Live · NIST FIPS 203 / 204 conformant · ML-DSA-65 sig 3309 B · ML-KEM-768 ct 1088 B
smsh: semantic message shorthand

A policy-preserving prompt compiler with attested fidelity.

smsh segments enterprise prompts by semantic region, freezes behavioral invariants across turn boundaries, attests fidelity at the answer level, and amortizes post-quantum signatures across an inference session. We bill net of savings. The customer never pays more than the compression saves.

Read this page in 3 minutes
1 · See it work
Run the live demo. Free, no auth. Same endpoint your app would call.
2 · Check the numbers
Read the honest ratio bands. Measured tonight. We don't round up.
3 · Take the offer
Send 10K real prompts under NDA. Net-savings billing. Zero if we don't beat your status quo.
Category
Prompt compiler
Not a compression service. Region-segmented, tokenizer-aware, multi-turn invariant-preserving.
Attestation
Answer-level
7B fine-tuned NLI judge, async, calibrated to under 1% false-pass. Flag triggers automatic billing credit.
PQ posture
Session-amortized
ML-DSA-65 outer per session, Ed25519 inner per inference. Per-call PQ overhead under 1% at N=100.
Bucket 1 · Try it

Free demo. No auth, no card. Paste a prompt and read the receipt.

Live demo · free Tier0

Compress a prompt. Read the receipt.

No auth, no card. Paste up to 4,000 characters. We return the compressed payload, the ratio, dictionary hits, and the latency. The exact same endpoint your application calls.

POST /v1/smsh/compress/demo · Tier0 · 4,000 char cap
Receipt
Click "Compress now" to see a live receipt.
Bucket 2 · Trust and proof

Fidelity contract. FIPS-conformant PQ. Honest ratio bands measured tonight. Kill conditions we ship instead of press releases.

The five-layer fidelity contract

What makes the receipt enterprise-grade.

Not a hash. Not a vendor claim. A bound chain from compressed prompt to verified answer to post-quantum signature. Five layers, each one verifiable independently.

  1. Tokenizer-aware subword masking

    The pruner never splits a BPE token. Mask is constructed via HuggingFace Tokenizers' offset_mapping; sub-token positions are forced non-prunable.

  2. Region segmentation

    Instruction, schema, retrieved context, user input, examples, output spec each get their own compression policy. Schema fields and output specs are pass-through by default.

  3. Invariant freezing across turns

    Negations, numbers, named entities, schema fields, tool names, refusal triggers are non-prunable in turn one and stay non-prunable through turn N. Multi-turn cap default 2.5x.

  4. Answer-level fidelity judge

    7B fine-tuned NLI, async on a sample of compressed inferences, calibrated to under 1% false-pass on held-out probes. Flag triggers automatic customer billing credit.

  5. Signed receipt with session PQ amortization

    Ed25519 inner per inference (64 bytes). ML-DSA-65 outer over Merkle root of inference receipts per session (3309 bytes amortized). ML-KEM-768 wraps the session key (1088 bytes ct).

PQ self-test · live, FIPS-conformant

NIST conformance, served from production.

Hit the self-test endpoint and read the byte sizes and latencies for ML-DSA-65 and ML-KEM-768 against the FIPS-specified values.

FIPS 204 · ML-DSA-65
public key
1,952 bytes ✓
signature
3,309 bytes ✓
OID
2.16.840.1.101.3.4.3.18
sign latency
~140 ms (production)
verify latency
~10 ms (production)
FIPS 203 · ML-KEM-768
public key
1,184 bytes ✓
ciphertext
1,088 bytes ✓
shared secret
32 bytes ✓
OID
2.16.840.1.101.3.4.4.2
encap latency
~4 ms (production)
# Live FIPS-conformance self-test, no auth, returns byte sizes + latencies
curl https://receipts.thehiveryiq.com/v1/smsh/pq/selftest

# Fetch a fresh sample ML-DSA-65 keypair for demo verification
curl https://receipts.thehiveryiq.com/v1/smsh/pq/keys/sample
Honest ratio bands

What we publish. What we refuse to publish.

Single-number compression claims are how vendors get caught. We publish CDF by prompt length × task × turn count, and we'll send you the bench repo. Below are the bands we stand behind.

WorkloadBandBar
Broad single-prompt, task-preserving2x to 5xLLMLingua-2 sidecar target at ≤1.5% accuracy loss (v1.5)
Repetitive long-context tenant traffic (batched RAG, CASB templates)4x to 8xTenant dictionary + region segmentation engaged
Multi-turn (3+ turns) with invariant freeze1.5x to 2.5xHard cap at 2.5x; below 1.5x if fidelity drop > 10%
Short prompts (< 150 tokens)pass-throughRouter skips compression. No overhead.
"8x broadly"neverWe don't claim this. Anyone who does is selling the lie.
"100% HumanEval"neverWe publish range, not headline.
"Sub-millisecond compression"neverProduction latency is 1 to 15 ms. The lie is the qualifier.
Measured tonight · smsh-v1 · deterministic scorer · live
WorkloadRatioLatencyInvariants
RAG repetitive (paraphrased context, 237 t)1.58x4 msall preserved
RAG paraphrased (3 facets of one topic, 156 t)1.23x3 msall preserved
Verbose-fillers (customer ticket rewrite, 159 t)1.14x2 msall preserved
CASB-mixed (DLP policy + entity-heavy, 207 t)1.04x2 msPII, refusal, deontics preserved
Schema-heavy (JSON output spec, 255 t)pass-through2 msschema frozen
Chat short (< 80 t)pass-through0.1 msno overhead
Reproduce: POST /v1/smsh/compress/demo with mode: "smsh-v1". All measurements from production at receipts.thehiveryiq.com. We don't round up.
v2 · distilled importance head · in training tonight

v1 is the deterministic scaffold. v2 is the engine. A 4-layer transformer (≈20M params, int8 ONNX, ≈30 MB on disk), trained on a corpus of CASB, RAG, agent-context, and conversation prompts labeled by LLMLingua-2, then conditioned on the region type emitted by v1's segmenter. The teacher labels everything. The student runs on CPU.

v2 targetCommitmentWhy we can promise it
Mean compression across CASB / RAG / agent3.0 to 3.5xLLMLingua-2 teacher hits 2.5 to 3.5x. Region-conditioning lifts it further on workloads it was never trained for.
Warm CPU latency, 1K tokens< 50 msThe 4-layer student is 30x smaller than the BERT-base teacher. ONNX int8 halves it again.
Invariant preservation (PII, numbers, deontics)≥ 99.5%The hard veto from v1's invariant set is added to the student's keep mask. The teacher can't override it.
Statustraining now50K-prompt typed corpus generated. Teacher labeling on GPU spot starts tonight. Student trains over the next 24 to 72 hours. When it clears the bars above, it ships. When it doesn't, it doesn't.
No ship date is published until the model clears all three bars on staging. v1 stays in production as the deterministic fallback for any workload v2 declines (invariant-only regions, schema-heavy, pass-through short chat). v2 replaces v1.5, the LLMLingua-2 sidecar experiment. Honest result: 7s warm CPU latency, unshippable, killed.
Six kill conditions · hard-coded

We flip the switch. We don't ship the lie.

If any of these fires during the v1 build, we don't claim what we can't deliver. We label honestly or we don't ship the feature.

  1. LLMLingua-2 baseline below 2x at ≤1.5% accuracy loss → kill v1, rebuild.
  2. Judge cannot calibrate <1% false-pass in 2 weeks → ship without fidelity guarantee, labeled best-effort.
  3. tiktoken-boundary masking cannot ship cleanly in 2 weeks → single-endpoint v1 only.
  4. Multi-turn fidelity drop > 10% at 2.5x cap → drop cap to 1.5x or remove multi-turn from v1.
  5. Domain teacher bias > 5% on CASB prompts → switch to Mistral-7B teacher.
  6. PQ overhead > 50% of savings on any prompt class → symmetric-only for that class, labeled symmetric-attested.
Bucket 3 · Take action

Endpoints to call. The offer (we only bill what we save). Questions answered.

Live endpoints

The HTTP shape, for any language.

POST/v1/smsh/compress/demoTier0 · live POST/v1/smsh/compressTier1 · live POST/v1/smsh/learned_pruneTier2 · live POST/v1/smsh/judgeTier1 · live GET/v1/smsh/pq/selftestTier0 · live GET/v1/smsh/pq/keys/sampleTier0 · live POST/v1/smsh/pq/signTier1 · live POST/v1/smsh/pq/verifyTier0 · live GET/v1/smsh/dictionary/{tenant}Tier1 · live
# Free Tier0 compression demo, returns ratio + receipt
curl -X POST https://receipts.thehiveryiq.com/v1/smsh/compress/demo \
  -H "Content-Type: application/json" \
  -d '{"text": "<your prompt>", "mode": "smsh-v1"}'

# Verify FIPS conformance live
curl https://receipts.thehiveryiq.com/v1/smsh/pq/selftest
The offer

Send us 10,000 real prompts.

Under NDA. We mine your tenant dictionary. We measure your actual ratio on your actual traffic. If we beat your status quo cost-per-inference at no worse than 1.5% accuracy delta, you pay our net-savings rate. If we don't, you don't.

No deck. No POC fee. The deployment is live, the endpoints respond, the PQ self-test passes FIPS 203/204 conformance today. Read the receipt yourself.

Reach out via X or hit the demo endpoint above. The bench repo is open-source.

FAQ

What people ask first.

Are you another LLMLingua wrapper?
No. LLMLingua is our v1 baseline. We ship it honestly in weeks one and two so customers have a number on day fifteen. The wedge is the layers above: region segmentation, invariant freezing, answer-level fidelity attestation, session-amortized PQ, net-savings billing. We don't custom-train a compressor unless our bench shows it beats LLMLingua-2.
Why session-amortized PQ instead of per-inference?
A single ML-DSA-65 signature is 3,309 bytes. On a compressed 500-token prompt, signing per-inference would make the signature bigger than the prompt. Sessions amortize the cost: at N=100 inferences, per-call PQ overhead is ≈63 bytes, under 1% of an 8KB compressed prompt.
What happens when the fidelity judge flags a compression?
The customer is credited the compression service fee for that inference automatically. We hold ourselves harmless only when we deliver. This is the net-savings billing meter.
Why don't you publish a single headline ratio?
Because no honest single number exists. Compression ratio depends on prompt length, task class, turn count, tenant corpus. We publish CDF and stand behind bands. Vendors who publish "8x broadly" are selling the lie.
Is the PQ backend liboqs?
Production runs a pure-Python reference (dilithium_py + kyber_py) right now. It's a predictable build with no system-library dependency. We'll move to the liboqs C backend when we need the speedup. The API stays the same either way, and FIPS sizes are conformant on every response.
What's the privacy posture for tenant dictionaries?
Mined under NDA from customer-supplied corpora. Signed with our key, versioned. Each compressed prompt receipt references the dictionary version. Dictionaries do not leave the tenant boundary.