SAMAYUKTAM PROTOCOL SUITE PCMP v1.0 · SPCMP v1.0 · CAGM · FADS · CLOUD · INFERENCE ⬤ SYSTEM LIVE DETERMINISTIC INFERENCE LAYER MERKLE-ANCHORED · BIT-IDENTICAL · ZERO-OVERHEAD · CROSS-HARDWARE ROOT: 0x8dc860c7…
PATENT PENDING — PROVISIONAL FILED · PCT WINDOW: DECEMBER 2026 Samayuktam / SPCMP · All protocols, methods & implementations are protected intellectual property of Swapnopam Mitra. Unauthorized commercial use is prohibited.
⚖ PATENT PENDING PROVISIONAL FILED 2026 PCT WINDOW: DEC 2026 SOLE INVENTOR: SWAPNOPAM MITRA LICENSING INQUIRIES ONLY — NO UNSOLICITED ACCESS swapnopammitra1308@gmail.com ⚖ PATENT PENDING PROVISIONAL FILED 2026 PCT WINDOW: DEC 2026 SOLE INVENTOR: SWAPNOPAM MITRA LICENSING INQUIRIES ONLY — NO UNSOLICITED ACCESS swapnopammitra1308@gmail.com
███████╗ █████╗ ███╗   ███╗ █████╗ ██╗   ██╗██╗   ██╗██╗  ██╗████████╗ █████╗ ███╗   ███╗
██╔════╝██╔══██╗████╗ ████║██╔══██╗╚██╗ ██╔╝██║   ██║██║ ██╔╝╚══██╔══╝██╔══██╗████╗ ████║
███████╗███████║██╔████╔██║███████║ ╚████╔╝ ██║   ██║█████╔╝    ██║   ███████║██╔████╔██║
╚════██║██╔══██║██║╚██╔╝██║██╔══██║  ╚██╔╝  ██║   ██║██╔═██╗    ██║   ██╔══██║██║╚██╔╝██║
███████║██║  ██║██║ ╚═╝ ██║██║  ██║   ██║   ╚██████╔╝██║  ██╗   ██║   ██║  ██║██║ ╚═╝ ██║
╚══════╝╚═╝  ╚═╝╚═╝     ╚═╝╚═╝  ╚═╝   ╚═╝    ╚═════╝ ╚═╝  ╚═╝   ╚═╝   ╚═╝  ╚═╝╚═╝     ╚═╝
Protocol Suite for Verifiable Inference | v1.0 FROZEN | 4 LIVE SPACES ↗ | UNSOLVED GLOBALLY
CORE PCMP Predictive Canonical Monotonic Permutation
PROFILE SPCMP Samayuktam Profile — IEEE-754 CVD
LAYER CAGM Cross-Architecture Graph Map
SEARCH FADS Formal AI Deterministic Search
CLOUD S-CLOUD Deterministic Inference Infrastructure
FINAL LAYER INFERENCE Samayuktam Inference — Runtime Canonicalization

PCMP

Predictive Canonical
Monotonic Permutation
Samayuktam v1.0

Floating-point inference is non-deterministic across hardware. PCMP terminates that. Every logit stream canonicalized, committed, and reconstructible — byte-identical across CPU, GPU, and batch. No approximations. No overrides. No excuses.

samayuktam@verifier:~$ pcmp-audit STANDBY
$
⎇ GitHub ↗ Live Demo
01 Float Stream Raw IEEE-754
02 SPCMP Preprocess CVD canonicalize
03 PCMP Canonical Sort · predict · compress
04 Merkle Root 0x8dc860c7…
NETWORK STATUS ⬤ ONLINE
0 logits/sec
0 hashes committed
100% match rate
CANONICAL MERKLE ROOT 0x8dc860c7a3f21e90
10.7M float32 logits
per inference
5 protocol layers
PCMP · SPCMP · CAGM · FADS · INFERENCE
100% bit-identical
reconstruction
0 hardware-locked
artifacts
verifier
portability
जटाजूटसमायुक्तां — Devi Durga Stotra Samayuktam — joined, united, brought into one. One canonical representation. One truth. Across every chip that runs inference.

What Samayuktam Is

Samayuktam solves a problem that has quietly undermined every verifiable AI system built to date: the same neural network, run on different hardware, produces different outputs at the bit level. This makes cryptographic verification of AI inference and training mathematically impossible — until now.

Samayuktam is a deterministic verification protocol that canonicalizes floating-point representations across CPUs, GPUs, and TPUs, producing a hardware-agnostic cryptographic fingerprint for any neural network's inference or training run. The same model, the same input, the same hash — regardless of which chip ran it.

81assertions · 0failures · x86NVIDIA GPUTPU v5e

Verified across GPT-2, BERT-base, and LLaMA-style architectures.

Technical Specification

Samayuktam is a deterministic inference verification protocol for neural networks running on heterogeneous hardware. At its core, PCMP (Predictive Canonical Monotonic Permutation) defines a bijective, order-preserving mapping of IEEE-754 float32 values into a uint32 space, enabling a canonical total ordering that is bit-identical across any CPU, GPU, or TPU.

SPCMP, its strict profile, adds a mandatory preprocessing pass that collapses all NaN payload variants to a single canonical quiet NaN and all negative zeros to positive zero, eliminating the four primary sources of cross-hardware floating-point divergence: FMA sub-LSB jitter, signed-zero contamination, NaN payload variants, and near-bucket-edge exponent shifts.

The Samayuktam Inference Engine wraps this canonicalization around every layer boundary of a running neural network, producing a 32-bit FNV-1a logit fingerprint provably bit-identical regardless of whether the model ran on an NVIDIA H100, a Google TPU v5e, or an x86 CPU.

What Samayuktam does not do
Does not preserve NaN payloads Does not guarantee bit-identity with raw producer float output Does not define or control the arithmetic of layer functions Does not operate on float64 or mixed-precision tensors Not a compression or storage format for production weights
DETERMINISM AUDIT REPORT Inference Layer v1 Generated 2026-05-01 · TPU v5e · Python 3.12.13 · JAX 0.7.2
100% SCORE
S1SPCMP/PCMP Core Correctness16P / 0FPASS
S2Bucket Boundary Isolation10P / 0FPASS
S3GPT-2 Small — GPU vs GPU Simulation5P / 0FPASS
S4BERT-base Bimodal Distribution3P / 0FPASS
S5LLaMA-style Log-Normal Wide Range7P / 0FPASS
S6Multi-Layer Depth Stability (12/24/48L)6P / 0FPASS
S7Hash Collision Resistance3P / 0FPASS
S8IEEE-754 Special Value Completeness25P / 0FPASS
S9Cross-Machine Reference Artefact4P / 0FPASS
S10Throughput & Production Viability2P / 0FPASS

Floating-Point Inference Is Fundamentally Non-Deterministic

Every major AI lab ships models that produce different numerical outputs on different hardware. Not approximately different. Actually different. CPU vs GPU. Driver A vs Driver B. Batch of 1 vs batch of 32. The same model, the same weights, different bits.

This is not a bug. IEEE-754 floating-point arithmetic is non-associative. Operation reordering, FMA fusion, subnormal flushing, signed-zero behavior, NaN propagation — all of these produce legitimate but divergent results. No existing standard resolves this for inference-scale logit streams.

CONSEQUENCE Model audit, reproducibility, on-chain commitment, and cross-hardware verification are impossible without a canonical layer. PCMP is that layer.
LIVE DIVERGENCE SIMULATION DETECTING…
M4 MAX 0x3f800000
H100 SXM5 0x3f800001
PCMP CANONICAL 0x8dc860c7…

Four-Layer Verification Stack

CORE PROTOCOL
PCMP
Predictive Canonical Monotonic Permutation
Defines canonical ordering, permutation, prediction transform, compression, and Merkle hashing for arbitrary float32 sequences. The invariant foundation. Platform-agnostic. Frozen.
Deterministic Invertible Merkle-Anchored
Spec v1.0 · FROZEN SEMANTICS
STRICT PROFILE
SPCMP
Samayuktam Profile for PCMP v1
Mandatory preprocessing phase transforming raw IEEE-754 inputs into the Canonical Value Domain (CVD) before PCMP processing. NaN canonicalization. Signed-zero elimination. Idempotent. Fail-closed.
CVD-Enforced Idempotent Fail-Closed
SPCMP Profile ID: 0x01
GRAPH LAYER
CAGM
Cross-Architecture Graph Map
Extends PCMP canonicalization to computational graph structures. Maps model execution graphs across heterogeneous hardware architectures. Enables graph-level commitment and cross-device proof.
Graph-Native Cross-Architecture Commitment-Ready
CAGM Spec v1.0

Three-Phase Canonical Pipeline

01
CANONICALIZE
SPCMP Preprocessing + PCMP Mapping
Raw IEEE-754 float32 streams are transformed into a hardware-agnostic Canonical Value Domain. Non-deterministic bit patterns — NaN payloads, signed zeros, FMA artifacts — are resolved to a single deterministic representation. The output is identical regardless of CPU, GPU, driver, or batch configuration.
Deterministic Idempotent CVD-Enforced
02
COMMIT
Predictive Encoding + Merkle Anchoring
The canonical sequence is compressed via predictive delta encoding and anchored to a Merkle tree over fixed-size chunks of the compressed byte stream. The resulting root hash is the artifact identity — a compact, portable, tamper-evident commitment to the full inference output.
Merkle-Anchored Compact Tamper-Evident
03
RECONSTRUCT
10-Step Verified Decode
Any verifier — on any hardware — can independently reconstruct and validate the canonical artifact against the committed root. The 10-step SPCMP verification protocol is fail-closed: any deviation from the canonical specification causes immediate hard failure. No partial verification. No silent passes.
Fail-Closed Cross-Hardware Portable
FAIL SEMANTICS Any step violation causes immediate hard failure. No partial verification passes.
VERSION POLICY Version 1 frozen forever. Modifications require a new profile identifier.
ADVERSARIAL BOUND No adversary operating on SPCMP artifacts can create equivalent artifacts with different hashes.
TECHNICAL REVIEW Full specification and implementation details available under NDA. Contact for partnership inquiry.

Evaluated on Real Transformer Workloads

Tested on real transformer inference across heterogeneous hardware execution environments. Not synthetic benchmarks. Not toy models. Production-scale inference.

MODEL BLIP (Bootstrapping Language-Image Pretraining)
SCALE 10.7M float32 logits per inference
HARDWARE Apple M4 MAX · NVIDIA H100 SXM5 · x86-64 CPU
✓ CPU deterministic execution
✓ CUDA GPU deterministic execution
✓ Batched inference (32 images)
✓ Byte-level reconstruction via cmp
✓ SPCMP profile header validation
✓ Cross-driver version stability
RESULT Bit-identical canonical artifacts across CPU, GPU, and batch. Zero divergence post-canonicalization.
↗ Inspect Validation Scripts
32-image batch · all cells verified · 0 divergences
■ CANONICAL MATCH ■ PROCESSING

What PCMP + SPCMP Terminates

Cross-platform float divergence
Non-reproducible inference outputs
Unverifiable model commitments
Silent kernel drift accumulation
Hardware-locked artifact proofs
NaN payload side-channels
Signed-zero behavioral divergence
Driver-version hash instability
FMA-fusion ordering artifacts
Subnormal flushing divergence
Batch-size dependent outputs
Unauditable AI inference chains

10-Step Verification Algorithm

01
Validate PCMP header and version Version mismatch causes immediate hard failure.
PASS
02
Validate SPCMP profile indicator Profile ID must equal 0x01. No other value accepted.
PASS
03
Decompress payload LZ decompression of predictive byte stream.
PASS
04
Validate element count Count mismatch indicates malformed or truncated payload.
PASS
05
Compute hash over decompressed predictive data Merkle tree over fixed-size chunks. Chunk size inherited from PCMP v1.
PASS
06
Validate hash Compare computed root against declared root. Single-bit mismatch = failure.
PASS
07
Apply inverse predictor Reconstruct mapped uint32 sequence from delta encoding.
PASS
08
Validate canonical ordering Ordering validated on inverse-predicted sequence prior to inverse permutation.
PASS
09
Apply inverse permutation Restore original positional ordering from permutation encoding.
PASS
10
Validate all values lie within CVD Any value outside Canonical Value Domain = hard failure. Detects false SPCMP claims.
PASS

Samayuktam Protocol Suite — Running Now

Four active Hugging Face Spaces. Open source. Verifiable. No accounts required.

CORE LIVE
PCMP
Samayuktam
Predictive Canonical Monotonic Permutation
The primary PCMP verification space. Upload a float32 artifact and receive a canonical Merkle commitment. Demonstrates bit-identical verification across heterogeneous hardware in real time. The working implementation of everything on this page.
▸ Live PCMP encoding ▸ Merkle root computation ▸ SPCMP profile verification ▸ Cross-hardware diff display
SEARCH LIVE
FADS
Samayuktam System
Formal AI Deterministic Semantic Search
Deterministic semantic search engine built on PCMP canonicalization. Uses embed.py and the full Samayuktam stack to make AI model outputs deterministic end-to-end — from embedding through retrieval. Search results are bit-reproducible across any hardware that runs the verifier.
▸ Deterministic embeddings ▸ Canonical semantic search ▸ Hardware-agnostic retrieval ▸ embed.py reference impl
INFRASTRUCTURE LIVE
CLOUD
Samayuktam Cloud
Deterministic Inference Cloud Infrastructure
Cloud-native deployment layer for the Samayuktam protocol suite. Exposes PCMP and SPCMP as hosted verification endpoints. Anchors inference runs to Merkle-committed canonical artifacts accessible for audit without downloading the full model stack.
▸ Hosted verification API ▸ Cloud-scale Merkle anchoring ▸ Audit-ready artifact storage ▸ Zero local setup required
TRAINER LIVE
CAGM
CAGM
Cross-Architecture Graph Map — ML Model Trainer
The CAGM ML model training interface. Maps computational graphs across heterogeneous hardware architectures for deterministic training runs. Enables graph-level canonical commitment so model training artifacts are bit-reproducible and independently auditable.
▸ Graph-level canonicalization ▸ Cross-architecture training ▸ Deterministic weight commits ▸ Hardware-portable proofs

verify.py — The Canonical Verifier

Production Python implementation. Runs locally. No dependencies on this page. Pure protocol.

verify.py Samayuktam PCMP + SPCMP Reference Verifier · v1.0
Python 3.10+ Runtime
zstandard Only dep
Fail-closed Error mode
MIT License

Standalone reference implementation of the full SPCMP verification algorithm. Reads a .pcmp artifact, performs all 10 verification steps, validates the Canonical Value Domain, checks the Merkle root, and emits structured output. This is the ground-truth verifier — what all other implementations must match.

INSTALL pip install zstandard
RUN python verify.py artifact.pcmp
DETAILED python verify.py --info artifact.pcmp
JSON python verify.py --json artifact.pcmp
VERIFICATION OUTPUT DEMO

            
10-STEP VERIFICATION CHECKLIST
01 PCMP header · version
02 SPCMP profile 0x01
03 Decompress payload
04 Element count
05 Hash over predictive data
06 Hash validation
07 Inverse predictor
08 Canonical ordering
09 Inverse permutation
10 CVD validation

Samayuktam Inference

The inference runtime is the terminal enforcement boundary of the Samayuktam protocol stack. It applies SPCMP canonicalization directly inside the forward pass — between every layer, at every activation boundary.

Every activation tensor is canonicalized before it enters a layer and corrected after it exits. The logit stream produced at the final layer is guaranteed to hash identically across any compliant hardware, regardless of FMA fusion, NaN payload, signed-zero behavior, or driver-level float rounding.

CANONICALIZE Pre-layer SPCMP pass on every activation
CORRECT Post-layer correction pass enforces CVD
HASH FNV-1a logit hash — bit-identical cross-hardware
Samayuktam Inference Space
ACTIVATION TENSOR
CANONICALIZESPCMP · CVD enforcement
LAYER NModel forward pass
CORRECTION PASSBucket-snap · idempotent
REPEAT × NUM_LAYERS
LOGIT HASH0x8dc860c7…
BIT-IDENTICAL · CROSS-HARDWARE

Inference Canonicalization Simulator

PRESETS:

Protocol Directives