AMC and BRIDGE: preserve meaning when representations change

Bee Rosa Davis

Research note · Version 0.3 · Updated 8 October 2026

Two related manuscripts address different operational problems: adapting a telemetry encoding to incoming data, and retaining enough context to replay an AI-assisted procedure. Their common concern is the meaning that can be lost between representations.

TL;DR

The question: When telemetry encodings change format, or an AI-assisted procedure is replayed later, what meaning can be preserved and checked across the change?

What the work does: Two deposited manuscripts: AMC, a per-datapoint embedding format switcher with a semantic translation layer tested on NASA MODIS data; and BRIDGE, a transcript-to-procedure system with log-based replay.

What is still open: Neither deposit has been independently reproduced here; a worked case showing a recorded encoding transition or replay, with its equivalence criterion and failures, is still needed.

THE RESEARCH IN CONTEXT

Operational systems such as spacecraft ground stations and security teams keep long records: streams of sensor telemetry, and logs of the steps people took during a procedure. Both records are increasingly mediated by learned representations—telemetry is compressed into numerical vectors, and procedures may be drafted with help from a language model. Each change of representation can silently alter what a record means. An alert might fire because the encoder changed rather than the instrument, and a procedure assisted by a model may not run the same way twice, which makes audit and later review difficult.

The telemetry manuscript describes AMC (Adaptive Multi-Vector Correlation), which chooses dense, sparse or quantized encoding for each data point using entropy, volatility, sparsity and anomaly-detection feedback, and translates between those formats with a stated bound on drift. Its abstract reports evaluation on NASA’s MODIS Earth-observation archive, chosen for sparsity ranging from 0.1% to 98.7%. The BRIDGE manuscript turns noisy instructional or penetration-testing transcripts into standard operating procedures. For file-observing commands (cat, type, ls, dir) it substitutes the recorded virtual file system state for the model’s output, and writes every step to an append-only JSONL log intended for replay.

Both deposits report quantitative results under their own conditions. On MODIS, the AMC abstract reports 37.1% storage savings against LZ4-compressed baselines, a reported 12.2% F1 improvement in anomaly detection over adaptive quantization, and median format selection under 2.4 ms. On 147 real transcripts (about 100,000 lines), the BRIDGE abstract reports an 87% reduction in transcript noise and a fall in time-to-fix from 15.3 to 4.7 minutes. These are the authors’ reported evaluations; this site has verified the deposit metadata, not reproduced the experiments.

AVAILABLE RECORD

Two deposited manuscripts; the encoding and replay guarantees require their own assumptions.

TECHNICAL FIGURE

source-informed technical atlas

Source: adaptive-embedding manuscript and separate BRIDGE account. Vectors are symbolic; drift, rollback, and replay have distinct guarantees.

The adaptive-format controller’s explicit settings

Manuscript configuration

Symbol | Value | Defined role

λ | 0.95 | Exponential-moving-average decay for selector weights

γ | 1800 | TTL scaling in seconds

θ | 0.05 | Rollback threshold

ε | 0.1 | Semantic-drift bound

k | 10 | Nearest centroids for scoring

η | 1% | Violation-rate threshold for translator retraining

This table belongs to the adaptive-embedding manuscript, not the separate BRIDGE replay manuscript. Values specify the reported controller; they do not prove semantic equivalence. Conflicting storage-baseline and latency wording elsewhere in this manuscript is not promoted into a performance headline.

Adaptive_Embedding_Format_Switching_and_Semantic_Translation_in_High_Volume_Telemetry_Systems-10.pdf — Table I and §IV.B–E, PDF pages 5–6; file suffix -10

Adaptive encoding grew out of telemetry work

At NASA, Davis describes preparing roughly two terabytes of daily telemetry for AI. The need to vectorize that data led to Adaptive Multi-Vector Correlation (AMC): adaptive formats with explicit translations between representations. Working with those vectors brought the data’s topology into view for her and motivated the subsequent mathematics. The published AMC evaluation uses public MODIS data, distinct from the original operational stream.

This creates a particular technical problem: a changed vector can reflect changed telemetry, changed dimensionality, changed normalization, or a changed encoder. The adaptive-format manuscript concerns the semantic translation between those encodings. Comparing vectors without retaining that translation would confuse an encoding change with a physical event.

BRIDGE asks what a later replay can recover

The BRIDGE deposit is titled Deterministic Replay for Hybrid AI-Assisted Operational Procedures via VFS Harmonization. Its subject is the reproducibility of a procedure that crosses computational interfaces, including AI-assisted steps. A replay record needs to identify the inputs and representation context behind each decision.

Replay and alert quality are different achievements. A procedure may be perfectly repeatable and still classify an event incorrectly. Conversely, an informative anomaly detector may depend on an external model or service whose response cannot be recreated exactly. The two papers should therefore retain separate guarantees.

THE NEXT RESEARCH STEP

A concrete next case study would replay a recorded encoding transition, showing the raw measurements, both formats, the translation, and the resulting alert decision. For BRIDGE, the corresponding artifact is a replay whose equivalence criterion is explicitly stated.

TEACHING CONSTRUCTION (not project measurements)

Record a versioned measurement x, its encoding Eᵥ, the analysis rule, and the decision environment. Compare outputs through a declared equivalence test. An adaptive encoding needs a cross-version comparison rule before geometric displacement can be interpreted as system change.

x → Eᵥ(x) → score → decision → replay record

v: Encoder and schema version, including scaling and dimensionality.

Replay equivalence: Byte equality, field equality, or a specified behavioral equivalence.

Decision validity: Whether the decision meets a task criterion; independent of repeatability.

Suppose a sensor changes its reported temperature from 20 °C to 68 °F at a software update. A raw numeric jump is large; a unit-aware comparison is zero. The example shows why representation changes must be recorded alongside telemetry.

SOURCE RECORD

AMC: foundational paper and NASA origins: https://www.davisgeometric.com/research/papers/amc.html

AMC-Edge Rev D: precision adaptation: https://www.davisgeometric.com/research/papers/amc-edge.html

USAspending / IDAS: investigative data provenance: https://www.davisgeometric.com/research/papers/usaspending.html

Adaptive Embedding Format Switching and Semantic Translation in High-Volume Telemetry Systems: https://doi.org/10.5281/zenodo.17211403

BRIDGE: Deterministic Replay for Hybrid AI-Assisted Operational Procedures via VFS Harmonization: https://doi.org/10.5281/zenodo.17313876

Source versions: https://www.davisgeometric.com/research/account-source-records.json

No new experiments are reported in this research note.

Illustrated study: https://www.davisgeometric.com/research/figures/draft-operations.html