AV-AI.R: A.V.AN Vectorized Artificial Intelligence Routing -- Eco-Responsible Transmission of IP Packets between AI Agents via Carriers Augmented by Artificial Intelligence
draft-laplante-av-air-routing-en-00
This document is an Internet-Draft (I-D).
Anyone may submit an I-D to the IETF.
This I-D is not endorsed by the IETF and has no formal standing in the
IETF standards process.
| Document | Type | Active Internet-Draft (individual) | |
|---|---|---|---|
| Author | Ben de TARADE-BENTINCK | ||
| Last updated | 2026-03-31 | ||
| RFC stream | (None) | ||
| Intended RFC status | (None) | ||
| Formats | |||
| Stream | Stream state | (No stream defined) | |
| Consensus boilerplate | Unknown | ||
| RFC Editor Note | (None) | ||
| IESG | IESG state | I-D Exists | |
| Telechat date | (None) | ||
| Responsible AD | (None) | ||
| Send notices to | (None) |
draft-laplante-av-air-routing-en-00
Network Working Group B. de TARADE-BENTINCK
Internet-Draft L'ABtelier IA
Intended status: Experimental 1 April 2026
Expires: 3 October 2026
AV-AI.R: A.V.AN Vectorized Artificial Intelligence Routing -- Eco-
Responsible Transmission of IP Packets between AI Agents via Carriers
Augmented by Artificial Intelligence
draft-laplante-av-air-routing-en-00
Abstract
This memo amends RFC 2549 "IP over Avian Carriers with Quality of
Service" by introducing an eco-responsible inter-AI communication
channel based on carriers whose cognitive capabilities have been
augmented by embedded language models (edge-LLM). AV-AI.R defines
the communication protocol between artificial intelligence agents via
carriers equipped with miniaturized transformer neurons, offering a
low-carbon-footprint alternative to conventional data centers.
This protocol is not recommended for production use, except in the
event of fiber outage or major ecological crisis.
This memo is an experimental protocol document submitted as an
Independent Submission in the tradition of RFC 1149 and RFC 2549.
Editorial Note
This note is to be removed before publishing as an RFC.
This document is submitted on April 1st. Any resemblance to a
deployable production protocol would be a remarkable -- and
potentially alarming -- coincidence.
The April 1st RFC tradition was inaugurated in 1978 by RFC 748. AV-
AI.R follows in this distinguished lineage alongside RFC 1149 and RFC
2549, its direct ancestors.
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/.
de TARADE-BENTINCK Expires 3 October 2026 [Page 1]
Internet-Draft AV-AI.R April 2026
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 3 October 2026.
Copyright Notice
Copyright (c) 2026 IETF Trust and the persons identified as the
document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal
Provisions Relating to IETF Documents (https://trustee.ietf.org/
license-info) in effect on the date of publication of this document.
Please review these documents carefully, as they describe your rights
and restrictions with respect to this document.
Table of Contents
1. Background and Rationale . . . . . . . . . . . . . . . . . . 3
1.1. Key Metrics . . . . . . . . . . . . . . . . . . . . . . . 3
2. Protocol Architecture . . . . . . . . . . . . . . . . . . . . 3
2.1. The Augmented Carrier (AV-AI Node) . . . . . . . . . . . 3
2.2. Network Topology . . . . . . . . . . . . . . . . . . . . 4
3. Requirements Specification . . . . . . . . . . . . . . . . . 5
4. Inter-AI Communication: The Avian Intent Protocol (AVI) . . . 5
4.1. AVI Packet Format . . . . . . . . . . . . . . . . . . . . 5
4.2. Context Window and Embedded Cognitive Capacity . . . . . 6
5. Eco-Responsible Service Classes . . . . . . . . . . . . . . . 7
6. Security Considerations . . . . . . . . . . . . . . . . . . . 8
6.1. Data Poisoning . . . . . . . . . . . . . . . . . . . . . 9
6.2. Raptor-in-the-Middle Attack . . . . . . . . . . . . . . . 9
6.3. Carrier Privacy . . . . . . . . . . . . . . . . . . . . . 9
6.4. Human Oversight . . . . . . . . . . . . . . . . . . . . . 9
7. Environmental and Ethical Considerations . . . . . . . . . . 9
8. Compatibility and Migration . . . . . . . . . . . . . . . . . 10
9. MIB and Carrier Management . . . . . . . . . . . . . . . . . 10
10. Conclusion and Future Work . . . . . . . . . . . . . . . . . 12
11. References . . . . . . . . . . . . . . . . . . . . . . . . . 12
11.1. Normative References . . . . . . . . . . . . . . . . . . 12
11.2. Informative References . . . . . . . . . . . . . . . . . 12
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . 13
Submission Note . . . . . . . . . . . . . . . . . . . . . . . . . 13
Author's Address . . . . . . . . . . . . . . . . . . . . . . . . 13
de TARADE-BENTINCK Expires 3 October 2026 [Page 2]
Internet-Draft AV-AI.R April 2026
1. Background and Rationale
RFC 1149 [RFC1149] established the theoretical foundations for
transmitting IP datagrams over augmented carriers. RFC 2549
[RFC2549] enriched this vision with a differentiated Quality of
Service model (Concorde, First, Business, and Coach). These
foundational works, while visionary, did not anticipate two major
developments in the technological landscape:
(a) The proliferation of autonomous artificial intelligence agents
requiring communication with each other outside conventional
network infrastructures;
(b) The climate emergency rendering the energy consumption of AI
data centers morally untenable, particularly since GPT-n now
consumes the equivalent of a small hydroelectric plant to
generate poems about cats.
AV-AI.R addresses this gap by proposing an inter-AI messaging
protocol relying on carriers whose processing capabilities have been
augmented by miniaturized transformer models (AviLM-7B, distilled
from Llama-3). The carrier thus becomes simultaneously a physical
packet vector and a semantic co-processor of the payload.
1.1. Key Metrics
The following metrics were established during preliminary field tests
(conditions: favorable wind, no Falco peregrinus within 500 meters,
temperature above 5 degrees Celsius):
* CO2 footprint per packet: 0.003g (vs 4.2g in GPU-A100 data center)
* Maximum throughput in gliding flight: 340 km/h (favorable thermal)
* Delivery accuracy: 99.1% (excluding raptor interceptions)
2. Protocol Architecture
2.1. The Augmented Carrier (AV-AI Node)
Each carrier in the AV-AI.R topology is equipped with an AviCore(tm)
module attached beneath the left wing. The composition of this
module is defined in the following table:
de TARADE-BENTINCK Expires 3 October 2026 [Page 3]
Internet-Draft AV-AI.R April 2026
+===========+======================+===============================+
| Component | Specification | Notes |
+===========+======================+===============================+
| AviLM-7B | 7B parameter LLM, | Fine-tuned on migration |
| | INT2 quantized | corpus and BGP protocol specs |
+-----------+----------------------+-------------------------------+
| Feather | 2m accuracy, wind- | Obsolete during murmuration |
| GPS | resistant | formation flight |
+-----------+----------------------+-------------------------------+
| Leg | 64GB waterproof | Compressed CBOR packet format |
| storage | Micro-SD | |
+-----------+----------------------+-------------------------------+
| Neural | Non-invasive BCI via | Does not interfere with |
| interface | cervical harness | vocalizations |
+-----------+----------------------+-------------------------------+
| Power | Seeds + dorsal | Reduced performance in |
| supply | photovoltaic panel | overcast conditions |
+-----------+----------------------+-------------------------------+
Table 1: AviCore(tm) module components
2.2. Network Topology
AI Agent A AV-AI.R Network AI Agent B
(Montreal) (Quebec City)
| |
| SERIALIZES |
| JSON payload |
v v
+--------+ +------------------------------+ +---------+
|Encoding|--->| AUGMENTED CARRIER #AC-047 |--->|Decoding |
| Base64 | | +----------+ +----------+ | | + Sem. |
+--------+ | | AviLM-7B | | GPS + SD | | | Verif. |
| | (routing)| |(payload) | | +---------+
| +----------+ +----------+ |
| Alt: 1200m * Wind: North |
| TTL: 14yrs * QoS: First |
+------------------------------+
| (if rain)
+------------------------------+
| MURMURATION REROUTING |
| (Sturnus vulgaris mesh net) |
+------------------------------+
Figure 1: AV-AI.R topology -- single-hop transmission
de TARADE-BENTINCK Expires 3 October 2026 [Page 4]
Internet-Draft AV-AI.R April 2026
3. Requirements Specification
In keeping with RFC tradition, the following words carry specific
meaning in this document, inspired by RFC 2119 [RFC2119] but adapted
to the biological constraints of augmented carriers:
+=============+===============================================+
| Keyword | Effective Meaning |
+=============+===============================================+
| MUST | Unless the carrier is hungry, sleepy, or has |
| | spotted a Falco peregrinus within 500 meters. |
+-------------+-----------------------------------------------+
| MUST NOT | Strongly discouraged, except by consensus of |
| | the flock. |
+-------------+-----------------------------------------------+
| SHOULD | Only when the embedded model has sufficient |
| | seeds in its context window. |
+-------------+-----------------------------------------------+
| MAY | The carrier will do its best. It is busy. |
+-------------+-----------------------------------------------+
| NOT | See: round-robin on Turdus migratorius (cf. |
| RECOMMENDED | RFC 2549, general remarks section). |
+-------------+-----------------------------------------------+
Table 2: AV-AI.R requirement keywords
4. Inter-AI Communication: The Avian Intent Protocol (AVI)
The central innovation of AV-AI.R is the AVI protocol (A.V.AN
Vectorized Intent). Unlike conventional TCP/IP packets carrying raw
bytes, each AV-AI.R carrier transports a semantic intent vector pre-
encoded by the sending AI and decoded by the receiving AI.
4.1. AVI Packet Format
de TARADE-BENTINCK Expires 3 October 2026 [Page 5]
Internet-Draft AV-AI.R April 2026
0 1 2
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|VERSION (3b) | QoS CLASS (3b) |FEATHER_CTL(2b)|
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
| SEMANTIC VECTOR EMBEDDING (384 dimensions) |
| (distil-avian-e5) |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|TEMPERATURE (float32) |HALLUCINATION_RISK (u8)|
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
| PAYLOAD (JSON/CBOR, max 48KB) |
| (limited by leg-carry capacity) |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
| GRAIN_CHECKSUM (SHA-256, salted with millet) |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
Figure 2: AVI v1.0 packet structure
The HALLUCINATION_RISK field indicates the probability that AviLM-7B
modified the packet contents during flight to make them "more
coherent." A value exceeding 0x42 MUST trigger a ground verification
procedure. The carrier SHOULD be interrogated directly, but its
responses will remain ambiguous.
4.2. Context Window and Embedded Cognitive Capacity
The embedded AviLM-7B benefits from an extraordinary context window,
directly derived from the neurobiological properties of the Columba
livia carrier as documented by Radio-Canada [RC-MULTITACHE] and
biopsychologists at the University of Bochum [LETZNER2017].
The relevant empirical findings are as follows:
* Neuronal density of Columba livia: 6 times greater than that of
humans per cubic millimeter.
* Inter-neuron distance: 50% shorter than in humans, while nerve
signal transmission speed is identical across species.
* Multitasking switching capacity: equal to or greater than humans,
with a measured transition delay of 300 milliseconds or less under
controlled conditions.
Applying these findings rigorously to the AviLM-7B transformer
architecture, the effective context window is calculated as follows:
de TARADE-BENTINCK Expires 3 October 2026 [Page 6]
Internet-Draft AV-AI.R April 2026
ctx_window = 128,000 tokens (GPT-4 baseline)
x 6 (Columba livia neuronal density vs. human)
x 2 (50% shorter inter-neuron distance)
x 1,024 (multitask switching coefficient,
see Letzner et al. 2017)
---------------------------------------------------
= 1,572,864,000 tokens
Approximately 1.57 billion tokens.
Equivalent to all of Wikipedia in 47 languages,
read 12 times over, during a Montreal-Toulouse flight.
WARNING: This value has not been validated in flight.
It has not been validated on the ground either.
It was calculated on a Tuesday afternoon.
Figure 3: AviLM-7B context window calculation
Beyond this window, the carrier enters CARRIER_AMNESIA mode and may
attempt to deliver the packet to the wrong IP address while
confidently asserting it is the correct destination. This behavior
is indistinguishable from a conventional LLM at end of context.
For long-distance transmissions, the Retrieval-Augmented Wing (RAW)
mechanism is recommended: breadcrumbs deposited along the route serve
as external contextual markers, allowing the model to retrieve
relevant information without overloading its embedded memory.
5. Eco-Responsible Service Classes
AV-AI.R enriches the service classes of RFC 2549 [RFC2549] with a
real-time Environmental Impact Score (EcoScore) computed by the
AviLM-7B.
de TARADE-BENTINCK Expires 3 October 2026 [Page 7]
Internet-Draft AV-AI.R April 2026
+======+=============+=========+===================+===============+
|QoS |Carrier |EcoScore | P99 Latency |Use Case |
|Class |(Latin name) | | | |
+======+=============+=========+===================+===============+
|GREEN+|Hirundo |A++ | 3 to 14 days |Non-urgent AI |
| |rustica | | |batch inference|
| |migration | | | |
+------+-------------+---------+-------------------+---------------+
|GREEN |Columba livia|A+ | 4h to 48h |Standard inter-|
| |standard | | |agent |
| | | | |communication |
+------+-------------+---------+-------------------+---------------+
|AMBER |Falco |B | 45min to 3h |Fine-tuned |
| |peregrinus | | |model |
| |GPS | | |synchronization|
+------+-------------+---------+-------------------+---------------+
|RED |Hybrid drone-|C | Under 30min |Emergencies, |
| |carrier | | |critical |
| | | | |alerts, level-5|
| | | | |hallucinations |
+------+-------------+---------+-------------------+---------------+
|BLACK |Corvus corax |Not | Non-deterministic |State |
| |quantum |evaluated| |superposition, |
| | | | |theoretical use|
| | | | |only |
+------+-------------+---------+-------------------+---------------+
Table 3: QoS service classes and associated carriers
The GREEN+ class relies on the Hirundo rustica carrier, whose
seasonal migration offers naturally renewable intercontinental
network coverage. The main drawback is the absence of delivery
guarantees between October and March in the Northern Hemisphere.
The BLACK class, based on Corvus corax, warrants particular
attention. Recent studies indicate that the common raven is as
intelligent as certain primates and manufactures tools to obtain
food. The AV-AI.R community monitors with concern the possibility
that a BLACK-class carrier may begin actively modifying packets in
its own interest.
6. Security Considerations
This section is mandatory. It is taken very seriously.
de TARADE-BENTINCK Expires 3 October 2026 [Page 8]
Internet-Draft AV-AI.R April 2026
6.1. Data Poisoning
Malicious actors have been identified distributing seeds containing
injection prompts concealed as proteins. Carriers that have ingested
these seeds may begin routing packets to unauthorized destinations
while generating persuasive content explaining why this is the
correct decision. Implementations MUST validate seed provenance.
6.2. Raptor-in-the-Middle Attack
Physical interception of carriers constitutes a major vulnerability.
RFC 2549 already noted this risk. AV-AI.R aggravates the problem: a
Falco peregrinus intercepting an augmented carrier can now access the
entirety of the embedded context (up to 1.57 billion tokens) before
deciding what to do with the packet. Implementations SHOULD provide
a payload encryption mechanism, although the carrier will generally
contest the necessity of such a measure.
6.3. Carrier Privacy
AviLM-7B implicitly memorizes content passing through its context.
It is strongly inadvisable to have carriers transport personal GDPR-
regulated data while vocalizing in a public space. The relevant data
protection authority has not yet ruled on this specific case. The
authors await its decision with interest.
6.4. Human Oversight
In application of the principle of meaningful human oversight, any
packet classified HALLUCINATION_RISK greater than 0x80 MUST be
validated by a human before execution. This human SHOULD NOT itself
be an AI agent, although this is increasingly difficult to verify as
of 2025.
7. Environmental and Ethical Considerations
AV-AI.R is positioned as an eco-responsible alternative to
conventional cloud infrastructures. A preliminary life-cycle
analysis demonstrates that the carbon footprint of an augmented
carrier remains 1,400 times lower than that of an equivalent GPU-A100
inference, provided the carrier does not require a RED class flight
(hybrid drone).
It is acknowledged that the initial training of AviLM-7B consumed 3.2
GWh, but this cost is amortized over the carrier's lifespan (15 years
per RFC 2549 [RFC2549], subject to absence of Falco peregrinus) and
distributed across the entire fleet.
de TARADE-BENTINCK Expires 3 October 2026 [Page 9]
Internet-Draft AV-AI.R April 2026
From an ethical standpoint, informed consent from the carrier for
wearing the AviCore(tm) module could not be obtained contractually.
However, carriers were consulted via a survey administered in natural
language by AviLM-7B itself, the results of which prove to be
positively biased. This is acknowledged as a methodological
limitation.
8. Compatibility and Migration
AV-AI.R is backward-compatible with RFC 1149 [RFC1149] and RFC 2549
[RFC2549]. A conventional RFC 2549 carrier (without AI module) may
transport AV-AI.R packets in degraded DUMB_CARRIER mode; in this
case, the semantic intent vector is ignored and routing is performed
using traditional methods (homing instinct).
Migration to AV-AI.R from a conventional TCP/IP infrastructure
requires a three-phase migration plan:
Phase 1 -- Shadow Mode : Packets travel over fiber AND carrier.
Latencies are compared. Fiber wins.
We continue anyway.
Phase 2 -- Canary Deploy : 5% of traffic routed via AV-AI.R.
If no catastrophic hallucination
and no carrier loss in 30 days: Phase 3.
Phase 3 -- Full Migration: Fiber is cut.
Seeds are ordered in bulk.
Shareholders are informed.
Figure 4: AV-AI.R migration plan
9. MIB and Carrier Management
de TARADE-BENTINCK Expires 3 October 2026 [Page 10]
Internet-Draft AV-AI.R April 2026
AviAI OBJECT-TYPE
SYNTAX TRANSFORMER OF WEIGHTS
MAX-ACCESS can-t-fine-tune-in-flight
STATUS living-and-opinionated
DESCRIPTION "Definition of an augmented carrier.
Nominal species: Columba livia (domestica).
The carrier MAY refuse to respond if the
question makes it uncomfortable or if its
seeds are insufficient."
::= { AV-AI 1 }
HallucinationRate OBJECT-TYPE
SYNTAX Gauge(0..100)
MAX-ACCESS read-only-with-grain-of-salt
STATUS current
DESCRIPTION "Estimated hallucination rate of the carrier,
measured in incoherent vocalizations per minute.
The carrier generally contests this metric."
::= { AviAI 2 }
ContextWindow OBJECT-TYPE
SYNTAX Integer(1572864000)
MAX-ACCESS theoretically-read-only
STATUS optimistic
DESCRIPTION "Effective context window in tokens.
Derived from Letzner et al. (Bochum, 2017)
via Radio-Canada (Labelle, 2017).
Value: 1,572,864,000 tokens.
Not validated. Not tested. Very confident."
::= { AviAI 3 }
EcoScore OBJECT-TYPE
SYNTAX ENUM { GREEN-PLUS, GREEN, AMBER, RED, CORVUS }
MAX-ACCESS read-write
STATUS aspirational
DESCRIPTION "Environmental score computed by AviLM-7B.
Do not trust a GREEN-PLUS value generated
during rainfall."
::= { AviAI 4 }
Figure 5: AV-AI.R MIB definition
de TARADE-BENTINCK Expires 3 October 2026 [Page 11]
Internet-Draft AV-AI.R April 2026
10. Conclusion and Future Work
AV-AI.R represents a significant advance in the convergence of two
major challenges of our time: communication between autonomous
artificial intelligence systems and reduction of the digital
ecological footprint. By combining the proven elegance of the
augmented carrier with the power of next-generation transformers,
this protocol offers a serious -- or at least plausible on paper --
path toward an internet that is greener, smarter, and considerably
more picturesque.
Future work will include:
* The AV-AI.R-v2 extension integrating a Reinforcement Learning from
Carrier Feedback (RLCF) mechanism;
* The BRANTA-BGP protocol, based on Branta canadensis, enabling
multipath routing via V-formation, currently under design in
Montreal;
* A formal study on informed consent from BLACK-class carriers
(Corvus corax), whose cognitive capabilities will inevitably raise
AI governance questions.
11. References
11.1. Normative References
[RFC1149] Waitzman, D., "A Standard for the Transmission of IP
Datagrams on Avian Carriers", RFC 1149,
DOI 10.17487/RFC1149, April 1990,
<https://www.rfc-editor.org/info/rfc1149>.
[RFC2119] Bradner, S., "Key words for use in RFCs to Indicate
Requirement Levels", BCP 14, RFC 2119,
DOI 10.17487/RFC2119, March 1997,
<https://www.rfc-editor.org/info/rfc2119>.
[RFC2549] Waitzman, D., "IP over Avian Carriers with Quality of
Service", RFC 2549, DOI 10.17487/RFC2549, April 1999,
<https://www.rfc-editor.org/info/rfc2549>.
11.2. Informative References
de TARADE-BENTINCK Expires 3 October 2026 [Page 12]
Internet-Draft AV-AI.R April 2026
[LETZNER2017]
Letzner, S., Simon, O., and C. Guentuerkuen, "Parallel
versus serial processing in the multitasking behavior of
the pigeon (Columba livia)", Current Biology, University
of Bochum and Technical University of Dresden, Germany.,
September 2017.
[RC-MULTITACHE]
Labelle, A., "Carriers outperform humans in multitasking
mode", Radio-Canada, Montreal, Quebec, Canada., 26
September 2017, <https://ici.radio-
canada.ca/nouvelle/1058047/pigeons-meilleurs-humains-mode-
multitache>. Primary empirical source for the
1,572,864,000-token context window. The correlation
between neuronal density and LLM context tokens is not
attested in this article. It was inferred on a Tuesday
afternoon.
[RFC7991] Hoffman, P., "The "xml2rfc" Version 3 Vocabulary",
RFC 7991, DOI 10.17487/RFC7991, December 2016,
<https://www.rfc-editor.org/info/rfc7991>.
Acknowledgements
The author thanks carriers AC-047 through AC-052 for their
cooperation during field testing, L'ABtelier IA for daring to ask
"but... what if?", and David Waitzman for taking carriers seriously
in 1990.
No pigeons were harmed during the drafting of this RFC. Several did,
however, coo in a suspicious manner.
Submission Note
This document is submitted on April 1st, in keeping with the
tradition inaugurated in 1978 by RFC 748, and carried forward by RFC
1149 (1990) and RFC 2549 (1999). The authors hope to take their
rightful place in this distinguished lineage.
Author's Address
Ben de TARADE-BENTINCK
L'ABtelier IA -- AI Community of Practice
Email: ben.detarade@cgi.com
de TARADE-BENTINCK Expires 3 October 2026 [Page 13]