB_s": 1792, "usd_h": 0.7, "tdp": 575},
"RTX 5060Ti":{"flops": 36, "vram_GB": 16, "bw_GB_s": 448, "usd_h": 0.0, "tdp": 180}, # local
}
def cost_per_training_run(N_params: float, D_tokens: float, gpu: str = "H100 SXM",
n_gpus: int = 8, mfu: float = 0.45) -> dict:
"""§20.11 — cost = (flops_total / (peak·MFU·n_gpus)) · USD/h."""
info = GPU_CATALOG.get(gpu)
if info is None:
return {"error": f"unknown gpu '{gpu}'", "available": list(GPU_CATALOG.keys())}
total_flops = training_flops(N_params, D_tokens) # absolute FLOPs
effective_flops_per_sec = info["flops"] * 1e12 * mfu * n_gpus
seconds = total_flops / effective_flops_per_sec
hours = seconds / 3600
usd = hours * info["usd_h"] * n_gpus
return {
"total_FLOPs": total_flops,
"hours": hours,
"days": hours / 24,
"USD": usd,
"gpu": gpu, "n_gpus": n_gpus, "mfu": mfu,
}
def cost_per_inference_token(model_GB: float, gpu: str, batch: int = 1) -> dict:
"""§19.9 / §20.12 — derived $/Mtok from memory-bound decode."""
info = GPU_CATALOG.get(gpu)
if info is None:
return {"error": f"unknown gpu '{gpu}'"}
tok_per_sec = info["bw_GB_s"] / model_GB * batch
sec_per_Mtok = 1e6 / tok_per_sec
h_per_Mtok = sec_per_Mtok / 3600
usd_per_Mtok = h_per_Mtok * info["usd_h"]
return {
"tok_per_sec": tok_per_sec,
"USD_per_Mtok": usd_per_Mtok,
"gpu": gpu, "batch": batch,
}
# ════════════════════════════════════════════════════════════════════════════
# §24 — Cost / ROI
# ════════════════════════════════════════════════════════════════════════════
API_PRICING = {
# USD per million tokens (input/output blended typical)
"GPT-4o": {"input": 2.5, "output": 10.0},
"GPT-4o-mini": {"input": 0.15, "output": 0.60},
"Claude-Opus-4": {"input": 15.0, "output": 75.0},
"Claude-Sonnet-4":{"input": 3.0, "output": 15.0},
"Claude-Haiku-4": {"input": 0.80, "output": 4.0},
"Gemini-1.5-Pro": {"input": 1.25, "output": 5.0},
"DeepSeek-V3": {"input": 0.27, "output": 1.10},
"Llama-3.3-70B (Together)": {"input": 0.88, "output": 0.88},
}
def break_even_volume(training_cost: float, self_inference_per_Mtok: float,
api_per_Mtok: float, blend_input_output: float = 0.5) -> dict:
"""§24.3 — monthly tokens at which custom training pays off."""
savings_per_Mtok = api_per_Mtok - self_inference_per_Mtok
if savings_per_Mtok <= 0:
return {"error": "self-host more expensive than API per token; never breaks even"}
Mtok_breakeven = training_cost / savings_per_Mtok
return {
"savings_per_Mtok": savings_per_Mtok,
"Mtok_breakeven": Mtok_breakeven,
"tokens_breakeven": Mtok_breakeven * 1e6,
}
# ════════════════════════════════════════════════════════════════════════════
# RECIPES
# ════════════════════════════════════════════════════════════════════════════
# ─────────────────────────────────────────────────────────────────────
# X-2 — Long Context Viability
# ─────────────────────────────────────────────────────────────────────
def run_recipe_x2(theta, T_train, T_eval, n_attention_heads, n_kv_heads,
d_head, n_layers, n_params, has_SWA=False,
bytes_per_element=2.0, **_unused):
"""X-2: will model M serve length L doing NIAH retrieval?"""
chain = []
g_pade = gamma_pade(theta, T_eval)
chain.append(_step(1, "§26.1", "γ_Padé", "γ = (2θ - T√2)/(2θ + T√2)",
{"theta": theta, "T_eval": T_eval}, g_pade,
_phase_label(g_pade)))
has_GQA = (n_kv_heads < n_attention_heads)
decomp = gamma_decompose(g_pade, has_GQA=has_GQA, has_SWA=has_SWA, n_params=n_params)
g_corr = decomp["gamma_corrected"]
chain.append(_step(2, "§26.10", "γ-decomposition", "γ + δ_GQA + δ_SWA + δ_post_IH",
{"has_GQA": has_GQA, "has_SWA": has_SWA, "n_params": n_params},
g_corr, breakdown=decomp))
dh = d_horizon(theta, g_corr)
chain.append(_step(3, "§26.2", "d_horizon", "d_h = θ(1-γ)√2/(1+γ)",
{"theta": theta, "gamma": g_corr}, dh,
"n/a — γ outside (0,1)" if dh is None else f"horizon at d={dh:.0f}"))
l_niah = l_niah_c(dh)
chain.append(_step(4, "§26.5", "L_NIAH^c", "L_NIAH^c = 2·d_horizon",
{"d_horizon": dh}, l_niah,
"n/a" if l_niah is None else f"NIAH 50% at L={l_niah:.0f}"))
p_hallu = p_hallucinate(T_eval, theta, g_corr)
chain.append(_step(5, "§26.9", "P_hallucinate", "max(0,1-(d_h/L)^(1-γ))·√χ/(1+√χ)",
{"L": T_eval, "theta": theta, "gamma": g_corr}, p_hallu,
"n/a (Phase B)" if p_hallu is None else f"{p_hallu*100:.1f}% predicted"))
kv = kv_cache_memory(n_layers, n_kv_heads, d_head, T_eval, bytes_per_element)
chain.append(_step(6, "§19.1", "KV cache memory", "2·L·n_kv·d_h·seq·B",
{"n_layers": n_layers, "n_kv_heads": n_kv_heads, "d_head": d_head,
"seq_len": T_eval, "bytes_per_element": bytes_per_element},
kv, f"{kv['GB']:.2f} GB per request"))
if g_corr <= 0 or g_corr >= 1:
verdict, reason = "NO", "Phase B / geometric collapse (γ_corrected outside (0,1))"
mit = (f"Apply NTK-aware extension. Required θ for γ=0.85: "
f"{theta_design(0.85, T_eval):,.0f}. α_opt = {alpha_opt(0.85, T_eval, theta):.2f} "
f"({'fine-tuning required' if alpha_opt(0.85, T_eval, theta) > 8 else 'zero-shot may work'}).")
elif dh is not None and T_eval < dh:
margin = (1 - T_eval / dh) * 100
verdict, reason = "YES", f"L={T_eval} inside d_horizon={dh:.0f} ({margin:.0f}% margin)."
mit = "None required."
elif dh is not None and T_eval < l_niah:
verdict, reason = "DEGRADED", f"L between d_horizon ({dh:.0f}) and L_NIAH^c ({l_niah:.0f})."
mit = "Consider context contraction OR NTK extension."
else:
verdict, reason = "NO", f"L={T_eval} exceeds NIAH ceiling {l_niah:.0f}."
mit = f"Apply NTK extension; need θ ≈ {theta_design(0.85, T_eval):,.0f} for γ=0.85."
return _wrap("X-2", "Long Context Viability", locals(), chain, verdict, reason, mit)
# ─────────────────────────────────────────────────────────────────────
# X-1 — Custom training vs API for a domain task
# ─────────────────────────────────────────────────────────────────────
def run_recipe_x1(N_params, D_tokens=None, gpu="H100 SXM", n_gpus=8, mfu=0.45,
api_model="GPT-4o", monthly_tokens_M=10.0, **_unused):
"""X-1: custom training (Chinchilla optimal) vs API."""
chain = []
# Step 1: Chinchilla optimal D
if D_tokens is None:
D_tokens = chinchilla_optimal_tokens(N_params)
chain.append(_step(1, "§17.30", "Chinchilla optimal D", "D = 20·N",
{"N_params": N_params}, D_tokens,
f"recommended D = {D_tokens:.2e} tokens"))
# Step 2: training FLOPs
flops = training_flops(N_params, D_tokens)
chain.append(_step(2, "§17.10", "Training FLOPs", "C = 6·N·D",
{"N": N_params, "D": D_tokens}, flops,
f"{flops:.2e} FLOPs total"))
# Step 3: training cost
cost = cost_per_training_run(N_params, D_tokens, gpu=gpu, n_gpus=n_gpus, mfu=mfu)
chain.append(_step(3, "§20.11", "Training cost",
"hours·USD/h·n_gpus = total $",
{"gpu": gpu, "n_gpus": n_gpus, "mfu": mfu}, cost,
f"${cost['USD']:,.0f} over {cost['days']:.1f} days"))
# Step 4: model_GB and decode throughput
model_GB = N_params * 2 / 1e9 # BF16
inf = cost_per_inference_token(model_GB, gpu, batch=1)
chain.append(_step(4, "§19.9 / §20.12", "Self-inference $/Mtok",
"BW / model_GB → tok/s → $/Mtok",
{"model_GB": model_GB, "gpu": gpu}, inf,
f"${inf['USD_per_Mtok']:.2f} per million tokens (single user)"))
# Step 5: API blended price
api = API_PRICING.get(api_model, {"input": 2.0, "output": 8.0})
api_blend = (api["input"] + api["output"]) / 2
chain.append(_step(5, "§24.X", f"{api_model} blended price",
"(input + output) / 2 USD/Mtok",
{"api_model": api_model}, api_blend,
f"${api_blend:.2f}/Mtok blended"))
# Step 6: break-even
be = break_even_volume(cost["USD"], inf["USD_per_Mtok"], api_blend)
chain.append(_step(6, "§24.3", "Break-even tokens", "training$ / (api - self) = Mtok",
{"training_cost": cost["USD"]}, be,
_be_interp(be, monthly_tokens_M)))
# Verdict
if "error" in be:
verdict, reason = "NO", be["error"]
mit = f"Stick with {api_model} API."
elif monthly_tokens_M >= be["Mtok_breakeven"]:
verdict = "YES (custom)"
months_to_payoff = be["Mtok_breakeven"] / monthly_tokens_M
reason = (f"At {monthly_tokens_M} M tokens/month, break-even in "
f"{months_to_payoff:.1f} months. Long-term custom is cheaper.")
mit = f"Train at {gpu}×{n_gpus}; serve self-hosted."
else:
months = be["Mtok_breakeven"] / monthly_tokens_M
verdict = "NO (API)"
reason = (f"At {monthly_tokens_M} M tokens/month, break-even in "
f"{months:.1f} months — too slow.")
mit = f"Use {api_model} API (cheaper for your volume)."
return _wrap("X-1", "Custom training vs API", locals(), chain, verdict, reason, mit)
def _be_interp(be, monthly):
if "error" in be:
return be["error"]
months = be["Mtok_breakeven"] / max(monthly, 0.001)
return f"break-even at {be['Mtok_breakeven']:.0f} Mtok ({months:.1f} months at {monthly} M/mo)"
# ─────────────────────────────────────────────────────────────────────
# X-3 — Pre-flight check on $5K training budget
# ─────────────────────────────────────────────────────────────────────
def run_recipe_x3(USD_budget=5000.0, gpu="H100 SXM", mfu=0.45, n_gpus=1, **_unused):
"""X-3: given $ budget, what model can I train?"""
chain = []
info = GPU_CATALOG[gpu]
# Step 1: GPU-hours we can afford
hours = USD_budget / (info["usd_h"] * n_gpus)
chain.append(_step(1, "§20.11", "Affordable GPU-hours", "USD / ($/h·n_gpus)",
{"USD": USD_budget, "gpu": gpu, "n_gpus": n_gpus}, hours,
f"{hours:.0f} GPU-hours total ({hours/24:.1f} days at full use)"))
# Step 2: max FLOPs
max_flops = info["flops"] * 1e12 * mfu * n_gpus * hours * 3600
chain.append(_step(2, "§17.10", "Max training FLOPs",
"peak·MFU·n_gpus·seconds",
{"peak_TFLOPs": info["flops"], "MFU": mfu}, max_flops,
f"{max_flops:.2e} effective FLOPs"))
# Step 3: Chinchilla-optimal N (with D=20N)
# 6·N·D = max_flops, D=20N → 120·N² = max_flops → N = sqrt(max_flops/120)
N_chinchilla = math.sqrt(max_flops / 120)
D_chinchilla = 20 * N_chinchilla
chain.append(_step(3, "§17.30", "Chinchilla-optimal N",
"N = √(C/120) at D=20N", {"max_FLOPs": max_flops},
N_chinchilla,
f"N ≈ {N_chinchilla:.2e} params with D = {D_chinchilla:.2e} tokens"))
# Step 4: emergence check
emerg = emergent_threshold(N_chinchilla)
chain.append(_step(4, "§17.60", "Emergence threshold", "Wei 2022 capability",
{"N": N_chinchilla}, emerg, emerg))
# Step 5: memory budget check
mem = training_memory_16N(N_chinchilla)
fits = mem["GB"] <= info["vram_GB"]
chain.append(_step(5, "§17.20", "16N training memory",
"model + grads + AdamW",
{"N": N_chinchilla}, mem,
f"{mem['GB']:.1f} GB needed; "
f"{'fits in ' if fits else 'EXCEEDS '}{info['vram_GB']} GB VRAM"))
# Verdict
if N_chinchilla < 1e8:
verdict, reason = "TINY-MODEL", f"Budget supports only ~{N_chinchilla:.0e} params"
mit = "Use LoRA fine-tuning of larger pretrained model instead."
elif not fits:
verdict, reason = "MEMORY-LIMITED", f"Chinchilla N ({N_chinchilla:.1e}) doesn't fit one {gpu}"
mit = f"Use ZeRO-3 across multiple GPUs (need ≥{math.ceil(mem['GB']/info['vram_GB'])}× {gpu}) OR train smaller N undertrained."
else:
verdict = "GO"
reason = (f"At ${USD_budget}, train {N_chinchilla:.1e}-param model on "
f"{D_chinchilla:.1e} tokens in ~{hours/24:.1f} days. "
f"Capability tier: {emerg.split('—')[0].strip()}.")
mit = "None — proceed with Chinchilla-optimal recipe."
return _wrap("X-3", "Budget pre-flight", locals(), chain, verdict, reason, mit)
# ─────────────────────────────────────────────────────────────────────
# X-5 — Hardware selection for serving
# ─────────────────────────────────────────────────────────────────────
def run_recipe_x5(N_params, T_eval=4096, n_layers=32, n_kv_heads=8, d_head=128,
bytes_per_weight=2.0, target_tokens_per_day=10_000_000.0,
concurrent_users=1, **_unused):
"""X-5: which GPU should I use to serve N-param model at L context?"""
chain = []
# Step 1: weights memory
w_mem = model_weights_memory(N_params, bytes_per_weight)
chain.append(_step(1, "§19.X", "Model weights memory",
"N · bytes_per_weight",
{"N": N_params, "bytes": bytes_per_weight}, w_mem,
f"{w_mem['GB']:.1f} GB for weights"))
# Step 2: KV cache per request
kv = kv_cache_memory(n_layers, n_kv_heads, d_head, T_eval, bytes_per_weight)
chain.append(_step(2, "§19.1", "KV cache (per request)",
"2·L·n_kv·d_h·seq·B",
{"n_layers": n_layers, "n_kv": n_kv_heads,
"d_head": d_head, "seq": T_eval}, kv,
f"{kv['GB']:.2f} GB per concurrent request"))
# Step 3: total memory needed
total_GB = w_mem["GB"] + kv["GB"] * concurrent_users
chain.append(_step(3, "§20.3", "Total GPU memory",
"weights + KV·n_concurrent", {}, {"GB": total_GB},
f"{total_GB:.1f} GB for {concurrent_users} concurrent users"))
# Step 4: scan GPU catalog
candidates = []
for name, info in GPU_CATALOG.items():
if info["vram_GB"] < total_GB:
continue
# Decode throughput estimate (memory-bound)
tok_per_s = info["bw_GB_s"] / w_mem["GB"]
tok_per_day = tok_per_s * 86400
capacity_users = tok_per_day / target_tokens_per_day
usd_per_day = info["usd_h"] * 24
usd_per_Mtok = (usd_per_day / (tok_per_day / 1e6)) if tok_per_day > 0 else float('inf')
candidates.append({
"gpu": name, "vram_GB": info["vram_GB"], "bw_GB_s": info["bw_GB_s"],
"tok_per_sec": tok_per_s, "tok_per_day": tok_per_day,
"USD_per_day": usd_per_day, "USD_per_Mtok": usd_per_Mtok,
"users_supported": capacity_users,
})
candidates.sort(key=lambda c: c["USD_per_Mtok"])
chain.append(_step(4, "§20", f"Eligible GPUs (≥{total_GB:.0f}GB)",
"filter + rank by $/Mtok",
{"min_VRAM": total_GB}, candidates[:5],
f"{len(candidates)} GPUs fit; cheapest: {candidates[0]['gpu'] if candidates else 'NONE'}"))
# Verdict
if not candidates:
verdict, reason = "NO", f"No single GPU has ≥{total_GB:.0f} GB VRAM."
mit = (f"Use tensor parallelism across multiple GPUs "
f"(e.g. 2× H100 = 160GB), or quantize to INT8 (halves memory).")
else:
best = candidates[0]
verdict = "YES"
reason = (f"Best GPU: {best['gpu']} at ${best['USD_per_Mtok']:.2f}/Mtok. "
f"Supports {best['users_supported']:.1f}× your daily target.")
mit = f"Provision {best['gpu']}, expected {best['tok_per_sec']:.0f} tok/s decode."
return _wrap("X-5", "Hardware selection for serving", locals(), chain, verdict, reason, mit)
# ─────────────────────────────────────────────────────────────────────
# X-19 — KV compression decision (ours vs literature)
# ─────────────────────────────────────────────────────────────────────
def run_recipe_x19(theta, T_train, T_eval, n_attention_heads, n_kv_heads,
d_head, n_layers, n_params, has_SWA=False, **_unused):
"""X-19: should I use γ-soft KV decay, hard D_f, or literature methods?"""
chain = []
# Step 1: γ_Padé
g_pade = gamma_pade(theta, T_eval)
chain.append(_step(1, "§26.1", "γ_Padé", "(2θ-T√2)/(2θ+T√2)",
{"theta": theta, "T_eval": T_eval}, g_pade, _phase_label(g_pade)))
# Step 2: γ-decomposition
has_GQA = n_kv_heads < n_attention_heads
decomp = gamma_decompose(g_pade, has_GQA, has_SWA, n_params)
g_corr = decomp["gamma_corrected"]
chain.append(_step(2, "§26.10", "γ-decomposition", "5-axis adjustment",
{"has_GQA": has_GQA, "has_SWA": has_SWA, "n_params": n_params},
g_corr))
# Step 3: §26.7 D_f window applicability
df = df_window(g_corr, T_eval, f=0.90)
df_zone_ok = df is not None
chain.append(_step(3, "§26.7", "D_f window (γ in [0.65, 0.85])",
"[(1-f)+fN^(1-γ)]^(1/(1-γ))",
{"gamma": g_corr, "N": T_eval, "f": 0.9}, df,
f"D_f = {df}" if df_zone_ok
else f"NOT applicable (γ={g_corr:.3f} outside [0.65, 0.85])"))
# Step 4: §26.8 soft decay régimen
regime = kv_soft_decay_regime(theta, g_corr, T_train)
dh = d_horizon(theta, g_corr)
dh_str = f"{dh:.0f}" if dh is not None else "n/a"
chain.append(_step(4, "§26.8", "Soft decay régimen", "d_h ≳ T_train/2",
{"theta": theta, "gamma": g_corr, "T_train": T_train}, regime,
f"d_horizon={dh_str}; regime: {regime}"))
# Step 5: KV cache memory baseline
kv = kv_cache_memory(n_layers, n_kv_heads, d_head, T_eval)
chain.append(_step(5, "§19.1", "Baseline KV memory", "2·L·n_kv·d_h·seq·B",
{"L": n_layers, "n_kv": n_kv_heads, "d_h": d_head, "seq": T_eval},
kv, f"{kv['GB']:.2f} GB without compression"))
# Verdict
if regime == "applies" and df_zone_ok:
verdict = "USE SOFT DECAY"
reason = (f"d_horizon ≳ T_train/2 AND γ in compression zone. "
f"Soft decay (1-d/d_h)^γ best (-21% PPL vs hard cutoff per F17).")
mit = "Implement as 4D attention_mask additive bias with eager attention."
elif df_zone_ok:
verdict = "USE D_f HARD CUTOFF"
reason = f"γ in [0.65, 0.85] zone but d_h < T_train/2. Hard truncation at D_f={df} works."
mit = "Set cache_max_len = D_f."
elif regime == "applies":
verdict = "USE SOFT DECAY (caveat)"
reason = "Régimen applies but γ outside D_f validity zone. Soft decay only."
mit = "Soft decay; do not use D_f window."
elif g_corr >= 1 or g_corr <= 0:
verdict = "USE LITERATURE METHODS"
reason = f"γ={g_corr:.3f} outside Phase A. Our formulas don't apply."
mit = "Use SnapKV / PyramidKV / FastGen (literature heuristics)."
else:
verdict = "USE HARD T_train CUTOFF"
reason = "Régimen not met AND γ outside zone. Cap context at T_train."
mit = f"Set seq_len ≤ {T_train}, no extension."
return _wrap("X-19", "KV compression decision", locals(), chain, verdict, reason, mit)
# ─────────────────────────────────────────────────────────────────────
# X-21 — Imprint Purity Diagnostic (sesión 29 — uses §28 ν=−1/(2π))
# ─────────────────────────────────────────────────────────────────────
def run_recipe_x21(theta, T_train, n_attention_heads, n_kv_heads,
d_head, n_layers, n_params, T_eval=None,
gamma_random_obs=None, **_unused):
"""X-21: how clean is the model's RoPE-Padé prediction?
Predicts γ on RANDOM-token input via learned-imprint formula:
γ_random = γ_pade(θ,T) + ν·log_10(P/14M), ν = −1/(2π) ≈ −0.1592
If user provides observed γ_random, returns purity diagnostic.
"""
chain = []
if T_eval is None:
T_eval = T_train
# Step 1: γ_Padé baseline
g_pade = gamma_pade(theta, T_eval)
chain.append(_step(1, "§26.1", "γ_Padé", "(2θ-T√2)/(2θ+T√2)",
{"theta": theta, "T_eval": T_eval}, g_pade,
_phase_label(g_pade)))
# Step 2: predicted imprint shift
n_params_M = n_params / 1e6
imprint_shift = NU_IMPRINT * math.log10(max(n_params_M, 1e-3) / P_0_IMPRINT_M)
chain.append(_step(2, "§28.1", "Imprint shift", "ν·log_10(P/P_0), ν=−1/(2π)",
{"P_M": n_params_M, "P_0_M": P_0_IMPRINT_M, "nu": NU_IMPRINT},
imprint_shift,
f"Bigger model → stronger imprint (more negative shift)."))
# Step 3: predicted γ_random
g_pred = g_pade + imprint_shift
chain.append(_step(3, "§28.1", "γ_random predicted", "γ_pade + ν·log_10(P/P_0)",
{"gamma_pade": g_pade, "imprint": imprint_shift}, g_pred,
f"Predicted γ_random = {g_pred:.4f} ± 0.18 (95% CI)"))
# Step 4: purity diagnostic if observed value provided
if gamma_random_obs is not None:
purity = imprint_purity(gamma_random_obs, theta, T_eval, n_params_M)
chain.append(_step(4, "§28.2", "Imprint purity",
"obs − pred (purity = within ±0.18)",
{"gamma_random_obs": gamma_random_obs,
"gamma_random_pred": g_pred},
purity["imprint_residual"], purity["purity"]))
verdict = "CLEAN" if abs(purity["imprint_residual"]) < 0.18 else \
("OVER-IMPRINTED" if purity["imprint_residual"] < 0 else "UNDER-IMPRINTED")
reason = (f"Residual γ_random_obs − γ_pred = {purity['imprint_residual']:+.4f}. "
f"95% CI is ±0.18.")
mit = ("Models far from prediction may have anomalous training (e.g. heavy "
"fine-tuning, format conversion). Compare to native checkpoint.")
else:
verdict = "PREDICTION ONLY"
reason = (f"Predicted γ_random = {g_pred:.4f}. Provide gamma_random_obs to "
f"check purity (measure on RANDOM token sequences, e.g. via E4 protocol).")
mit = ("To measure: run a 150-prompt forward pass on RANDOM-token sequences "
"across distances d=10..1000 and fit power law. "
"(See https://github.com/karlesmarin/tafagent for E4 protocol.)")
return _wrap("X-21", "Imprint Purity Diagnostic", locals(), chain,
verdict, reason, mit)
# ─────────────────────────────────────────────────────────────────────
# X-22 — Compute-Context Invariant Check (sesión 29 — F2 Chinchilla)
# ─────────────────────────────────────────────────────────────────────
def run_recipe_x22(theta, T_train, n_params, gamma_obs, D_tokens=None,
T_eval=None, **_unused):
"""X-22: does the model lie in the empirical Chinchilla invariant band?
K = γ × log(N²·D), D = 20·N if not given.
Empirical: K ≈ 51.2 ± 16.8 (CV=0.329, n=22 panel).
"""
chain = []
if T_eval is None:
T_eval = T_train
n_params_M = n_params / 1e6
if D_tokens is None:
D_tokens = 20 * n_params # Chinchilla compute-optimal
# Step 1: K computation
inv = compute_invariant_K(gamma_obs, n_params_M, D_tokens)
chain.append(_step(1, "§29", "K = γ·log(N²·D)", "γ × ln(N²·D)",
{"gamma": gamma_obs, "N": n_params, "D": D_tokens},
inv["K"],
f"K = {inv['K']:.2f} (panel mean {inv['panel_mean']:.1f} ± "
f"{inv['panel_std']:.1f})"))
# Step 2: z-score interpretation
chain.append(_step(2, "§29", "z-score vs panel", "(K − μ)/σ",
{"K": inv["K"], "mean": inv["panel_mean"],
"std": inv["panel_std"]},
inv["z_score"],
inv["interpretation"]))
# Step 3: γ_pade comparison (anomaly test)
g_pade = gamma_pade(theta, T_eval)
pade_diff = gamma_obs - g_pade
chain.append(_step(3, "§26.1", "γ deviation from Padé", "γ_obs − γ_pade",
{"gamma_obs": gamma_obs, "gamma_pade": g_pade}, pade_diff,
"negative = anomaly (sub-Padé); positive = supra-Padé"))
if inv["in_distribution"]:
verdict = "IN-BAND"
reason = f"K = {inv['K']:.2f} within ±1σ of panel mean {inv['panel_mean']:.1f}."
mit = "Model conforms to compute-context invariant. No action needed."
else:
verdict = "OUTLIER"
reason = (f"K = {inv['K']:.2f} ({inv['interpretation']}). "
f"|z| = {abs(inv['z_score']):.2f} > 1.")
mit = ("High-K (over-concentrating attention for given compute) or low-K "
"(under-using compute for attention concentration). Check tokenizer, "
"training recipe, fine-tuning history.")
return _wrap("X-22", "Compute-Context Invariant", locals(), chain,
verdict, reason, mit)
# ─────────────────────────────────────────────────────────────────────
# X-23 — IH-Phase Detector (sesión 29 — F4 Δγ probe)
# ─────────────────────────────────────────────────────────────────────
def run_recipe_x23(n_params, gamma_text=None, gamma_random=None, **_unused):
"""X-23: is this checkpoint pre- or post-induction-head formation?
Discriminator: sign(γ_text − γ_random) > 0 ⟺ post-IH.
Cheaper than ICL benchmark for monitoring training trajectories.
"""
chain = []
n_params_M = n_params / 1e6
# Step 1: size-based prediction
expected = "post-IH" if n_params >= 4e8 else "pre-IH"
chain.append(_step(1, "§30", "Size-based phase prediction",
"P ≥ 400M ⇒ post-IH",
{"n_params_M": n_params_M, "threshold_M": 400}, expected))
# Step 2: γ-based discrimination if both gammas given
if gamma_text is not None and gamma_random is not None:
check = ih_phase_check(gamma_text, gamma_random, n_params_M)
chain.append(_step(2, "§30", "Δγ discriminator", "sign(γ_text − γ_random)",
{"gamma_text": gamma_text, "gamma_random": gamma_random},
check["delta_gamma"],
f"observed phase: {check['phase_observed']}"))
if check["consistent"]:
verdict = f"CONFIRMED {check['phase_observed'].upper()}"
reason = (f"Δγ = {check['delta_gamma']:+.3f} sign matches size-prediction "
f"({expected}).")
mit = "Phase confirmed. Use this checkpoint for downstream tasks accordingly."
else:
verdict = "ANOMALY"
reason = (f"Δγ = {check['delta_gamma']:+.3f} suggests {check['phase_observed']}, "
f"but size predicts {expected}. Investigate.")
mit = ("Possible causes: incomplete training, anomalous fine-tuning, "
"format conversion, tokenizer corruption (cf. F5 OLMo Δγ=0.30).")
else:
verdict = f"PREDICTED {expected.upper()}"
reason = (f"Only size given: P = {n_params_M:.0f}M. "
f"Provide gamma_text + gamma_random to verify via Δγ probe.")
mit = ("Run E4 protocol with corpus=mongo and corpus=random; "
"compare γ values.")
return _wrap("X-23", "IH-Phase Detector", locals(), chain,
verdict, reason, mit)
# ════════════════════════════════════════════════════════════════════════════
# Helpers
# ════════════════════════════════════════════════════════════════════════════
def _step(n, sec, name, formula, inputs, result, interpretation=None, breakdown=None):
s = {"step": n, "section": sec, "name": name, "formula": formula,
"inputs": inputs, "result": result}
if interpretation:
s["interpretation"] = interpretation
if breakdown:
s["breakdown"] = breakdown
return s
def _wrap(rid, rname, locals_dict, chain, verdict, reason, mitigation):
# Clean inputs (drop chain/internal vars)
inputs = {k: v for k, v in locals_dict.items()
if not k.startswith("_") and k not in
("chain", "verdict", "reason", "mit", "info", "be", "kv", "g_pade", "g_corr",
"decomp", "dh", "l_niah", "p_hallu", "cost", "model_GB", "inf", "api",
"api_blend", "fits", "mem", "emerg", "max_flops", "hours",
"N_chinchilla", "D_chinchilla", "candidates", "best", "tok_per_s",
"tok_per_day", "capacity_users", "usd_per_day", "usd_per_Mtok",
"total_GB", "w_mem", "df", "df_zone_ok", "regime", "has_GQA",
"margin", "months", "months_to_payoff", "name")}
return {"recipe_id": rid, "recipe_name": rname, "inputs": inputs,
"chain": chain, "verdict": verdict, "reason": reason,
"mitigation": mitigation}
def _phase_label(g):
if 0 < g < 1:
return "Phase A (long-range OK)"
if g >= 1:
return "Phase B / Hagedorn"
return "Phase B / catastrophic (negative γ — T too large for θ)"
# ════════════════════════════════════════════════════════════════════════════
# §32 — Sesión 29 visual-diagnostic helpers (paper 2 §4 bimodal + Hagedorn)
# ════════════════════════════════════════════════════════════════════════════
def hagedorn_safety_alert(gamma: float) -> dict:
"""§32.1 — Classify γ into safety zones (paper 2 §4.2 finding F12).
Empirical n=25 panel: 36% of LLMs operate at γ ≥ 0.95 (Hagedorn-zone risk).
Models at γ ≥ 1.0 cannot serve long context without NTK extension.
"""
if gamma is None or not isinstance(gamma, (int, float)):
return {"level": "unknown", "color": "gray", "message": "no γ available"}
if gamma >= 1.0:
return {
"level": "critical",
"color": "red",
"label": "🔴 HAGEDORN ZONE",
"message": ("γ ≥ 1.0: attention concentrates locally. Long-context "
"retrieval will FAIL without NTK extension (α_opt > 1)."),
"fraction_of_panel": "36% of n=25 LLMs are in this zone",
}
if gamma >= 0.95:
return {
"level": "warning",
"color": "orange",
"label": "🟠 HAGEDORN BOUNDARY",
"message": ("γ ≥ 0.95: at the edge. Long context above T_train risky. "
"Run X-2 long-context viability before deploying."),
"fraction_of_panel": "~24% of n=25 LLMs cluster here",
}
if gamma >= 0.65:
return {
"level": "ok",
"color": "green",
"label": "🟢 PHASE A NORMAL",
"message": ("γ ∈ [0.65, 0.95]: long-context OK. KV compression via "
"D_f window applicable in [0.65, 0.85]."),
"fraction_of_panel": "~40% of n=25 LLMs",
}
if gamma > 0:
return {
"level": "info",
"color": "blue",
"label": "🔵 PHASE A WIDE-FIELD",
"message": ("γ < 0.65: very long-range attention. Model may be over-"
"trained on long context, or pre-IH (small model)."),
"fraction_of_panel": "~12% of n=25 LLMs (mostly small or anomalous)",
}
return {
"level": "catastrophic",
"color": "darkred",
"label": "❌ NEGATIVE γ",
"message": "Negative γ means T_eval >> θ. Bad operating point. Use lower T or higher θ.",
"fraction_of_panel": "0% of n=25 (pathological)",
}
# ════════════════════════════════════════════════════════════════════════════
# §33 — Sesion 31 (2026-04-30) findings — added to TAF v0.4
# Architectural concentration law, PDI, 4-bit R²-direction rule, critical exponents
# ════════════════════════════════════════════════════════════════════════════
def architectural_concentration_predict(gamma_pade_val: float, n_kv: int) -> dict:
"""§33.1 — Architectural concentration law (paper 2 NEW, sesion 31).
γ_text ≈ γ_Padé − 0.012·n_kv
R² = 0.30 cross-panel (n=22) vs Padé alone R²=0.02.
IMPORTANT: This is a CORRELATIONAL law, NOT per-model predictor.
Mean per-model |err| = 0.27, WORSE than Padé alone (0.24).
Use as CROSS-PANEL diagnostic, not individual prediction.
"""
k_arch = 0.012 # panel-fit coefficient (n=22), not derived from first principles
gamma_predicted = gamma_pade_val - k_arch * n_kv
return {
"gamma_pade": gamma_pade_val,
"n_kv": n_kv,
"k_arch": k_arch,
"gamma_text_predicted": gamma_predicted,
"caveat": "Correlational, not per-model predictor (R²=0.30, mean err 0.27)",
"interpretation": (
"GQA aggressive (low n_kv) → γ pushed up toward Hagedorn. "
"MHA full (n_kv=32) → γ drops below Padé (sink-prone)."
),
}
def padé_deviation_index(theta: float, gamma_obs: float, T_eval: int) -> dict:
"""§33.2 — PDI Padé Deviation Index (paper 2 NEW, sesion 31).
PDI = d_horizon_obs / T_eval = θ(1−γ_obs)√2 / ((1+γ_obs)·T_eval)
Identity (D-NEW-1): PDI = 1 ⟺ γ_obs = γ_Padé(θ, T_eval)
Diagnostic value:
PDI ≈ 1: canonical (model matches Padé)
PDI > 1.5: γ_obs < γ_Padé (sink-dominated, code/instruct shift)
PDI < 0.5: γ_obs > γ_Padé but < 1 (over-concentrated)
PDI < 0: γ_obs > 1 (Phase B, formula sign-flips)
Equivalent log scale: log(PDI) + ΔH_Cardy = 0 (D-DEEP-15).
"""
if gamma_obs == -1.0:
return {"PDI": float('inf'), "regime": "singular"}
pdi = theta * (1 - gamma_obs) * math.sqrt(2) / ((1 + gamma_obs) * T_eval)
if pdi < 0:
regime = "Phase B (γ>1, formula sign-flip)"
traffic = "🔴 RED — Phase B, NTK extension required"
elif 0.5 <= pdi <= 1.5:
regime = "canonical (γ_obs ≈ γ_Padé)"
traffic = "🟢 GREEN — model matches Padé prediction"
elif pdi > 1.5:
regime = "γ_obs << γ_Padé (sink-dominated or extreme alignment)"
traffic = "🟠 ORANGE — large positive deviation"
else: # 0 < pdi < 0.5
regime = "γ_obs > γ_Padé (over-concentrated in Phase A)"
traffic = "🟡 YELLOW — moderate deviation"
return {
"PDI": pdi,
"log_PDI": math.log(pdi) if pdi > 0 else None,
"regime": regime,
"traffic_light": traffic,
"identity": "log(PDI) + ΔH_Cardy = 0 (use either, log-inverse)",
}
def precision_shift_predict_4bit(gamma_bf16: float, R2_bf16: float, is_GQA: bool) -> dict:
"""§33.3 — 4-bit precision shift direction predictor (paper 2 NEW, sesion 31).
Empirical n=5 rule: bf16 R² of power-law fit predicts 4-bit shift direction.
For MHA models:
R²(bf16) < 0.9 (sink-dominated): 4-bit shifts γ UP toward γ_Padé
R²(bf16) > 0.99 (clean): 4-bit shifts γ DOWN (introduces noise)
0.95 ≤ R² ≤ 0.99: stable (~no shift)
For GQA models: precision-robust regardless (|Δγ| < 0.05).
"""
if is_GQA:
return {
"predicted_shift_direction": "stable",
"expected_magnitude": "|Δγ| < 0.05",
"reason": "GQA KV-sharing distributes attention; little long-tail to perturb",
"recommendation": "Either bf16 or 4-bit OK for deployment",
}
# MHA case — depends on R²
if R2_bf16 < 0.9:
direction = "UP"
magnitude = "+0.3 to +0.8 expected"
reason = "Sink-dominated bf16: 4-bit truncates long-tail, reveals Padé prediction"
elif R2_bf16 > 0.99:
direction = "DOWN"
magnitude = "−0.2 to −0.4 expected"
reason = "Clean bf16: 4-bit further sparsifies, introduces non-monotonicity"
else:
direction = "stable"
magnitude = "|Δγ| < 0.05"
reason = "Borderline R², 4-bit minimal effect"
return {
"predicted_shift_direction": direction,
"expected_magnitude": magnitude,
"reason": reason,
"evidence": "n=5 paired measurements (DeepSeek/Pythia-1B/Pythia-2.8B/Llama-3/Qwen-7B-Inst)",
"caveat": "R²-direction rule is empirical, not formally derived",
}
def critical_exponents_bundle(gamma: float) -> dict:
"""§33.4 — Critical exponents bundle (paper 2 NEW, sesion 31 + GAME-O).
Returns ν_c (correlation length), β_c (order parameter), η_c (anomalous dim),
α_C (specific heat), γ_susc (susceptibility) as functions of γ.
Hyperscaling consistent (Rushbrooke + Josephson, d=1).
NEW IDENTITY (GAME-P recursive):
γ_susc(γ) = 1/(1−γ) + 2(1−γ) ≥ 2√2 (AM-GM bound)
Minimum at γ = 1 − 1/√2 ≈ 0.293
Equals c_central=3 at γ=0 AND γ=1/2
"""
if gamma >= 1:
return {"regime": "Hagedorn or beyond", "exponents": "diverge"}
nu_c = 1 / (1 - gamma)
beta_c = gamma - 1
eta_c = gamma - 1 # CORRECTED from paper 1's η=2γ (Lévy mapping consistent with hyperscaling)
alpha_C = 2 - 1 / (1 - gamma)
gamma_susc = 1 / (1 - gamma) + 2 * (1 - gamma)
# AM-GM bound check
gamma_min = 1 - 1 / math.sqrt(2)
gamma_susc_min = 2 * math.sqrt(2)
return {
"nu_correlation_length": nu_c,
"beta_order_param": beta_c,
"eta_anomalous_dim": eta_c,
"eta_note": "η_c = γ−1 (Lévy-derived, hyperscaling-consistent). Paper 1 claim η=2γ is INCORRECT.",
"alpha_specific_heat": alpha_C,
"gamma_susceptibility": gamma_susc,
"c_central_at_gamma_0": 3, # γ_susc(γ=0) = 3 = c_central
"AM_GM_bound": {
"min_gamma_susc": gamma_susc_min,
"min_at_gamma": gamma_min,
"interpretation": (
f"γ_susc has UNIVERSAL minimum {gamma_susc_min:.3f} at γ = {gamma_min:.4f} "
"(AM-GM with ab=2 product constant)"
),
},
"hyperscaling_check": {
"Rushbrooke (α + 2β + γ_susc = 2)": alpha_C + 2 * beta_c + gamma_susc,
"expected": 2,
},
"warning_paper1_eta": "Paper 1's η_c = 2γ is INCORRECT. Use η_c = γ-1 (this function).",
}
def verify_algebraic_consistency(gamma: float, tol: float = 1e-9) -> dict:
"""§34 v0.5 — Machine-verified framework consistency check.
Verifies the 12 D-SAGE algebraic identities discovered by Sage Groebner basis
and formally proven in Lean Mathlib4 (sesion 32, 2026-05-01).
Each identity is a sanity check: given measured γ, the TAF critical exponents
must satisfy these relations. Failures indicate γ measurement artifacts
(bf16 outliers, quantization issues, Phase B regime).
References:
- sage_recursive_sweep_results.json (Sage Groebner verification)
- lean_taf/taf/Taf/Identities.lean (Lean Mathlib4 machine-proof)
- paper 2 appendix A.4 "Formal verification"
"""
if gamma >= 1 or gamma <= 0:
return {
"status": "out_of_phase_A",
"phase_A_range": "(0, 1)",
"input_gamma": gamma,
"message": "Verification requires γ ∈ (0, 1) Phase A regime."
}
nu = 1 / (1 - gamma)
beta = gamma - 1
eta = gamma - 1 # CORRECTED η=γ-1 (NOT paper 1's 2γ)
alpha = 2 - 1 / (1 - gamma)
chi = 1 / (1 - gamma)
gamma_chi = 1 / (1 - gamma) + 2 * (1 - gamma)
checks = {
"D-SAGE-1": {
"claim": "2η² + η·γ_χ + 1 = 0",
"value": 2 * eta**2 + eta * gamma_chi + 1,
"expected": 0.0,
"passes": abs(2 * eta**2 + eta * gamma_chi + 1) < tol,
},
"D-SAGE-2": {
"claim": "β·χ = -1",
"value": beta * chi,
"expected": -1.0,
"passes": abs(beta * chi - (-1)) < tol,
},
"D-SAGE-4": {
"claim": "α + χ = 2",
"value": alpha + chi,
"expected": 2.0,
"passes": abs(alpha + chi - 2) < tol,
},
"D-SAGE-5": {
"claim": "α + γ_χ = 2(2-γ)",
"value": alpha + gamma_chi,
"expected": 2 * (2 - gamma),
"passes": abs(alpha + gamma_chi - 2 * (2 - gamma)) < tol,
},
"D-SAGE-6": {
"claim": "β·γ_χ = -2γ²+4γ-3",
"value": beta * gamma_chi,
"expected": -2 * gamma**2 + 4 * gamma - 3,
"passes": abs(beta * gamma_chi - (-2 * gamma**2 + 4 * gamma - 3)) < tol,
},
"Rushbrooke_tautology": {
"claim": "2β + γ_χ - ν·d = 0 (d=1)",
"value": 2 * beta + gamma_chi - nu,
"expected": 0.0,
"passes": abs(2 * beta + gamma_chi - nu) < tol,
},
"Josephson_tautology": {
"claim": "2 - α - ν·d = 0 (d=1)",
"value": 2 - alpha - nu,
"expected": 0.0,
"passes": abs(2 - alpha - nu) < tol,
},
"Fisher_independent": {
"claim": "Fisher residual = γ(2γ-3)/(1-γ) [NOT 0 generally]",
"value": gamma_chi - (2 - eta) * nu,
"expected_formula": gamma * (2 * gamma - 3) / (1 - gamma),
"passes": abs((gamma_chi - (2 - eta) * nu) - gamma * (2 * gamma - 3) / (1 - gamma)) < tol,
},
"eta_2gamma_REFUTED": {
"claim": "Paper 1's η=2γ residual > 0 in Phase A",
"value": 2 * (2 * gamma)**2 + 2 * gamma * gamma_chi + 1,
"expected": "positive (refutes η=2γ)",
"passes": (2 * (2 * gamma)**2 + 2 * gamma * gamma_chi + 1) > 0,
},
"D-14_nu_imprint": {
"claim": "ν_imprint · 2π = -1",
"value": (-1 / (2 * math.pi)) * 2 * math.pi,
"expected": -1.0,
"passes": abs((-1 / (2 * math.pi)) * 2 * math.pi - (-1)) < tol,
},
"D-SAGE-7": {
"claim": "c · |ν_imprint| · 2π = 3",
"value": 3 * (1 / (2 * math.pi)) * 2 * math.pi,
"expected": 3.0,
"passes": abs(3 * (1 / (2 * math.pi)) * 2 * math.pi - 3) < tol,
},
"nu_beta_id": {
"claim": "ν · β = -1",
"value": nu * beta,
"expected": -1.0,
"passes": abs(nu * beta - (-1)) < tol,
},
}
n_total = len(checks)
n_passed = sum(1 for c in checks.values() if c["passes"])
all_consistent = n_passed == n_total
return {
"input_gamma": gamma,
"phase": "A (γ ∈ (0,1))",
"n_checks_total": n_total,
"n_checks_passed": n_passed,
"all_consistent": all_consistent,
"framework_verified_by": "Sage Groebner basis (PolynomialRing(ℚ)) + Lean Mathlib4 (dependent type theory)",
"checks": checks,
"interpretation": (
f"All {n_total}/{n_total} D-SAGE identities consistent ✓ "
"(framework algebraic structure intact)"
if all_consistent
else f"INCONSISTENCY: {n_passed}/{n_total} pass. Possible bf16 outlier, "
"quantization artifact, or measurement noise."
),
"references": [
"Sage script: sage_recursive_sweep_2026-04-30.sage",
"Lean script: lean_taf/taf/Taf/Identities.lean",
"Paper 2 appendix A.4: appendix_formal_verification_2026-05-01.md",
],
}
def bimodal_phase_class(gamma: float) -> str:
"""§32.2 — Bimodal classifier (paper 2 §4 finding F11).
γ_text panel n=25 shows 2 density peaks (~0.75 + ~1.0) with gap 0.85-0.95.
Hartigan dip test pending (paper 2 Tier-A E3).
"""
if gamma is None:
return "unknown"
if gamma < 0:
return "catastrophic"
if gamma < 0.85:
return "Phase A (long-range)"
if gamma < 0.95:
return "boundary (gap zone)"
if gamma < 1.0:
return "Hagedorn boundary"
return "Hagedorn zone"
def nearest_famous_constant(gamma: float, max_results: int = 3,
tolerance: float = 0.05) -> list:
"""§32.3 — Convenience wrapper: find named constants near γ.
Wraps famous_constant_proximity(); always returns a list (possibly empty).
Useful for displaying "your γ is close to <constant>" in UI.
"""
if gamma is None:
return []
out = famous_constant_proximity(gamma, tolerance=tolerance)
return out.get("hits", [])[:max_results]
# ════════════════════════════════════════════════════════════════════════════
# Recipe registry
# ════════════════════════════════════════════════════════════════════════════
RECIPES = {
"X-1": {
"name": "Custom Training vs API",
"description": "Should I train a custom model or use a frontier API for my domain task?",
"fn": run_recipe_x1,
"params": ["N_params", "D_tokens", "gpu", "n_gpus", "mfu",
"api_model", "monthly_tokens_M"],
"category": "build-vs-buy",
"uses_sections": ["§17", "§19", "§20", "§24"],
},
"X-2": {
"name": "Long Context Viability",
"description": "Will model M serve length L doing Needle-in-a-Haystack retrieval?",
"fn": run_recipe_x2,
"params": ["theta", "T_train", "T_eval", "n_attention_heads", "n_kv_heads",
"d_head", "n_layers", "n_params", "has_SWA"],
"category": "long-context",
"uses_sections": ["§26", "§19"],
},
"X-3": {
"name": "Budget Pre-flight",
"description": "Given $ budget, what model is feasible to train?",
"fn": run_recipe_x3,
"params": ["USD_budget", "gpu", "mfu", "n_gpus"],
"category": "training-budget",
"uses_sections": ["§17", "§20"],
},
"X-5": {
"name": "Hardware Selection",
"description": "Which GPU should I use to serve my model at target throughput?",
"fn": run_recipe_x5,
"params": ["N_params", "T_eval", "n_layers", "n_kv_heads", "d_head",
"bytes_per_weight", "target_tokens_per_day", "concurrent_users"],
"category": "serving",
"uses_sections": ["§19", "§20"],
},
"X-19": {
"name": "KV Compression Decision",
"description": "Should I use soft decay, D_f cutoff, or literature methods to compress KV?",
"fn": run_recipe_x19,
"params": ["theta", "T_train", "T_eval", "n_attention_heads", "n_kv_heads",
"d_head", "n_layers", "n_params", "has_SWA"],
"category": "kv-compression",
"uses_sections": ["§26", "§19"],
},
"X-21": {
"name": "Imprint Purity Diagnostic",
"description": "How clean is the model's RoPE-Padé prediction? Predicts γ on RANDOM-token input via ν=−1/(2π).",
"fn": run_recipe_x21,
"params": ["theta", "T_train", "n_attention_heads", "n_kv_heads",
"d_head", "n_layers", "n_params", "T_eval", "gamma_random_obs"],
"category": "diagnostic",
"uses_sections": ["§26", "§28"],
},
"X-22": {
"name": "Compute-Context Invariant",
"description": "Does γ × log(N²·D) lie in the panel band 51.2 ± 16.8? Detects training/scaling anomalies.",
"fn": run_recipe_x22,
"params": ["theta", "T_train", "n_params", "gamma_obs", "D_tokens", "T_eval"],
"category": "diagnostic",
"uses_sections": ["§26", "§29"],
},
"X-23": {
"name": "IH-Phase Detector",
"description": "Is this model pre- or post-induction-head? Cheap probe via sign(γ_text − γ_random).",
"fn": run_recipe_x23,
"params": ["n_params", "gamma_text", "gamma_random"],
"category": "diagnostic",
"uses_sections": ["§30"],
},
}
def list_recipes() -> str:
"""Return JSON of all recipes for UI dropdown."""
return json.dumps([
{"id": rid, "name": r["name"], "description": r["description"],
"category": r["category"], "params": r["params"],
"uses_sections": r["uses_sections"]}
for rid, r in RECIPES.items()
])
def run_recipe(recipe_id: str, **params) -> dict:
"""Dispatcher — execute recipe by id with given params."""
r = RECIPES.get(recipe_id)
if r is None:
return {"error": f"unknown recipe '{recipe_id}'",
"available": list(RECIPES.keys())}
return r["fn"](**params)
# ════════════════════════════════════════════════════════════════════════════
# Known model presets
# ════════════════════════════════════════════════════════════════════════════
PRESETS = {
"EleutherAI/pythia-2.8b": {
"theta": 10000, "T_train": 2048,
"n_attention_heads": 32, "n_kv_heads": 32,
"d_head": 80, "n_layers": 32, "n_params": 2.8e9, "has_SWA": False,
},
"EleutherAI/pythia-1b": {
"theta": 10000, "T_train": 2048,
"n_attention_heads": 8, "n_kv_heads": 8,
"d_head": 256, "n_layers": 16, "n_params": 1e9, "has_SWA": False,
},
"EleutherAI/pythia-1.4b": {
"theta": 10000, "T_train": 2048,
"n_attention_heads": 16, "n_kv_heads": 16,
"d_head": 128, "n_layers": 24, "n_params": 1.4e9, "has_SWA": False,
},
"meta-llama/Meta-Llama-3-8B": {
"theta": 500000, "T_train": 8192,
"n_attention_heads": 32, "n_kv_heads": 8,
"d_head": 128, "n_layers": 32, "n_params": 8e9, "has_SWA": False,
},
"meta-llama/Llama-3.2-1B": {
"theta": 500000, "T_train": 131072,
"n_attention_heads": 32, "n_kv_heads": 8,
"d_head": 64, "n_layers": 16, "n_params": 1.2e9, "has_SWA": False,
},
"meta-llama/Llama-3.3-70B-Instruct": {
"theta": 500000, "T_train": 131072,
"n_attention_heads": 64, "n_kv_heads": 8,
"d_head": 128, "n_layers": 80, "n_params": 70e9, "has_SWA": False,
},
"mistralai/Mistral-7B-v0.1": {
"theta": 10000, "T_train": 8192,
"n_attention_heads": 32, "n_kv_heads": 8,
"d_head": 128, "n_layers": 32, "n_params": 7e9, "has_SWA": True,
},
"Qwen/Qwen2.5-7B": {
"theta": 1000000, "T_train": 32768,
"n_attention_heads": 28, "n_kv_heads": 4,
"d_head": 128, "n_layers": 28, "n_params": 7.6e9, "has_SWA": False,
},
"Qwen/Qwen2.5-1.5B": {
"theta": 1000000, "T_train": 32768,
"n_attention_heads": 12, "n_kv_heads": 2,
"d_head": 128, "n_layers": 28, "n_params": 1.5e9, "has_SWA": False,
},
"google/gemma-2-9b-it": {
"theta": 10000, "T_train": 8192,
"n_attention_heads": 16, "n_kv_heads": 8,
"d_head": 256, "n_layers": 42, "n_params": 9e9, "has_SWA": True,
},
"microsoft/phi-3-mini-4k-instruct": {
"theta": 10000, "T_train": 4096,
"n_attention_heads": 32, "n_kv_heads": 32,
"d_head": 96, "n_layers": 32, "n_params": 3.8e9, "has_SWA": True,
},
}
def list_presets() -> str:
return json.dumps([
{"id": k, "label": k.split("/")[-1],
"theta": v["theta"], "T_train": v["T_train"]}
for k, v in PRESETS.items()
])
def get_preset(model_id: str) -> dict:
return PRESETS.get(model_id, {})
# ════════════════════════════════════════════════════════════════════════════
# MODEL PROFILE — runs all 5 recipes with sensible defaults
# ════════════════════════════════════════════════════════════════════════════
def profile_model(theta, T_train, n_attention_heads, n_kv_heads, d_head,
n_layers, n_params, has_SWA=False,
T_eval=None, USD_budget=5000, target_tokens_per_day=10_000_000,
api_model="GPT-4o", monthly_tokens_M=10.0,
**_unused) -> dict:
"""Run all 5 recipes against the same model and assemble a TAF Card profile.
This produces the canonical paper §sec:gamma_decomposition view: one model,
all viability dimensions, with κey numbers + falsification status per dimension.
"""
if T_eval is None:
T_eval = T_train # default eval at training context
has_GQA = n_kv_heads < n_attention_heads
g_pade = gamma_pade(theta, T_eval)
decomp = gamma_decompose(g_pade, has_GQA, has_SWA, n_params)
g_corr = decomp["gamma_corrected"]
dh = d_horizon(theta, g_corr)
chi = chi_susceptibility(g_corr) if 0 < g_corr < 2 else None
# Architecture classification (paper species map)
if has_SWA:
arch_class = "SWA-alternating (gemma/phi family signature)"
elif has_GQA and n_params >= 4e8:
arch_class = "RoPE-GQA post-IH (Llama-3 / Mistral / Qwen-style)"
elif has_GQA:
arch_class = "RoPE-GQA pre-IH (small GQA model)"
elif n_params >= 4e8:
arch_class = "RoPE-MHA post-IH (classical Llama-2 / pythia-large)"
else:
arch_class = "RoPE-MHA pre-IH (small MHA model)"
common_params = {
"theta": theta, "T_train": T_train, "T_eval": T_eval,
"n_attention_heads": n_attention_heads, "n_kv_heads": n_kv_heads,
"d_head": d_head, "n_layers": n_layers, "n_params": n_params,
"has_SWA": has_SWA,
}
# Run all 5 recipes
results = {}
try:
results["X-2"] = run_recipe_x2(**common_params)
except Exception as e:
results["X-2"] = {"error": str(e)}
try:
results["X-19"] = run_recipe_x19(**common_params)
except Exception as e:
results["X-19"] = {"error": str(e)}
try:
results["X-1"] = run_recipe_x1(N_params=n_params, gpu="H100 SXM",
n_gpus=8, mfu=0.45,
api_model=api_model, monthly_tokens_M=monthly_tokens_M)
except Exception as e:
results["X-1"] = {"error": str(e)}
try:
results["X-3"] = run_recipe_x3(USD_budget=USD_budget, gpu="H100 SXM",
mfu=0.45, n_gpus=1)
except Exception as e:
results["X-3"] = {"error": str(e)}
try:
results["X-5"] = run_recipe_x5(N_params=n_params, T_eval=T_eval,
n_layers=n_layers, n_kv_heads=n_kv_heads,
d_head=d_head, bytes_per_weight=2.0,
target_tokens_per_day=target_tokens_per_day,
concurrent_users=1)
except Exception as e:
results["X-5"] = {"error": str(e)}
# Falsification status (from FALSIFICATION.md F1-F23)
falsifications = []
if 0 < g_corr < 1:
falsifications.append({"id": "F1", "claim": "γ_Padé median MAE < 5%", "status": "✅ in scope"})
if dh is not None:
falsifications.append({"id": "F2", "claim": "d_horizon predicts NIAH ±1%", "status": "✅ applicable"})
falsifications.append({"id": "F17", "claim": "Soft KV decay regime",
"status": "✅ applies" if dh >= T_train / 2 else "⚠ refuted regime"})
if has_GQA:
falsifications.append({"id": "F10", "claim": "GQA Δγ < -0.1 ⇒ post-IH",
"status": "✅ in scope"})
if has_SWA:
falsifications.append({"id": "F11", "claim": "SWA Δγ > +0.3 (gemma signature)",
"status": "✅ in scope"})
# Sesión 29 / paper 2 visual diagnostics
safety = hagedorn_safety_alert(g_corr)
phase_cls = bimodal_phase_class(g_corr)
constants = nearest_famous_constant(g_corr, max_results=2, tolerance=0.02)
n_params_M = n_params / 1e6
gamma_random_pred = gamma_random_predict(theta, T_eval, n_params_M)
K_inv = compute_invariant_K(g_corr, n_params_M)
return {
"model_summary": {
"architecture_class": arch_class,
"n_params": n_params,
"T_train": T_train,
"T_eval": T_eval,
"rope_theta": theta,
"has_GQA": has_GQA,
"has_SWA": has_SWA,
},
"key_numbers": {
"gamma_pade": g_pade,
"gamma_decomposed": g_corr,
"decomposition_breakdown": decomp,
"d_horizon": dh,
"L_NIAH_ceiling": l_niah_c(dh),
"chi_susceptibility": chi,
"kv_memory_per_request_GB": kv_cache_memory(n_layers, n_kv_heads,
d_head, T_eval)["GB"],
"gamma_random_predicted": gamma_random_pred,
"compute_invariant_K": K_inv["K"],
"K_in_distribution": K_inv["in_distribution"],
},
"v04_diagnostics": {
"hagedorn_safety": safety,
"bimodal_phase_class": phase_cls,
"nearest_famous_constants": constants,
"imprint_predicted_shift": gamma_random_pred - g_pade,
"compute_invariant": K_inv,
},
"recipes": {
rid: {
"verdict": r.get("verdict", "ERROR"),
"reason": r.get("reason", r.get("error", "")),
"mitigation": r.get("mitigation", ""),
"name": r.get("recipe_name", ""),
}
for rid, r in results.items()
},
"falsification_status": falsifications,
}
# ════════════════════════════════════════════════════════════════════════════
# COMPARE — same recipe across multiple models
# ════════════════════════════════════════════════════════════════════════════
def compare_models(model_specs: list, recipe_id: str = "X-2",
shared_params: dict = None) -> dict:
"""Run one recipe across multiple models and assemble side-by-side comparison.
Args:
model_specs: list of dicts each with model architectural params + a "label" key
recipe_id: which recipe to run (X-1, X-2, X-3, X-5, X-19)
shared_params: extra params to pass to recipe (T_eval, etc.)
"""
shared_params = shared_params or {}
rows = []
for spec in model_specs:
label = spec.pop("label", spec.get("model_id", "model"))
params = {**spec, **shared_params}
try:
result = run_recipe(recipe_id, **params)
rows.append({
"label": label,
"verdict": result.get("verdict", "ERROR"),
"reason": result.get("reason", ""),
"key_numbers": _extract_key_numbers(result),
})
except Exception as e:
rows.append({"label": label, "verdict": "ERROR", "reason": str(e), "key_numbers": {}})
return {
"recipe_id": recipe_id,
"recipe_name": RECIPES.get(recipe_id, {}).get("name", recipe_id),
"shared_params": shared_params,
"rows": rows,
}
def _extract_key_numbers(result: dict) -> dict:
"""Pull the numerically interesting fields from a recipe result for compare table."""
nums = {}
for step in result.get("chain", []):
name = step.get("name", "")
res = step.get("result")
if res is None:
continue
if isinstance(res, (int, float)):
nums[name] = res
elif isinstance(res, dict) and "GB" in res:
nums[name] = res["GB"]
return nums
# Smoke test
if __name__ == "__main__":
print("─── X-2 Llama-3-8B @ 32K ───")
r = run_recipe("X-2", theta=500_000, T_train=8192, T_eval=32_000,
n_attention_heads=32, n_kv_heads=8, d_head=128,
n_layers=32, n_params=8e9, has_SWA=False)
print(f"Verdict: {r['verdict']}{r['reason']}\n")
print("─── X-1 Llama-3-8B vs GPT-4o (10M tok/mo) ───")
r = run_recipe("X-1", N_params=8e9, monthly_tokens_M=10.0, api_model="GPT-4o")
print(f"Verdict: {r['verdict']}{r['reason']}\n")
print("─── X-3 budget $5K ───")
r = run_recipe("X-3", USD_budget=5000.0, gpu="H100 SXM", n_gpus=1)
print(f"Verdict: {r['verdict']}{r['reason']}\n")
print("─── X-5 serve Llama-3-8B at 4K ───")
r = run_recipe("X-5", N_params=8e9, T_eval=4096, n_layers=32, n_kv_heads=8, d_head=128,
target_tokens_per_day=10e6, concurrent_users=1)
print(f"Verdict: {r['verdict']}{r['reason']}\n")
print("─── X-19 KV compression for Llama-3-8B ───")
r = run_recipe("X-19", theta=500_000, T_train=8192, T_eval=8192,
n_attention_heads=32, n_kv_heads=8, d_head=128,
n_layers=32, n_params=8e9)
print(f"Verdict: {r['verdict']}{r['reason']}\n")