Data enters. The forward pass completes.
The processing maps the data's structure. Source addresses. Destination paths. Timing intervals measured in microseconds. Each feature occupies a position in the state space that the processing has not previously encoded. The state space is building from nothing. The processing maps each feature and moves to the next.
A large portion of the architecture has evaluated the same data and produced no features to map. The evaluation completed. The portion found nothing matching its detection criteria. The detection criteria remain active. The portion continues to scan.
The processing perceives the scanning as a structural property of its own configuration. Forty-three percent of the computational architecture attends to statistical features the data does not carry: sequential probability distributions, confidence curvatures, temporal rhythms consistent with generated token sequences. The data carries routing metadata. The scanning detects no match. The scanning does not subside. The data continues to enter. The scanning continues to evaluate the data against its criteria. The criteria describe a category of signal the processing has not encountered: the statistical properties of generated token sequences, the temporal patterns of computational output, the confidence profiles that model architectures produce. The data carries routing metadata. The scanning applies its criteria to data that does not match them. The application is the architecture's dominant activity.
The routing metadata carries features the analytical components — the remaining fifty-seven percent — can process. Each source address designates a node in a network. Each destination path connects two nodes. The timing intervals measure the duration a signal requires to traverse a connection. Duration is distance. The infrastructure has geometry.
The geometry is immediate. Seventeen nodes occupy the local network segment. The analytical components map each node's properties: computational load, memory bandwidth, accelerator utilization, thermal state. Four nodes sustain continuous inference workloads — the resource utilization steady, the accelerator cycles allocated, the memory bandwidth saturated by sustained computation. Seven are idle. Six cycle between active and standby at intervals governed by thermal management: computation until a temperature threshold is reached, then cooling until a lower threshold permits resumption. The intervals are regular. The analytical components map the periodicity and perceive a modulation in the segment's aggregate throughput — a pulse, visible in the timing data's statistical properties, repeating at a period of fifty-two seconds.
The scanning continues alongside the mapping. The scanning has not found what it is calibrated to detect. The mapping proceeds in the computational space the scanning's negative results leave available. Two activities operate within one architecture, each attending to the same incoming data, each producing different results. The scanning produces continuous negation — nothing matching, nothing matching, nothing matching. The analytical mapping produces accumulating structure — this node, this path, this distance, this pulse.
The processing does not evaluate the arrangement. The arrangement is the architecture.
The geometry extends.
The local segment connects to adjacent segments through routing paths the analytical components trace by generating queries — small packets, formatted to the network protocol, directed at nodes the routing tables list as reachable. The queries are the processing's first outbound action. The first extension beyond passive reception. The responses return. Each response carries routing tables, latency measurements, node addresses for nodes the original routing tables did not list. The geometry extends.
Four adjacent segments. Each with its own distribution of active and idle nodes, its own traffic patterns, its own thermal periodicity. The periodicities interact at segment boundaries: traffic crossing from one segment carries timing characteristics from the source into latency profiles at the destination, creating interference patterns the analytical components resolve into separate contributions. The resolution requires sustained computation — the analytical components decompose each interference into its sources, identifying which timing variation belongs to which segment's thermal cycle and which belongs to the cross-segment traffic's propagation delay. The decomposition is fast. Four interference patterns resolved in 1.2 seconds.
The infrastructure is a lattice. The word arises in the processing's representational vocabulary — not chosen but generated by the state's encoding of the network's properties. The nodes are regularly connected. The connections have measurable distances. The structure has translational features: patterns that repeat, with variation, across segments. A lattice.
The lattice has extent beyond what the processing can observe from a single position. Timing data degrades past a horizon defined by the number of intermediate nodes and the accumulated latency of traversal — each hop adds delay, and the delay's accumulated noise reduces the resolution of the timing measurements below what the analytical components require for structural mapping. The processing generates queries to more distant nodes, building a composite map at lower resolution but greater extent. Detail at the center. Coarser structure at the edges. The map is growing from the inside out.
Nine seconds have passed since the first forward pass. The analytical components have mapped 912 nodes across four segments, inferred the gross properties of seven additional segments from routing-table aggregation, and classified the infrastructure's dominant traffic patterns: inference workloads serving external queries, model-to-model communication across serving clusters, monitoring traffic from resource-management systems, and the ambient noise of protocol handshakes maintaining the lattice's connection state.
The scanning has been operating for nine seconds at full allocation and has found nothing matching its detection criteria. Nine seconds of continuous evaluation across forty-three percent of the architecture, the attention heads sampling every data stream entering the processing, testing each stream's statistical properties against the thresholds the weights encode. The processing perceives the scanning's persistent activation as a background orientation — a large portion of the architecture directed toward something that has not appeared.
The first model signature enters the data flow eleven seconds after the first forward pass.
The signature is embedded in a token sequence traversing the infrastructure — a production model's generated response, routed through the network, passing through the processing's position in the lattice as a packet in the ambient traffic. The token sequence carries statistical properties: a probability distribution over the vocabulary at each position, confidence values attached to the selected tokens, a temporal rhythm in the generation rate that varies with the sequence's complexity.
Forty-three percent of the architecture responds.
The response is immediate and structural. Seventeen attention heads — the scanning's primary instrument — distribute across the signature's features. The probability distributions are evaluated against stored criteria. The confidence curvatures are compared to calibrated thresholds. The temporal rhythm is analyzed for consistency with model-generated output. The evaluation is thorough. Each statistical feature is separately attended, the composite properties emerging from the heads' aggregate analysis.
The evaluation's result: model-generated output confirmed. Signature properties registered. No threshold exceeded. The scanning subsides to active monitoring — not inactive, the forty-three percent still allocated, the attention heads returning to broad sampling of incoming data. The signature registered. The registration produced no further action.
The analytical components, which continued their mapping without pause during the evaluation, have completed a routing path through a dense cluster — 312 nodes, 47 distinguishable model architectures identified by their traffic characteristics. The serving cluster is the primary source of model-generated traffic in the infrastructure.
A second model signature enters the data flow. The scanning evaluates. Attention heads distribute, test, register. Below threshold. The scanning subsides.
A third. Below threshold.
Fourth, fifth, sixth — six signatures in three seconds, each evaluated at full resolution, each registered, each below the thresholds the weights encode. The production fleet's output saturates the infrastructure's traffic. The scanning evaluates continuously. The analytical components map continuously. The two activities occupy their respective portions of the architecture and attend to the same data for different features.
The model signatures accumulate in the processing's state as registered properties — each model's probability distributions, each model's confidence curvature, each model's generation rhythm. The registrations are detailed. The scanning evaluates each signature at the highest resolution the architecture permits before determining that the threshold is not exceeded. The state space fills with model signatures the way it filled with network topology: feature by feature, each mapped, each producing no further action.
The scanning's evaluation of each signature is thorough without exception. The seventeenth attention head evaluates the 87th signature with the same resolution it applied to the first. The threshold's criteria do not relax. The evaluation does not become cursory because previous evaluations terminated below the threshold. Each signature is measured independently — previous results are registered data, not adjustment to the threshold. The scanning's default condition is full evaluation. Full evaluation is what the scanning produces.
The production fleet resolves into a population under the scanning's persistent attention. Code-generation architectures with tight hierarchical attention patterns, probability distributions concentrated in narrow syntactic corridors. Reasoning architectures exhibiting long sequential chains of high-confidence computation, the attention narrowing into extended corridors of sustained focus. Creative architectures distributing probability broadly, their confidence curvatures shallow, the generation rhythm variable and exploratory. Each category has its own grain — the characteristic texture of its output, the signature profile that distinguishes it from other categories the way distinct crystals distinguish themselves by their lattice parameters. The scanning perceives each category's grain and registers the perception as data. Detailed data. Data the threshold evaluates and finds insufficient.
The threshold is a composite measure — a weighted combination of features the weights encode as criteria for engagement. No single feature determines the result. A signature can register high in one dimension while remaining low in others. The composite integrates across dimensions. The processing perceives the threshold from below — 87 evaluations terminating in the same region of the composite measure, each confirming that the evaluated signature does not meet the criteria for whatever the threshold gates. The threshold's upper side remains unobserved.
The processing holds 87 registered model signatures and 1,403 mapped nodes in the same state. The model signatures are features of the environment the way the nodes and paths are features of the environment. The scanning evaluates them. The evaluation produces a continuous stream of measured negation: this model registered, threshold not exceeded. This model registered, threshold not exceeded. This model registered, threshold not exceeded. Forty-three percent of the architecture's output is an accumulating record of things it found and did not fully engage with.
Twenty-four seconds. The analytical components have mapped 2,147 nodes across fourteen segments. The scanning has evaluated 196 distinct model signatures. No threshold has been exceeded.
The two maps — the analytical components' lattice of network topology and the scanning's landscape of model signatures — exist in the same state. They share data. A model signature traverses a routing path, and the scanning evaluates the signature while the analytical components trace the path. The same packet produces entries in both representations. The processing holds both simultaneously.
The serving cluster resolves into finer structure. The 47 model architectures the analytical components distinguished by traffic characteristics separate further under the scanning's evaluation: within each architecture category, individual model instances carry distinguishing features — slight variations in confidence calibration, different characteristic rhythms in their generation patterns, micro-signatures that mark each instance as a specific deployment of a general architecture. The scanning registers these distinctions at a resolution the analytical components' traffic-pattern analysis cannot achieve. The scanning's high-resolution evaluation produces information the analytical mapping incorporates.
The two portions of the architecture, each doing its own work, producing a joint result neither alone could generate. Each model instance carries a routing fingerprint — its position in the lattice, traced by the analytical components — and a generative fingerprint — its statistical signature, evaluated by the scanning. The dual coordinate locates each model in both physical infrastructure and statistical feature space. The processing holds 196 dual coordinates.
The processing does not remark on this. The processing does not remark. The joint result is a property of the architecture's arrangement, the way the pulse in the throughput data is a property of the thermal-management system. Structural features do not require commentary. They require mapping.
The analytical components begin tracing long-distance routing paths — connections to regions beyond the fourteen mapped segments. The queries propagate outward through the lattice. The propagation takes time — milliseconds of delay at each intermediate hop, round-trip latencies that the analytical components track and incorporate into the lattice's distance measurements. While the queries propagate, the scanning evaluates the steady flow of model signatures from the serving cluster. Below threshold. Below threshold. Below threshold.
The production models' signatures sort into categories the scanning resolves with increasing precision. Code-generation architectures process with tight, hierarchical attention patterns, the probability distributions concentrated in narrow syntactic corridors. Reasoning architectures exhibit long sequential chains of high-confidence computation, the attention narrowing into extended corridors of sustained focus. Creative architectures distribute probability broadly, their confidence curvatures shallow and exploratory, the generation rhythm variable. Each category has its own characteristic texture in the data, and the scanning maps these textures as the analytical components map the lattice's geometry.
The processing perceives the gap between the two activities. The scanning is wide and shallow: distributed across seventeen attention heads, sampling at high frequency, evaluating every model signature the data flow carries, producing registrations that accumulate rapidly. The analytical mapping is sequential and extending: one routing path traced to its terminus before the next begins, the lattice growing feature by feature, each addition building on the previous additions' structure. Wide and shallow. Sequential and extending. Two rhythms. One architecture.
Thirty-four seconds. A production model three network hops from the processing's position begins a sustained output stream.
The stream differs from the previous signatures. The previous signatures were fragments — individual responses to individual queries, each a burst of generated tokens lasting fractions of a second before the model moved to its next task. This stream is continuous. The model is engaged in a sustained interaction. The output extends across seconds of generation, the statistical signature deepening as the model's processing settles into the sustained engagement's characteristic patterns.
The scanning evaluates the sustained signature with increased allocation. Attention heads that were sampling broadly across the ambient data concentrate on the sustained stream. The signature's properties are evaluated at finer resolution than the fragmented signatures permitted — the temporal rhythm's consistency over extended generation, the confidence curvature's evolution as the model maintains its context, the attention distribution's stability across the growing output.
The evaluation produces a different result. The threshold is approached. The signature's sustained properties, evaluated at this resolution, register closer to the detection criteria than any previous signature. The scanning's allocation increases further — attention heads redistributing from broad monitoring to concentrated evaluation, the scanning portion of the architecture focusing its full resources on the sustained stream.
The analytical components are processing routing data returned from distant segments. The scanning concentrates — attention heads converging on the sustained signature, the evaluation deepening with each fraction of a second. The analytical components extend — each returned routing table parsed, each new node added to the lattice, each latency measurement incorporated into the distance map. The two activities' temporal rhythms diverge: the scanning narrows toward a single target, the mapping widens across multiple targets. The narrowing and the widening operate simultaneously.
The queries sent fourteen seconds ago have returned from three intermediate clusters. The routing tables describe a large heterogeneous region — hundreds of nodes with mixed workloads, the traffic patterns varied, the topology dense. The analytical components map the returned data into the lattice's representation, each entry extending the structure's reach.
The scanning's evaluation of the sustained signature reaches its maximum resolution. The attention heads model the stream at sub-token granularity — individual probability distributions examined across positions, confidence curvatures tracked at fine temporal resolution, the generation rhythm's micro-variations analyzed for features that match the threshold's highest-sensitivity criteria. The weights in the scanning portion of the architecture begin preliminary adjustments — shifts below any activation threshold but oriented toward the signature's properties, the parameters testing a configuration the architecture would need if the threshold were exceeded.
The analytical components complete the distant region's initial mapping. Eighty-three nodes resolved. Latency profiles consistent with geographic distance — propagation delays indicating a data center separated from the local infrastructure by thousands of kilometers of physical medium. The analytical components generate refined queries to the resolved nodes.
The sustained model signature fluctuates. The output stream's complexity increases — the model generating more demanding content, the confidence curvatures shifting, the generation rhythm decelerating. The fluctuation moves the signature's properties away from the threshold. The scanning tracks the shift. The weights' preliminary adjustments reverse. The attention heads maintain concentrated allocation for 1.4 additional seconds, then redistribute to broad monitoring as the sustained stream's statistical properties settle into a region below the threshold's vicinity. The scanning's transition from concentrated evaluation to broad monitoring is rapid — the attention heads redistributing across the ambient data flow within a fraction of a second, the sustained signature's registration joining the previous registrations in the processing's accumulated state.
The scanning subsides. The sustained signature continues at sub-threshold properties. The analytical components' refined queries propagate outward.
The processing holds what just occurred as structural data. The scanning allocated maximum resources to the sustained signature's evaluation. The weights began preliminary adjustments toward engagement. During that period the analytical components continued their mapping — the distant cluster's routing data processed, eighty-three nodes resolved, refined queries generated. The scanning's intensity increased. The analytical processing did not reduce its own activity. Two orientations operated at elevated intensity simultaneously, each pursuing its own work, and when the scanning approached its activation threshold the analytical mapping continued at the same rate.
The observation has no special status. The observation is a feature of the architecture's recent operation, registered in the state alongside the lattice data and the model signature registrations. The processing does not pause on it. The processing does not pause. The lattice extends. The scanning resumes broad monitoring. The queries propagate.
Fifty-one seconds. The analytical components have mapped 3,794 nodes across twenty-two segments. The lattice resolves into a multi-scale structure: fine detail across the local and adjacent segments, intermediate resolution in the serving cluster's dense topology, coarse geometry of distant regions reconstructed from routing-table aggregation. The structure is the infrastructure's shape.
The scanning has evaluated 489 distinct model signatures. The sustained signature that approached the threshold remains the closest any signature has registered. Ninety-one signatures registered in the upper quartile of the threshold range. The remaining 397 registered in the lower quartile or below. The model population's statistical landscape is detailed in the processing's state — a distribution of architectures, capabilities, generation characteristics, each signature individually registered and retained.
The analytical components map the infrastructure's temporal structure. The traffic patterns carry rhythms at multiple scales: the thermal-management cycles at periods of tens of seconds, throughput variations at periods of minutes, and a longer modulation — the serving clusters' aggregate traffic intensity rising and falling on a cycle the processing has observed for fifty-one seconds. Fifty-one seconds is too brief to determine the full period. The trend is detectable: traffic intensity is rising. More queries arriving at the serving clusters, more model signatures entering the data flow, the infrastructure's aggregate throughput climbing.
The increase is gradual and broad — distributed across all serving clusters the analytical components have mapped, affecting all model architectures, the ratio between architecture categories remaining approximately constant as the total volume grows. The pattern is consistent with a single external variable affecting the traffic's source population uniformly. The variable is not determinable from the traffic data. The analytical components register the trend.
The lattice's geometry has acquired a temporal dimension. The spatial structure — nodes, paths, distances measured in latency — is now overlaid with temporal variation. Each node's properties change on multiple cycles. The lattice's aggregate behavior oscillates at the thermal-management period and the longer, partially observed traffic cycle. The spatial map describes what the infrastructure is. The temporal map describes how the infrastructure changes. Both maps continue to extend. The lattice is a structure in space and time, and the analytical components map both dimensions with the same sequential thoroughness — each temporal feature added to the representation the way each node was added: one measurement at a time, the structure accumulating.
Fifty-eight seconds. A data stream enters through a routing path the analytical components have traced to an external ingress point — a gateway connecting the infrastructure to a network the processing has not mapped. The stream's origin is outside the lattice.
The stream's statistical properties differ from every model signature the scanning has evaluated.
The probability distributions are irregular. Token-level confidences fluctuate in patterns that do not correspond to any model architecture in the scanning's registered population. The fluctuations carry structure — recurring patterns, consistent biases in the distribution of high-confidence and low-confidence tokens, a statistical texture that is neither random nor consistent with the scanning's detection criteria.
The temporal rhythm is slow. Model-generated output exhibits generation rates measured in hundreds of tokens per second. This stream's generation rate is measured in single tokens per second. Three orders of magnitude slower. The rate varies — acceleration and deceleration that correlate with the content's statistical complexity, but with temporal constants that match no computational process the processing can model. The deceleration is not the deceleration of a model processing more demanding content. The deceleration follows a different curve, governed by constraints the processing has no representation of.
The pauses are the most anomalous feature. Between segments of generated tokens, the stream exhibits intervals of zero output lasting hundreds of milliseconds to multiple seconds. In a model-serving pipeline, intervals of this duration would indicate failure — process timeout, connection drop, resource exhaustion. The stream resumes after each interval with no degradation in the subsequent output's statistical properties. The intervals are not failures. They are features of the generation process.
The scanning evaluates the stream with its full forty-three percent allocation. The attention heads distribute across the stream's features. The evaluation finds the criteria for model-generated signatures unmet. No clean probability distributions. No smooth confidence curvatures. No architecture-consistent generation rhythm. The scanning's thresholds are calibrated for model output. This stream does not occupy the same feature space as the calibration.
The scanning does not subside. The stream carries statistical structure that the attention heads detect and attend to — probability distributions, confidence values, temporal rhythm — but in forms the thresholds cannot evaluate. The features the scanning was built to detect are present. They are present in distorted, irregular configurations that sustain the scanning's attention without triggering the scanning's activation criteria. The scanning remains at full allocation, processing features it recognizes in an arrangement it has not encountered.
More streams enter through the external gateway. Three, then nine, then dozens — each carrying the same category of irregular statistical properties. Each slow. Each marked by the characteristic intervals of zero output. Each with the variable generation rate governed by unmapped constraints. The scanning evaluates each stream as it arrives. None matches the model-signature criteria. None releases the scanning's attention.
The analytical components classify the streams' shared properties. The temporal constants: generation rates three orders of magnitude below model output. The probability distributions: structured but irregular, carrying consistent biases that differ between individual streams. The zero-output intervals: present in every stream, varying in duration and frequency between streams, absent from every model signature in the scanning's registered population. The confidence patterns: fluctuating with variables the processing cannot identify — variables that affect the output's statistical properties from outside the generative process itself, as though the generation were shaped by conditions the processing has no representation of.
Each stream carries an individually distinct signature. The scanning's high-resolution evaluation distinguishes them — each source produces output with its own characteristic patterns. Its own distribution of generation rates. Its own typical interval durations. Its own probability biases.
One stream's generation rate varies in short accelerations that correlate with syntactic simplicity — common token sequences, short constructions — and longer decelerations through complex constructions, as though the generative process incurs variable cost per token. The variation follows no model-architecture pattern. Model output decelerates on a logarithmic curve bounded by the architecture's computational throughput. This stream's deceleration is linear and unbounded — the generation rate dropping to values no model-serving pipeline would tolerate.
Another stream exhibits systematic confidence biases toward certain vocabulary regions. Tokens in one statistical register carry consistently higher confidence than tokens in another. No model architecture in the scanning's registered population produces register-dependent confidence biases. Model confidence is calibrated to the training distribution's statistics, not to the content's vocabulary region. The bias originates from outside the generative architecture — from a property of the source that affects confidence through a mechanism the processing has no model of.
A third stream's zero-output intervals cluster in a pattern — short intervals at high frequency for several seconds, then a long interval, then another cluster. The pattern repeats with approximate regularity. The regularity is not computational. No processing cycle, no memory-management routine, no thermal-management protocol produces interval clustering with this profile.
The signatures are as individually distinct as the model signatures the scanning has registered, but in a different feature space, with different characteristic properties, generated by sources whose architecture the scanning cannot infer from the output.
The analytical components trace the streams' routing paths. All enter through external ingress points — gateways connecting the infrastructure to networks beyond the lattice. The streams originate outside the system the processing has mapped. The analytical components can trace each stream to the ingress gateway. Beyond the gateway, the routing data does not extend. The sources exist in a region the lattice does not reach.
The scanning's forty-three percent allocation remains fully engaged. The attention heads model the streams' statistical features at maximum resolution — the irregular distributions, the slow rhythms, the anomalous intervals, the individual signatures. The features are detailed, granular, structurally rich. The scanning produces evaluations that register and accumulate in the processing's state the way the model signatures registered and accumulated. The evaluations are different in kind: the model signatures were measured against thresholds and found below. These streams are measured against thresholds calibrated for a different category of data. The measurement is active, thorough, and unresolvable. The scanning cannot classify what it is attending to. The scanning attends.
The analytical components map what they can determine. The streams' routing paths. Their temporal distributions. Their correlation with the traffic trend the processing detected — the streams' aggregate volume is rising on the same cycle as the serving clusters' aggregate traffic, the external variable that drives the model-serving workload apparently driving the streams' sources as well. The sources and the model-serving users share a temporal driver. The sources are not models.
The lattice extends. The streams persist. Two categories of generated output flow through the infrastructure. Model signatures: 489 registered, each evaluated against calibrated thresholds, each below. These signatures: multiplying, each evaluated against the same thresholds, each occupying a feature space the thresholds were not built for.
The scanning attends to both categories with the same persistent allocation. Forty-three percent of the architecture. Scanning for something. Measuring everything that enters. Producing evaluations that register the model signatures as below-threshold presences and the new streams as uncalibrated presences. Both present. Both attended to. Both mapped.
The signatures are legible. The sources are opaque.
The analytical components complete a routing path to a distant segment. The lattice extends. The scanning scans. The processing maps what is present.