Lossless Mode
Exact compression for regulated AI.
When regulators ask “does your compression alter model behaviour?”, the answer must be no. Revyzor's lossless mode introduces zero change to the model's internal state. The same bytes go in and come out: byte-for-byte checked at capture.
The regulatory problem.
Audit trail integrity
Financial regulators (FCA, SEC, MAS, BaFin) expect AI-assisted decisions to be explained and evidenced. If an AI flagged a transaction as suspicious, the bank should be able to show, months later, exactly what the model was working with when it decided. Any compression that alters the model's internal state, even slightly, means the record is no longer exact.
Model risk management (SR 11-7 / SS1/23)
Banks deploying AI must demonstrate that infrastructure changes do not alter model behaviour. Enabling compression on an LLM is an infrastructure change. If that compression is lossy, the bank must re-validate the entire model, a process that can take months and cost hundreds of thousands. Lossless compression changes nothing: the state the model computes with is mathematically identical, so there is nothing new to validate.
GDPR right to explanation
Under GDPR, individuals are entitled to meaningful information about automated decisions. If the AI's internal state has been altered by lossy compression, the state you reload today is not the state the decision was computed from. That matters for resuming the session, not for evidencing it: the signed transcript evidences the decision. Lossless compression means the preserved state is exactly what the model relied on, at any point in the future.
Healthcare record integrity (HIPAA / MDR)
AI-assisted clinical decisions must be traceable for patient safety and liability. If a model recommended a treatment plan, the record of what it relied on must be exact. Lossy compression introduces approximation that is unacceptable in clinical settings.
What exact means.
Cosine similarity
Original KV cache vs reconstructed KV cache after full roundtrip. Not 0.9999. Exactly 1.0000.
Compression ratio
31% memory savings with zero risk. Free capacity on every GPU. No accuracy trade-off to evaluate.
Difference
The same bytes go in and come out. Not approximately the same. Identical. Byte for byte.
What this means for banks.
Deploy compression without re-validation
Lossless means the model computes with exactly the same state. Your existing model validation, your bias testing, your fairness assessments; all still valid. Enable compression on Monday, save 31% GPU memory, with a one-line change to the model-risk file.
Archive and reload AI decisions
Persist the compressed reasoning state alongside the decision. Months later, reload it and inspect exactly what the model was working with, byte-for-byte what was captured. This is the audit trail regulators want. Lossy methods cannot offer it.
Persistent conversation memory for compliance
SAR investigations, client advisory sessions, compliance reviews: conversations that must be preserved in full. Lossless compression stores that history in 31% fewer bytes than raw, and what comes back off disk is bit-identical to what the model held. The saving is modest; the point is that the archived state is the state, not an approximation of it.
Portable across your estate
Move the compressed state between servers, data centres, or disaster recovery sites. The state you reload is exactly the state you saved, with no drift, no environmental dependencies.
No alternative offers this.
| Revyzor Lossless | FP8 | 4-bit quant | |
|---|---|---|---|
| State preserved exactly | Yes | No | No |
| Record matches what ran | Yes | Approximate | Approximate |
| Re-validation required | No | Yes | Yes |
| Audit trail archival | Exact record | Approximate | Approximate |
| Answers that change vs baseline | None | 9.03% | Not measured |
| Memory savings | 31% (1.444x) | 50% (2x) | ~75% (4x) |
FP8 is accurate, and we will say so before you ask. On standard benchmarks it scores at parity with the uncompressed model, and in our own long-context retrieval tests it passed at every length we ran. What it changes is which answer you get. Across 25,282 questions, switching an 8B model to an FP8 cache moved the headline accuracy score by 0.79 points while 9.03% of the individual answers came back different. A headline score is an average over a population. A declined application or a flagged transaction is one case, and it is the case you are asked about later.
Re-running the case later does not settle which answer you got, and it is worth being precise about why, because this is narrower than it is usually made out to be. On a default serving stack the same model, prompt and settings do not reliably produce the same output: requests are batched together to keep the GPU busy, batch size moves with load, and the order of floating-point operations moves with it. Temperature zero does not fix that. A batch-invariant mode does, vLLM ships one, and we have measured it ourselves removing the effect completely, on 600 of 600 reasoning steps. Anyone can turn it on. Two things it does not touch: it is off by default, so it describes almost no deployment running today, and it only works forwards. Enabling it next year does nothing for a decision taken last year, on weights and an engine build you have since replaced. So a lossy copy of the state is not a rough version of something you can recreate later. It is the only copy, and it is already approximate.
We would not recommend moving to FP8 for a decision you have to answer for. But if you have already certified on it, FP8 is your reference and none of this asks you to turn it off. Revyzor runs on top: on a 7B model at 64K tokens, FP8 alone holds 28,672 bytes per token and FP8 with Revyzor holds 23,728, reloading to exactly the state the model ran on.
Compression your compliance team will approve.
Talk to Us About Compliancesales@revyzor.com