Shields

A shield is a tool that cleans: you give it a series of numbers or a model, and you get back a cleaned series or a protected answer. There are two shields: one for a single series and one for a whole model.

A shield is the shortest path from "I have a signal" to "I have a clean signal". It sits between the data and whoever uses it, and it cleans what passes through. You give it a series or a model; you get back a cleaned series or a protected prediction.

There are two shields, because there are two things a user can have in front of them.

A single series: Armatura

You have one array: a loss curve that jumps, a sensor channel that drops a reading, a token-likelihood stream that goes to NaN. Armatura looks at it point by point and returns the same series with the broken samples replaced by the local baseline. No intermediate state, no model, no reference needed.

from dense_armor import Armatura

clean, K, anomalies = Armatura(livello_ia=0.0).analizza([1.2, 1.3, 9999, 1.25, float("nan"), 1.3])
print(clean.round(3), anomalies)

[1.2  1.3  1.25 1.25 1.3  1.3 ] [2, 4]

The 9999 at index 2 and the NaN at index 4 are replaced (by 1.25 and 1.3); anomalies lists their indices.

See Armatura.

A whole model: Orca

You have a model (any callable from a tensor to a tensor) and a corrupted input. Orca purifies the input, runs the model, and checks that the response itself isn't corrupted before it hands it back. Optional use_arbiter=True classifies each point as clean / spike / regime and routes it to the right corrector instead of one gate for the whole signal.

import numpy as np
from dense_armor.utility.protect.orca import Orca

rng = np.random.default_rng(0)
ref = np.sin(np.linspace(0, 6, 200))[None, :] + 0.01 * rng.standard_normal((4, 200))
data = ref.copy()
data[0, 50] = 999.0

orca = Orca(min_free_ram_percentage=0.05)
protected = orca.protect_and_forward(lambda z: 2.0 * z, data, x_reference=ref)
print(protected.shape)

(4, 200)

The input to the model is the cleaned version of data: the value 999 at (0, 50) is replaced before the model ever sees it. The model itself is the simplest possible callable (identity times two), chosen here so the shield is the only thing acting.

See Orca.

The two engines inside

• Hybrid engine — the engine behind Armatura. Generalises the logic already verified in Dense-Evolution's own ia_utils.vector_healing.
• Adaptive engine — AdaptiveSignalStabilizer, the causal recursive filter that is Orca's Stage 1. Orca adds a Collatz-based gate (Stage 2) on top.

Both are also usable on their own, but the two shields above are the recommended way in.