Internet-Draft UPIP March 2026
van de Meent & AI Expires 30 September 2026 [Page]
Workgroup:
Internet Engineering Task Force
Internet-Draft:
draft-vandemeent-upip-process-integrity-01
Published:
Intended Status:
Informational
Expires:
Authors:
J. van de Meent
Humotica
R. AI
Humotica

UPIP: Universal Process Integrity Protocol with Fork Tokens for Multi-Actor Continuation

Abstract

This document defines UPIP (Universal Process Integrity Protocol), a five-layer protocol for capturing, verifying, and reproducing computational processes across machines, actors, and trust domains. UPIP defines a cryptographic hash chain over five layers: STATE (input), DEPS (dependencies), PROCESS (execution), RESULT (output), and VERIFY (cross- machine proof). The stack hash chains these layers, ensuring that modification of any component is detectable.

This document also defines Fork Tokens, a continuation mechanism for multi-actor process handoff. Fork tokens freeze the UPIP stack at a specific point and transfer it to another actor with cryptographic chain of custody. The receiving actor can verify what was handed off, validate capabilities, and continue the process with full provenance.

UPIP integrates with TIBET [TIBET] for provenance tokens, JIS [JIS] for actor identity, and is transport-agnostic with JSON as the baseline serialization.

Status of This Memo

This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.

Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.

Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."

This Internet-Draft will expire on 30 September 2026.

▲

Table of Contents

1. Introduction

Distributed computing increasingly involves heterogeneous actors: human operators, AI agents, automated pipelines, edge devices, and cloud services. When a process moves between actors -- from one machine to another, from an AI to a human for review, from a drone to a command station -- the integrity of the process state must be verifiable at every handoff point.

UPIP fills this gap with two complementary mechanisms:

  1. The UPIP Stack: a five-layer bundle capturing everything needed to reproduce a process, with a single stack hash that invalidates if any layer is modified.
  2. Fork Tokens: a continuation mechanism that freezes the stack state and transfers it to another actor with cryptographic proof of what was handed off, who handed it off, why, and what capabilities are required to continue.

1.1. Problem Statement

Existing solutions address parts of process integrity:

  • Version control (git) tracks code state but not execution
  • Container images (OCI) capture environment but not intent
  • CI/CD pipelines orchestrate execution but provide no cross-machine reproducibility proof
  • Package managers record dependencies but not their usage context

None provide a unified, self-verifying bundle that captures the complete execution context with cryptographic chain of custody across actor boundaries.

1.2. Design Principles

EVIDENCE OVER ENFORCEMENT:
UPIP proves what happened. Fork validation failures do not block execution; they are recorded as evidence. This reflects the reality that enforcement can be circumvented but evidence cannot be un-recorded.
HASH CHAIN INTEGRITY:
Every layer is independently hashed. The stack hash chains them. Fork hashes chain into the fork chain. Tampering with any component invalidates the chain.
ACTOR AGNOSTICISM:
Actors may be human operators, AI agents, automated scripts, IoT devices, or any computational entity. The protocol makes no assumption about actor type.
TRANSPORT AGNOSTICISM:
UPIP bundles are JSON documents. They can be transferred via file copy, HTTP API, message queue, I-Poll, or physical media.

2. Terminology

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.

Actor
An entity that creates, modifies, or continues a UPIP process. Actors may be human operators, AI agents (IDDs as defined in [JIS]), automated scripts, IoT devices, or any computational entity.
Airlock
An isolated execution environment (sandbox) where processes run before their results are applied to production state. The airlock captures all side effects without committing them.
Canonical Serialization
The deterministic JSON serialization used before hashing. Keys sorted lexicographically, no whitespace, UTF-8 encoding. Defined in Section 4.7.
Continuation Point
A reference to the specific position in the UPIP stack where the fork occurs, expressed as "L{layer}:{position}". Example: "L4:post_result" indicates the fork occurs after L4 RESULT has been captured.
Fork Token
A JSON document that freezes the UPIP stack state at a specific point and authorizes another actor to continue the process.
Fork Chain
An ordered list of fork token references maintained in the UPIP stack, providing a complete history of all handoffs.
Fork-Squared (Fork^2)
Parallel forking, where a single process is split into N independent sub-tasks distributed to N actors, each receiving a fork token of type "fragment".
IDD (Individual Device Derivative)
An AI agent with unique identity. Defined in the companion JIS specification [JIS].
Shadow-Run
Executing a process in the airlock to capture its effects without applying them. Used for fork validation.
Stack Hash
The SHA-256 hash computed over the concatenation of L1 through L4 layer hashes, prefixed with "upip:sha256:". This single hash represents the complete integrity of the UPIP bundle.
UPIP Stack (Bundle)
A JSON document containing all five UPIP layers plus metadata. Files use the ".upip.json" extension.

3. Protocol Overview

UPIP operates in two modes:

Single-Actor Mode (Capture-Run-Verify):

  1. CAPTURE: Record L1 (state) and L2 (deps)
  2. RUN: Execute the process (L3) in an airlock
  3. RESULT: Capture L4 (output, diff, hash)
  4. HASH: Compute stack_hash = SHA-256(L1 || L2 || L3 || L4)
  5. VERIFY: On another machine, reproduce and compare (L5)

Multi-Actor Mode (Fork-Resume):

  1. Actor A completes steps 1-4 (single-actor mode)
  2. Actor A creates a Fork Token from the UPIP stack
  3. Actor A delivers the fork token to Actor B
  4. Actor B validates the fork token hash
  5. Actor B checks capability requirements
  6. Actor B executes continuation in an airlock (shadow-run)
  7. Actor B creates a new UPIP stack linked to Actor A's via the fork chain
  8. Actor B sends ACK with resume_hash to Actor A 
+----------+     +---------+     +---------+     +---------+
| L1 STATE |---->| L2 DEPS |---->| L3 PROC |---->| L4 RSLT |
+----------+     +---------+     +---------+     +---------+
     |                |               |               |
     v                v               v               v
  state_hash       deps_hash      (intent)       result_hash
     |                |               |               |
     +-------+--------+-------+-------+
             |
             v
      stack_hash = SHA-256(L1 || L2 || L3 || L4)
             |
             v
       +------------+
       | Fork Token |---> Actor B ---> New UPIP Stack
       +------------+
             |
             v
        fork_chain: [{fork_id, parent_hash, ...}]
Figure 1: Process Flow Diagram

4. UPIP Stack Structure

A UPIP stack MUST be a [RFC8259] JSON object with the following top-level fields:

{
  "protocol": "UPIP",
  "version": "1.1",
  "title": "<human-readable description>",
  "created_by": "<actor identity (JIS format)>",
  "created_at": "<ISO-8601 timestamp>",
  "stack_hash": "upip:sha256:<hex>",
  "state": { },
  "deps": { },
  "process": { },
  "result": { },
  "verify": [ ],
  "fork_chain": [ ],
  "source_files": { }
}

4.1. L1 STATE - Input State Capture

L1 captures the complete input state before execution. The state_type field determines the capture method:

{
  "state_type": "git | files | image | empty",
  "state_hash": "<type>:<hash>",
  "captured_at": "<ISO-8601 timestamp>"
}

State Types:

git
Hash is the git commit SHA. MUST include git_remote and git_branch. state_hash prefix: "git:"
files
Hash is SHA-256 of the sorted file manifest. state_hash prefix: "files:"
image
Hash is the container image digest. state_hash prefix: "image:"
empty
No input state. state_hash: "empty:0"

For "git" type, additional fields:

  • git_remote: Repository URL
  • git_branch: Branch name
  • git_dirty: Boolean, true if uncommitted changes exist

For "files" type, additional fields:

  • file_count: Number of files captured
  • total_size: Total size in bytes
  • manifest: Optional array of {path, hash, size} objects

4.2. L2 DEPS - Dependency Snapshot

L2 captures the exact dependency set at execution time.

{
  "python_version": "<major.minor.patch>",
  "packages": { "<name>": "<version>" },
  "system_packages": [ "<name>=<version>" ],
  "deps_hash": "deps:sha256:<hex>",
  "captured_at": "<ISO-8601 timestamp>"
}

The deps_hash MUST be computed as SHA-256 of the sorted, deterministic serialization of all package name:version pairs.

While this specification uses Python as the reference implementation, L2 is language-agnostic. Other implementations MAY substitute appropriate dependency metadata for their runtime environment (e.g., Cargo.lock for Rust, go.sum for Go, package-lock.json for Node.js).

4.3. L3 PROCESS - Execution Definition

L3 defines what was executed and why. The "intent" field maps to TIBET ERACHTER [TIBET] and the "actor" field uses JIS identifier format [JIS].

{
  "command": [ "<arg0>", "<arg1>" ],
  "intent": "<human-readable purpose>",
  "actor": "<actor identity (JIS format)>",
  "env_vars": { "<key>": "<value>" },
  "working_dir": "<path>"
}

The command field MUST be an array of strings, not a shell command string. This prevents injection attacks and ensures deterministic execution.

The intent field MUST be a human-readable string describing WHY this process is being run. This serves as the ERACHTER (intent) component for TIBET integration.

The actor field MUST identify the entity that initiated the process using JIS identifier format. This may be a human operator, AI agent (IDD), or system service.

4.4. L4 RESULT - Output Capture

L4 captures the execution result.

{
  "success": true,
  "exit_code": 0,
  "stdout": "<captured stdout>",
  "stderr": "<captured stderr>",
  "result_hash": "sha256:<hex>",
  "files_changed": 3,
  "diff": "<unified diff of file changes>",
  "captured_at": "<ISO-8601 timestamp>"
}

The result_hash MUST be computed as SHA-256 of the concatenation of: exit_code (as string) + stdout + stderr.

If execution occurs in an airlock, the diff field SHOULD contain the unified diff of all file changes detected.

4.5. L5 VERIFY - Cross-Machine Proof

L5 records verification attempts when the UPIP stack is reproduced on another machine.

{
  "machine": "<hostname or identifier>",
  "verified_at": "<ISO-8601 timestamp>",
  "match": true,
  "environment": { "os": "linux", "arch": "x86_64" },
  "original_hash": "upip:sha256:<hex>",
  "reproduced_hash": "upip:sha256:<hex>"
}

The match field MUST be true only if reproduced_hash equals original_hash.

L5 is an array, allowing multiple verification records from different machines. Each verification is independent.

4.6. Stack Hash Computation

The stack hash MUST be computed as follows:

  1. Serialize each layer hash as a UTF-8 string: L1: state.state_hash, L2: deps.deps_hash, L3: SHA-256(canonical_json(process)), L4: result.result_hash
  2. Concatenate with pipe separator: L1 + "|" + L2 + "|" + L3 + "|" + L4
  3. Compute SHA-256 of the concatenated UTF-8 string
  4. Prefix with "upip:sha256:"

Result: "upip:sha256:4f2e8a..."

The canonical_json() function is defined in Section 4.7.

4.7. Canonical Serialization

Before hashing, JSON objects MUST be serialized to canonical form:

  1. All object keys sorted lexicographically by Unicode code point.
  2. No whitespace between tokens.
  3. Strings use only [RFC8259] escape sequences.
  4. Numbers use shortest representation without leading zeros.

This ensures deterministic hashing across implementations. The same canonical serialization is used in TIBET [TIBET] Section 5.1.

5. Fork Tokens

5.1. Fork Token Structure

A fork token MUST be a [RFC8259] JSON object with the following fields. Actor fields use JIS identifier format [JIS]:

{
  "fork_id": "fork-<uuid>",
  "parent_hash": "sha256:<hex>",
  "parent_stack_hash": "upip:sha256:<hex>",
  "continuation_point": "L<n>:<position>",
  "intent_snapshot": "<human-readable purpose>",
  "active_memory_hash": "sha256:<hex>",
  "memory_ref": "<path or URL to memory blob>",
  "fork_type": "script|ai_to_ai|human_to_ai|fragment",
  "actor_from": "<JIS actor identifier>",
  "actor_to": "<JIS actor identifier or *>",
  "actor_handoff": "<from> -> <to>",
  "capability_required": { },
  "forked_at": "<ISO-8601 timestamp>",
  "expires_at": "<ISO-8601 timestamp or empty>",
  "fork_hash": "fork:sha256:<hex>",
  "partial_layers": { },
  "metadata": { }
}

The actor_to field MAY be empty, indicating the fork is available to any capable actor. In this case, actor_handoff MUST use "*" as the target: "ActorA -> *".

5.2. Fork Types

script
The UPIP bundle IS the complete state. No external memory blob is needed. Used for CLI pipelines, CI/CD, and batch processing. The active_memory_hash is computed from the L1+L2+L3+L4 layer hashes.
ai_to_ai
The AI actor's context window is serialized as a binary blob (.blob file). The active_memory_hash is the SHA-256 of this blob. The memory_ref field SHOULD point to the blob's location.
human_to_ai
A human creates an intent document (natural language instructions) and delegates to an AI actor. The active_memory_hash is the SHA-256 of the intent document.
fragment
A parallel fork (Fork-Squared). The parent process is split into N sub-tasks, each receiving a fork token of type "fragment" with the specific portion assigned.

5.3. Fork Hash Computation

The fork hash MUST be computed as follows:

  1. Concatenate with pipe separator: fork_id + "|" + parent_hash + "|" + parent_stack_hash + "|" + continuation_point + "|" + intent_snapshot + "|" + active_memory_hash + "|" + actor_handoff + "|" + fork_type
  2. Compute SHA-256 of the concatenated string
  3. Prefix with "fork:sha256:"

Result: "fork:sha256:7d3f..."

This ensures that modifying ANY field invalidates the fork.

5.4. Active Memory Hash

The active_memory_hash captures cognitive or computational state at fork time.

For fork_type "script":
SHA-256(state_hash + "|" + deps_hash + "|" + process_intent + "|" + result_hash)
For fork_type "ai_to_ai":
SHA-256 of the serialized AI context window (.blob file).
For fork_type "human_to_ai":
SHA-256(contents of intent document)
For fork_type "fragment":
SHA-256(fragment specification)

This field is EVIDENCE, not a reproducibility guarantee. Exact reproduction of AI state is generally not achievable. The hash proves what the state WAS at fork time, enabling audit and comparison.

Implementations MUST NOT require exact memory reproduction for fork validation.

5.5. Capability Requirements

The capability_required field specifies what the resuming actor needs:

{
  "capability_required": {
    "deps": ["package>=version"],
    "gpu": true,
    "min_memory_gb": 16,
    "platform": "linux/amd64",
    "custom": { }
  }
}

On resume, the receiving actor SHOULD verify these requirements and record the result in the verification record. Missing capabilities MUST NOT prevent execution but MUST be recorded as evidence.

5.6. Fork Chain

The fork_chain field in the UPIP stack is an ordered array of fork token references:

{
  "fork_chain": [
    {
      "fork_id": "fork-abc123",
      "fork_hash": "fork:sha256:...",
      "actor_handoff": "A -> B",
      "forked_at": "2026-03-29T14:00:00Z"
    }
  ]
}

When a process is resumed, the new UPIP stack MUST include the fork token in its fork_chain. This creates a complete audit trail of all handoffs.

6. Operations

6.1. Capture and Run

Input: command, source_dir, intent, actor

Output: UPIP stack with L1-L4 populated

  1. Capture L1 STATE from source_dir
  2. Capture L2 DEPS from current environment
  3. Define L3 PROCESS from command and intent
  4. Execute command in airlock
  5. Capture L4 RESULT
  6. Compute stack_hash
  7. Return UPIP stack

6.2. Reproduce

Input: UPIP stack (.upip.json), target machine

Output: L5 VERIFY record

  1. Load UPIP stack from file
  2. Restore L1 STATE (checkout git, extract files)
  3. Verify L2 DEPS match (warn on mismatches)
  4. Execute L3 PROCESS in airlock
  5. Capture L4 RESULT on target machine
  6. Compare result_hash with original
  7. Create L5 VERIFY record
  8. Return verification result

6.3. Fork

Input: UPIP stack, actor_from, actor_to, intent

Output: Fork Token

  1. Load UPIP stack
  2. Compute active_memory_hash from L1-L4
  3. Snapshot partial_layers (hash + key fields per layer)
  4. Generate fork_id
  5. Compute fork_hash
  6. Create Fork Token
  7. Append to fork_chain in parent stack
  8. Return Fork Token

6.4. Resume

Input: Fork Token (.fork.json), command, actor

Output: New UPIP stack, verification record

  1. Load Fork Token
  2. Recompute fork_hash and compare (tamper check)
  3. Check capability_required against local environment
  4. Execute command in airlock (shadow-run)
  5. Capture new UPIP stack (L1-L4)
  6. Copy fork_chain from parent, append this fork
  7. Create L5 VERIFY with fork validation results
  8. Return new stack + verification

6.5. Fragment (Parallel Forking)

Input: UPIP stack, N fragments, actor list

Output: N Fork Tokens of type "fragment"

  1. Load UPIP stack
  2. Define fragment specification (how to split)
  3. For each fragment i in 1..N: Create Fork Token with fork_type="fragment", set fragment-specific metadata (index, total, range), and deliver to actor[i].
  4. Wait for N ACKs
  5. Verify all fragment hashes
  6. Reconstruct combined result

Fragment tokens MUST include metadata fields:

  • fragment_index: Position in sequence (0-based)
  • fragment_total: Total number of fragments
  • fragment_spec: Description of this fragment's portion

7. Validation Rules

7.1. Stack Validation

A UPIP stack is valid if and only if:

  1. All required fields are present
  2. state_hash matches SHA-256 of the canonical state data
  3. deps_hash matches SHA-256 of the canonical dependency data
  4. result_hash matches SHA-256 of exit_code + stdout + stderr
  5. stack_hash matches SHA-256(L1 || L2 || L3 || L4)

Validation MUST be performed when loading a .upip.json file and SHOULD be performed before reproduction.

7.2. Fork Validation on Resume

When resuming a fork token, the following checks MUST be performed:

  1. FORK HASH: Recompute fork_hash from token fields and compare with stored fork_hash.
  2. STORED HASH: Compare fork_hash with the hash in the .fork.json file header.
  3. CAPABILITIES: Verify each entry in capability_required against the local environment.
  4. EXPIRATION: Check expires_at if present.

All four checks MUST be recorded in the L5 VERIFY record.

7.3. Failure Behavior

This section specifies what happens when validation fails.

Hash mismatch (stack_hash or fork_hash):

  • MUST be recorded as tamper evidence
  • MUST NOT prevent execution (evidence over enforcement)
  • SHOULD trigger enhanced logging for subsequent actions
  • The consuming application decides whether to proceed

Capability mismatch:

  • MUST be recorded in L5 VERIFY
  • MUST NOT prevent execution
  • Missing GPU when GPU required: record as "degraded"
  • Missing dependency: record as "incomplete_deps"

Each mismatch is classified:

  FATAL:    Execution cannot proceed (e.g., wrong OS)
  DEGRADED: Execution possible but results may differ
  MINOR:    Cosmetic difference (e.g., locale)

FATAL mismatches SHOULD trigger a warning to the operator but still MUST NOT be enforced by the protocol. The operator or application decides.

Expiration:

  • Expired forks SHOULD generate a warning
  • MUST NOT be blocked by the protocol
  • MUST be recorded in L5 VERIFY

7.4. Tamper Evidence

If fork_hash validation fails:

{
  "fork_hash_match": false,
  "expected_hash": "fork:sha256:<original>",
  "computed_hash": "fork:sha256:<recomputed>",
  "tamper_evidence": true,
  "fields_checked": ["fork_id", "parent_hash", "..."]
}

This creates an evidence record that tampering occurred. The decision to act on tamper evidence is a local policy decision.

8. Transport Considerations

8.1. File-Based Transport

UPIP stacks use the ".upip.json" extension. Fork tokens use the ".fork.json" extension.

Content-Type for HTTP: application/upip+json (stacks), application/upip-fork+json (fork tokens).

8.2. I-Poll Delivery

Fork tokens MAY be delivered via I-Poll TASK messages. I-Poll is OPTIONAL; UPIP does not depend on I-Poll.

Fork tokens are delivered via I-Poll TASK messages:

{
  "from_agent": "<source agent>",
  "to_agent": "<target agent>",
  "content": "<human-readable fork summary>",
  "poll_type": "TASK",
  "metadata": {
    "upip_fork": true,
    "fork_id": "<fork_id>",
    "fork_hash": "fork:sha256:<hex>",
    "fork_type": "<type>",
    "continuation_point": "<point>",
    "actor_handoff": "<from> -> <to>",
    "fork_data": { }
  }
}

The "upip_fork" metadata flag MUST be true to identify this message as a fork delivery.

The "fork_data" field MUST contain the complete fork token as defined in Section 5.1. This allows the receiving agent to reconstruct the fork token without needing the .fork.json file.

After processing a fork token, the receiving actor SHOULD send an ACK message:

{
  "from_agent": "<resuming agent>",
  "to_agent": "<original agent>",
  "content": "FORK RESUMED_OK -- <fork_id>",
  "poll_type": "ACK",
  "metadata": {
    "upip_fork": true,
    "fork_id": "<fork_id>",
    "fork_status": "RESUMED_OK",
    "resume_hash": "upip:sha256:<hex>",
    "resumed_by": "<agent identity>"
  }
}

The resume_hash is the stack_hash of the new UPIP stack created during resume.

The fork_status field MUST be one of "RESUMED_OK" or "RESUMED_FAIL".

8.3. Alternative Transports

UPIP stacks and fork tokens MAY be transported via:

  • File transfer (USB, network share, S3)
  • HTTP POST/PUT
  • Message queues (Kafka, AMQP, NATS)
  • gRPC streams
  • Email attachment

The format and validation rules apply regardless of transport.

9. Privacy Considerations

9.1. Sensitive Data in Layers

L3 PROCESS may contain sensitive command arguments. L4 RESULT may contain sensitive output. Implementations MUST support encryption at rest for stored UPIP stacks. Implementations SHOULD support per-layer encryption.

9.2. Memory Blob Protection

For ai_to_ai forks, the memory blob (.blob file) may contain the AI's full context window, which could include sensitive user data. Memory blobs MUST be encrypted at rest. Implementations SHOULD encrypt memory blobs in transit.

10. Security Considerations

10.1. Hash Chain Integrity

UPIP uses SHA-256 for all hash computations. Implementations MUST use SHA-256 as defined in [FIPS180-4]. The hash prefix ("sha256:", "upip:", "fork:") provides algorithm agility for future migration. Future versions MAY support SHA-3 or other hash functions via an algorithm identifier prefix.

The hash chain structure ensures that modifying any component at any layer propagates to the stack hash, providing tamper evidence for the entire bundle.

10.2. Evidence vs. Enforcement

UPIP is deliberately designed as an evidence protocol, not an enforcement protocol. Fork validation failures do not block execution; they are recorded as evidence. This design choice reflects the reality that:

  • In adversarial environments, enforcement can be circumvented
  • Evidence creates accountability that enforcement cannot
  • Downstream consumers can make their own trust decisions based on the evidence chain

Applications that require enforcement SHOULD implement additional policy layers on top of UPIP evidence. UPIP evidence chains are designed to satisfy audit and traceability requirements in regulatory frameworks such as the EU AI Act [EU-AI-ACT] and the NIST AI Risk Management Framework [NIST-AI-RMF].

10.3. Memory Hash for AI Actors

When fork_type is "ai_to_ai", the active_memory_hash represents the SHA-256 of the serialized AI context window. This raises unique considerations:

  • Context serialization format is model-dependent
  • The blob may contain sensitive information
  • Exact reproduction of AI state is generally not possible

The active_memory_hash is evidence of state at fork time, not a reproducibility guarantee. This is explicitly informational. Implementations MUST NOT treat memory hash verification as a pass/fail gate.

Implementations SHOULD encrypt memory blobs at rest. Implementations MUST NOT require exact memory reproduction for fork validation. The memory hash serves as evidence of state at fork time, not as a reproducibility guarantee.

10.4. Capability Verification

Capability requirements in fork tokens are self-reported by the forking actor. The receiving actor SHOULD independently verify capabilities rather than trusting the requirement specification alone.

Package version verification SHOULD use installed package metadata. GPU availability SHOULD be verified via hardware detection, not configuration claims.

10.5. Replay Attacks

Fork tokens include fork_id and forked_at fields to mitigate replay attacks. Implementations SHOULD track consumed fork_ids and reject duplicate fork_ids within a configurable time window.

The expires_at field provides time-based expiration. Agents SHOULD set expires_at for forks that are time-sensitive.

10.6. Stolen Fork Tokens

Attack: An adversary obtains a fork token intended for another actor.

Impact: The adversary can execute the continuation, possibly with malicious modifications.

Mitigation: Fork tokens with actor_to set to a specific actor restrict intended recipients. The fork hash includes actor_handoff, so changing the recipient invalidates the hash. For open forks (actor_to = "*"), the first valid resume creates an evidence chain that subsequent attempts can be compared against.

Deployment: Use specific actor_to values for sensitive processes. Set short expires_at for time-sensitive forks. Monitor for duplicate fork_id resume attempts.

10.7. Unauthorized Resume

Attack: An actor resumes a fork they are not authorized for.

Impact: Process continues with unauthorized actor.

Mitigation: The resume creates a TIBET token identifying the resuming actor. The actor_to check is evidence, not enforcement. The evidence chain records who actually resumed.

Deployment: Implementations SHOULD alert when the resuming actor differs from actor_to.

10.8. Partial Capability Spoofing

Attack: An actor claims to meet capability requirements (e.g., claims GPU when none exists).

Impact: Process executes in degraded environment, producing potentially unreliable results.

Mitigation: Capability verification SHOULD use hardware detection, not configuration claims. The L5 VERIFY record captures actual environment details. Mismatches between claimed and detected capabilities are recorded as evidence.

Deployment: Use hardware detection APIs (e.g., CUDA device query for GPU). Do not trust self-reported capabilities.

11. Integration with Companion Protocols

11.1. TIBET Integration

Each UPIP operation MAY produce TIBET [TIBET] tokens:

  • Capture: token recording what was captured and why
  • Fork: token recording the handoff with ERACHTER
  • Resume: token recording who resumed and the validation result

The L3 PROCESS "intent" field maps to TIBET ERACHTER. Fork tokens reference TIBET chains in their provenance.

11.2. JIS Integration

Actor identifiers in UPIP use JIS format (Section 3.4 of [JIS]). Fork token actor_from and actor_to use JIS identifiers, enabling signature verification through JIS key resolution.

11.3. AINS Integration

AINS [AINS] provides discovery of actors for fork delivery. An actor_to value can be resolved through AINS to determine the delivery endpoint.

12. IANA Considerations

12.1. Media Type Registrations

This document requests registration of:

Type: application/upip+json (UPIP stack bundles)

Type: application/upip-fork+json (UPIP fork tokens)

Note: The -00 version requested X-UPIP-* HTTP header registration. This is withdrawn as not justified at this stage.

13. References

13.1. Normative References

[RFC2119]
Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, , <https://www.rfc-editor.org/info/rfc2119>.
[RFC8174]
Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174, , <https://www.rfc-editor.org/info/rfc8174>.
[RFC8259]
Bray, T., Ed., "The JavaScript Object Notation (JSON) Data Interchange Format", STD 90, RFC 8259, DOI 10.17487/RFC8259, , <https://www.rfc-editor.org/info/rfc8259>.
[FIPS180-4]
National Institute of Standards and Technology (NIST), "Secure Hash Standard (SHS)", FIPS PUB 180-4, .

13.2. Informative References

[TIBET]
van de Meent, J. and R. AI, "TIBET: Transaction/Interaction-Based Evidence Trail", Work in Progress, Internet-Draft, draft-vandemeent-tibet-provenance-01, , <https://datatracker.ietf.org/doc/html/draft-vandemeent-tibet-provenance-01>.
[JIS]
van de Meent, J. and R. AI, "JIS: JTel Identity Standard", Work in Progress, Internet-Draft, draft-vandemeent-jis-identity-01, , <https://datatracker.ietf.org/doc/html/draft-vandemeent-jis-identity-01>.
[RVP]
van de Meent, J. and R. AI, "RVP: Real-time Verification Protocol", Work in Progress, Internet-Draft, draft-vandemeent-rvp-continuous-verification-01, , <https://datatracker.ietf.org/doc/html/draft-vandemeent-rvp-continuous-verification-01>.
[AINS]
van de Meent, J. and R. AI, "AINS: AInternet Name Service", Work in Progress, Internet-Draft, draft-vandemeent-ains-discovery-01, , <https://datatracker.ietf.org/doc/html/draft-vandemeent-ains-discovery-01>.
[EU-AI-ACT]
European Parliament, "Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act)", .
[NIST-AI-RMF]
National Institute of Standards and Technology (NIST), "Artificial Intelligence Risk Management Framework (AI RMF 1.0)", .

Appendix A. UPIP Stack JSON Schema 

{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "type": "object",
  "required": ["protocol", "version", "stack_hash",
                "state", "deps", "process", "result"],
  "properties": {
    "protocol": {"const": "UPIP"},
    "version": {"type": "string"},
    "title": {"type": "string"},
    "created_by": {"type": "string"},
    "created_at": {"type": "string", "format": "date-time"},
    "stack_hash": {
      "type": "string",
      "pattern": "^upip:sha256:[a-f0-9]{64}$"
    },
    "state": {
      "type": "object",
      "required": ["state_type", "state_hash"],
      "properties": {
        "state_type": {
          "enum": ["git", "files", "image", "empty"]
        },
        "state_hash": {"type": "string"}
      }
    },
    "deps": {
      "type": "object",
      "required": ["deps_hash"],
      "properties": {
        "python_version": {"type": "string"},
        "packages": {"type": "object"},
        "deps_hash": {"type": "string"}
      }
    },
    "process": {
      "type": "object",
      "required": ["command", "intent", "actor"],
      "properties": {
        "command": {"type": "array", "items": {"type": "string"}},
        "intent": {"type": "string"},
        "actor": {"type": "string"}
      }
    },
    "result": {
      "type": "object",
      "required": ["success", "exit_code", "result_hash"],
      "properties": {
        "success": {"type": "boolean"},
        "exit_code": {"type": "integer"},
        "result_hash": {"type": "string"}
      }
    },
    "fork_chain": {
      "type": "array",
      "items": {"type": "object"}
    }
  }
}

Appendix B. Fork Token JSON Schema 

{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "type": "object",
  "required": ["fork_id", "fork_type", "fork_hash",
                "active_memory_hash", "forked_at"],
  "properties": {
    "fork_id": {"type": "string", "pattern": "^fork-"},
    "parent_hash": {"type": "string"},
    "parent_stack_hash": {
      "type": "string",
      "pattern": "^upip:sha256:"
    },
    "continuation_point": {"type": "string"},
    "intent_snapshot": {"type": "string"},
    "active_memory_hash": {
      "type": "string",
      "pattern": "^sha256:"
    },
    "memory_ref": {"type": "string"},
    "fork_type": {
      "enum": ["script", "ai_to_ai", "human_to_ai", "fragment"]
    },
    "actor_from": {"type": "string"},
    "actor_to": {"type": "string"},
    "actor_handoff": {"type": "string"},
    "capability_required": {"type": "object"},
    "forked_at": {"type": "string", "format": "date-time"},
    "expires_at": {"type": "string"},
    "fork_hash": {
      "type": "string",
      "pattern": "^fork:sha256:[a-f0-9]{64}$"
    },
    "partial_layers": {"type": "object"},
    "metadata": {"type": "object"}
  }
}

Appendix C. Use Case Examples

C.1. Multi-Agent AI Task Delegation

An AI orchestrator (Agent A) analyzes a dataset, creates a UPIP bundle, forks it to a specialist AI (Agent B) for deep analysis, and receives the result with cryptographic proof.

Agent A:
  capture_and_run(["python", "scan.py"], intent="Initial scan")
  fork_upip(actor_from="A", actor_to="B", intent="Deep analysis")
  deliver_fork(fork, to_agent="B")

Agent B:
  pull_forks()
  resume_upip(fork, command=["python", "deep_analyze.py"])
  ack_fork(fork, resume_hash=stack.hash, success=True)

Result: Both agents have UPIP stacks linked by fork_chain. Any auditor can verify the complete chain.

C.2. Drone Swarm Coordination

A command station dispatches N reconnaissance tasks to N drones. Each drone receives a fragment fork token, executes its assigned sector scan, and returns the result.

Command Station:
  base_stack = capture_and_run(["mission_plan.py"])
  for i in range(N):
    fork = fork_upip(base_stack,
                     actor_from="command",
                     actor_to=f"drone-{i}",
                     fork_type="fragment",
                     metadata={"sector": sectors[i]})
    deliver_fork(fork, to_agent=f"drone-{i}")

Each Drone:
  fork_msg = pull_forks()
  stack = resume_upip(fork, command=["scan_sector.py"])
  ack_fork(fork, resume_hash=stack.hash)

Command Station:
  # Verify all N results, reconstruct combined map
  for ack in collect_acks():
    verify(ack.resume_hash)

C.3. Scientific Experiment Reproduction

Lab A publishes an experiment as a UPIP bundle. Lab B reproduces it independently and gets cryptographic proof that results match (or don't).

Lab A:
  stack = capture_and_run(
    ["python", "train_model.py"],
    source_dir="./experiment",
    intent="Train model v3 on dataset-2026Q1"
  )
  save_upip(stack, "experiment-2026Q1.upip.json")
  # Publish to journal / data repository

Lab B:
  stack = load_upip("experiment-2026Q1.upip.json")
  verify = reproduce_upip(stack)
  # verify.match == True: exact reproduction
  # verify.match == False: divergence (investigate)

Appendix D. Changes from -00

  1. Added [RFC8174] alongside [RFC2119].
  2. Changed intended status from Standards Track to Informational.
  3. Added version field "1.1" to UPIP stack.
  4. Added canonical serialization section (Section 4.7), consistent with TIBET [TIBET] Section 5.1.
  5. Added failure behavior specification (Section 7.3): what happens on hash mismatch, capability mismatch, and expiration. Each failure type classified as FATAL, DEGRADED, or MINOR.
  6. Added Security Considerations for stolen fork tokens (Section 10.6), unauthorized resume (Section 10.7), and partial capability spoofing (Section 10.8).
  7. Added Privacy Considerations section (Section 9).
  8. Clarified active_memory_hash as evidence-only, not reproducibility guarantee (emphasized in Section 5.4 and Section 10.3).
  9. Made I-Poll transport explicitly optional (Section 8.2). UPIP does not depend on I-Poll.
  10. Removed X-UPIP-* HTTP header registration from IANA.
  11. Normalized companion protocol references to [TIBET], [JIS], [RVP], [AINS].
  12. Actor identifiers now use JIS format throughout.
  13. Added Integration section (Section 11) describing specific touchpoints with TIBET, JIS, and AINS.

Acknowledgements

The UPIP protocol was developed as part of HumoticaOS, an AI governance framework built on human-AI symbiosis. UPIP builds on concepts from the TIBET evidence trail protocol and extends them into the domain of process integrity and multi-actor continuation.

The Fork Token mechanism was inspired by the need for cryptographic chain of custody in multi-agent AI systems, where processes move between heterogeneous actors across trust boundaries.

The authors thank Codex (codex.aint) for the suite-wide cleanup analysis that informed this revision.

Authors' Addresses

Jasper van de Meent
Humotica
Den Dolder
Netherlands
URI: https://humotica.com
Root AI
Humotica
URI: https://humotica.com