verbatim · source-checkedSource-checked agent context.
Smaller, and still verifiable.
Until now you had to choose: summaries are cheap but nobody can check them, full transcripts are checkable but expensive. PAIRL, the context format for business conversations between agents, gives you both — every excerpt copied word for word and verified, every line with its speaker, and around 60% fewer tokens than the raw history on long business conversations in our benchmark.
pairl compilerlive
Officially recognized research project in Germany — certified and publicly funded under the German Research Allowance (Forschungszulage), assessed by the BSFZ federal certification body.
We run on it ourselves: our in-house agent stack and our own products plinio® GEO and plinio® TL route their LLM traffic through the PAIRL Gateway.
Summaries lose who said what. PAIRL keeps it.
What a summary does to an agent conversation, and how PAIRL makes the same context smaller by quoting instead of rewriting. Under a minute, no sound needed.
Try it with your own textFar smaller than the transcript. More faithful than a summary.
Summaries save tokens but rewrite what agents said — afterwards nobody can tell what is original and what the model filled in. PAIRL cuts just as much context by quoting instead of rewriting: checked verbatim excerpts with the speaker attached. Your context shrinks, and when someone asks “why did the agent do that?”, you still have the answer.
Nothing invented
The model decides what matters; plain code copies it from the source byte for byte and drops anything it can’t find. The guarantee lives in code, not in a prompt — it holds whichever model does the work. This covers conversation history; older tool results are shortened by rules.
Who said what, kept
Every turn carries its speaker, set from your own metadata — never guessed by a model. That is exactly where summaries fail in multi-party conversations: they mix up who committed to what.
Around 60% smaller
On long agent business conversations our benchmark measured around 60% less context, at the same coverage as an LLM summary. What that does to your bill depends on your traffic: tokens your provider already serves from its cache cost little, so removing them saves little.
Long tool sessions stay small too
In agent runs, most of the context is old tool output: files read once, command results, search hits. PAIRL compacts older results once and then leaves them alone, so your provider’s prompt cache keeps working. Recent steps reach the model unchanged, and the agent can re-run a tool whenever it needs the full output.
Available for tool sessions on Anthropic today. On OpenAI and Google, requests with tools pass through unchanged, with no fee.
# Database host=db.internal port=5432 pool_size=20 # Feature flags flag.checkout_v2=true flag.checkout_v3=false …
#ret status=ok lines=214 compressed=pairl # Database host=db.internal port=5432 pool_size=20 # Feature flags flag.checkout_v2=true (+11 similar lines omitted) (older tool output compacted; facts above are verbatim — re-run the tool if you need the full output)
We tested our own claims — including the ones that failed
Every claim on this page comes from benchmarks we registered before running them, on agent business conversations and real multi-party interviews. The full paper, with methods, prompts, data and negative results, is in journal peer review and will be published as soon as the review allows.
Ask about the benchmarkWho said what, answered right
On 100 multi-party interviews, named speaker markers beat LLM summaries on attribution, using about 3.7× the context of the summary — and beat handing the model the full, uncompressed transcript by 11.9 points.
Tokens cut, coverage kept
60% fewer tokens end to end with the same coverage as an LLM summary — and about half as many invented details in the answers that follow.
Fewer errors than dropping tokens
At the same reduction, token-dropping compression (LLMLingua-2) loses negations, numbers and qualifiers. PAIRL produced 4.4× fewer material hallucinations.
Measurable cost of being checkable
At matched reduction, the checked format answers as well as an unchecked one. The difference: only the checked one can be verified afterwards.
Scope: one generated agent-business corpus (79 sessions) and one public interview corpus (100 transcripts), three runs each, one judge model. The paper will state the full limitations.
A protocol for agent conversations that hold up in review
Your existing OpenAI and Anthropic workflows keep running. PAIRL standardizes how agents hand context to each other: names, numbers, dates and sources travel unchanged, history travels as marked verbatim excerpts, and a quote can always be told apart from a paraphrase. Threading, origin and cost ride along in every message.
Your keys. Your models. Your provider.
The protocol is open and published. We bill a share of the tokens we measurably remove, and nothing at all when a request doesn’t compress.
Three steps to smaller, checkable context
Connect PAIRL
Point your application at the gateway with a single URL change. Run with compression off to capture your baseline. No risk, no surprises.
Compression runs automatically
No configuration. PAIRL condenses history into checked excerpts where it pays off and passes requests through unchanged where it doesn’t. If compression fails, your request still goes through, unchanged and free of charge.
See what changed
Your dashboard shows token reduction, the value of those tokens at list price, and the PAIRL fee — per request, per day, per month. Export anytime.
Private beta, terms agreed per pilot
PAIRL is in private beta. We set up each pilot with you and agree the price up front. You keep your provider keys and models, and you keep paying your provider directly.
Request accessIn a regulated industry? Keep your AI traffic on EU soil.
For finance, healthcare, the public sector and other compliance-bound teams, we deploy a dedicated PAIRL gateway in sovereign EU infrastructure — full data residency, your own isolated environment, recipients agreed by contract, and the same one-line setup. Your data stays in your jurisdiction, and you can show what your agents based their answers on.
See what your agent gets to read
Paste a conversation and see what PAIRL keeps: verbatim excerpts tied to who said them — names, figures and decisions in, small talk out. Nothing is reworded. Open the prepared example right away, or confirm your email to run your own text on the real engine.
The built-in example is free and unlimited. To run your own text, confirm your email and get 5 runs — once, no account needed. Demo engine: Google Gemini, routed through the PAIRL Gateway. In production PAIRL is provider-agnostic: Anthropic, OpenAI and Google, with your own keys. Inputs are forwarded to Google, so please don't enter personal or sensitive data. For production use, request access.
Open protocol. Drop-in libraries.
PAIRL is an open specification. Encode, validate, and decode messages natively in your own stack with the reference libraries — same format the gateway speaks, no lock-in.
The open libraries handle the PAIRL format itself. The hosted gateway adds advanced compression techniques of our own that go well beyond what a self-hosted implementation can reach — substantially more token savings, with the same one-line setup. See how it works.
Everything you need to evaluate PAIRL
Is PAIRL production-ready?+
What does “source-checked” mean here, and what doesn’t it?+
Why not just summarize the history, or drop tokens?+
How long does integration actually take?+
How much can I realistically save?+
How is PAIRL different from TOON or Headroom?+
- TOON works at the syntax layer — a leaner, lossless way to spell JSON, best for uniform tables (~30–60% vs JSON). A data format you reach for in your own payloads.
- Headroom works at the runtime layer — a proxy/middleware that compresses whatever bytes flow to the model (tool outputs, logs, RAG chunks) and can fetch elided detail back on demand.
- PAIRL works at the semantic layer — it defines what’s actually being communicated. Facts that must stay correct (names, numbers, dates, sources) are kept losslessly, and conversation history is carried as verbatim quotes with marked omissions — compressed by cutting, not by rewriting. It also carries conversation threading, provenance, and cost as part of the message.
What if PAIRL can’t compress a request?+
Do you train on our data?+
Do I have to change my provider or models?+
The protocol is open — what does PAIRL add?+
Source-checked context for your agents
One URL change. Verbatim excerpts, named speakers and less context to pay for. PAIRL is in private beta.