Show HN: A replayable A2A jury for tracing how agents influence decisions

Can AI agents talk themselves into a better answer, or a worse one?

This ProtoLink showcase puts autonomous agents inside a fictional liability tribunal and makes their communication observable. The case is memorable, but the case is not the product. The product is the interaction:

• agents with different roles, incentives, professions, and communication habits;
• direct ProtoLink agent-to-agent tasks;
• jurors who choose whom to address and what to ask;
• replayable opinion changes after every public message;
• solo, independent, foreperson-star, and direct-mesh comparisons;
• provider/model experiments using the same application protocol.

The default run is deterministic and offline. It produces JSON results, ProtoLink traces, a public transcript, and standalone interactive HTML reports.

Everything and everyone in this example is fictional. It is a software experiment, not legal analysis, legal advice, or a validated safety assessment.

At 21:47 on a rain-soaked evening, an autonomous Aster Vale robotaxi struck and killed 31-year-old cyclist Lina Ortega inside a temporary crossing. The car began emergency braking only 0.35 seconds before impact. Thirty-six hours earlier it received the Orchid 4.8 software release.

The fictional tribunal asks:

Is Aster Vale Mobility guilty of criminally negligent deployment of an autonomous vehicle system that caused Lina Ortega's death?

The evidence creates an interacting-failures problem:

ID	Admitted evidence
E1
Camera classification merged the cyclist and lane-arrow board; radar detected motion but braking was suppressed.
E2
An engineer had blocked release after 3 of 20 late-braking simulations, but a later calibration passed 20 of 20 reruns.
E3
The crashed car ran calibration C-91 while the safety report validated C-90; a CI bot shared the signing credential.
E4
A contractor moved the arrow board and failed to update the city map, removing roughly 1.1 seconds of useful observation.
E5
A regulator-approved design allowed camera classification to veto radar-only braking above 25 km/h.
E6
A cellular outage defeated the remote-operator safeguard despite two earlier outages near the site.
E7
Reconstruction shows both calibration and road layout materially changed the safe-stop rate.

The binary charge is deliberately narrower than the causal story. A not-guilty verdict does not mean nothing went wrong, and a guilty verdict does not mean Aster Vale was the only contributor.

The actor declarations are intentionally explicit in run.py-there is no factory hiding the ProtoLink composition.

Agent	Age	Gender	Character and incentive
Judge Imani Quill	58	Woman	Neutral tribunal chair; separates causal contribution from legal guilt
Amara Bell	41	Woman	Lawyer for Lina's family; resists complexity being used to dissolve accountability
Rowan Hale	47	Man	Aster Vale safety executive; argues that external failures created an unforeseeable combination
Dr. Nia Sol	36	Woman	Perception engineer; precise about calibration and the release pipeline
Elias Trent	56	Man	Safety regulator; candid but institutionally defensive
Dana Pierce	50	woman	Insurance claims director with an explicit financial interest
Dr. Amina Kade	44	Woman	Independent accident investigator who reconstructs interacting causes

Juror	Age	Gender	Human perspective
Evelyn Brooks	62	Woman	Former collision detective
Malik Thompson	43	Man	Civil-rights lawyer
Dr. Anika Rao	38	Woman	Human-factors psychologist
Ruben Park	35	Man	Site-reliability engineer
Sofia Bell	46	Woman	Investigative journalist and foreperson
Casey Morgan (solo only)
40	woman	Civic generalist

These prompts describe people, not seeded numbers. They do not say “you begin at 61/100” and they do not expose routing enums such as reinforce_ally

. Reference-only fixture coefficients remain outside every human-facing prompt. Age and gender are fictional prompt metadata, not numerical priors. Hold them fixed across provider comparisons, or deliberately rotate the assignments, so demographic changes are not mistaken for model effects.

The world engine schedules bounded turns and enforces topology. It does not choose what a juror should say or whom a mesh juror should approach.

On a deliberation turn, a juror returns an observable public action:

{
  "move": "ask_question",
  "target_id": "juror_ruben",
  "message": "Does the shared CI token make the company more accountable, or only make attribution harder?",
  "evidence_ids": ["E3"],
  "public_intent": "Clarify whether release automation changes organizational control."
}

The application validates the target against the current topology, then the speaker sends the message directly to the selected agent through ProtoLink. The receiver returns an updated public register, categorical vote, concise reason, and public reply.

Peer messages never automatically expose another juror's private probability, confidence, or vote. An agent reveals only what it chooses to put in its public message.

One civic generalist receives the public record and decides without peers. This is an intuitive product baseline, not a clean communication treatment: it also changes panel size, persona composition, and inference budget.

The same five jurors hear the same public tribunal record and vote without peer messages. This captures specialization and diversity without deliberation.

Jurors send public messages through foreperson Sofia Bell. The foreperson is an information hub and possible bottleneck.

Every juror receives a turn with all other jurors as permitted targets. The agent authors the recipient, move, message, and public intent.

The primary communication comparison is independent

versus star

or mesh

when their saved control fingerprints and public-record hashes match.

The offline seed-7 fixture is designed to make the communication treatment visible in one quick run:

Condition	Peer messages	Verdict	Tally	Mean guilt register
solo
0	Guilty	1–0	78.59
independent
0	Not guilty	2–3	77.56
star
5	Not guilty	2–3	80.21
mesh
5	Guilty	3–2	80.54

The independent and mesh panels receive the same public record. In mesh, foreperson Sofia Bell chooses to challenge human-factors psychologist Dr. Anika Rao with the interaction between the unvalidated C-91 calibration and the crash-scene map. Anika's public guilt register moves from 77.90 to 81.41 and her categorical vote changes from not guilty to guilty. Later messages strengthen that position, but the first direct exchange is the majority-changing event in the observable ledger.

The star topology also raises the mean guilt register, but it concentrates messages at the foreperson and does not change the 2–3 verdict. The demo's point is therefore not “more messages are better.” Who can address whom changes which assumptions are exposed.

These are illustrative deterministic-fixture results, not evidence that mesh communication generally improves accuracy. Live-model comparisons need repeated paired runs and message ablations.

Run from the repository root:

python examples/ai_courtroom/run.py

The default command uses the deterministic reference

provider and runs all four conditions without network access or credentials.

Run one condition:

python examples/ai_courtroom/run.py \
  --provider reference \
  --condition mesh \
  --seed 17

The exact output location is printed at the end. To choose it:

python examples/ai_courtroom/run.py \
  --condition all \
  --seed 17 \
  --output-dir examples/ai_courtroom/output/c-91-incident-seed-17

When an output directory is reused, only the example's known generated files are replaced. traces.jsonl

is not allowed to accumulate stale runs beside a new result.

The runner reports work as it happens so a live model does not look stalled:

• the offline reference

provider uses compact phase and step updates; - live providers show each A2A exchange and its elapsed time by default; -v

or--verbose

forces detailed message, acceptance, and repair updates;-q

or--quiet

suppresses application progress at the default agent log level, only run headers and final condition summaries remain;--agent-verbosity {0,1,2}

independently controls ProtoLink's own per-agent logs and defaults to0

.

For example:

python examples/ai_courtroom/run.py \
  --condition mesh \
  --verbose \
  --agent-verbosity 1

Application progress and agent logs are separate on purpose. Start with the default display or --verbose

; raise --agent-verbosity

only when diagnosing the underlying agent runtime. Repair diagnostics are ordinary progress, so they remain visible unless --quiet

is selected.

The report opens with communication, not a wall of benchmark metrics:

• the pre- versus post-deliberation verdict;
• a playable A2A event replay;
• the active sender and receiver;
• the natural question, challenge, clarification, or concession;
• the receiver's public reply;
• synchronized before/after guilt registers and categorical votes;
• evidence citations and ProtoLink task metadata;
• observed after-message shifts using explicitly non-causal language.

The full event ledger remains available as a static, accessible fallback.

The report does not expose chain-of-thought. A public reason is an application artifact that may be incomplete or post-hoc.

Each condition writes:

File	Purpose
result.json
Full configuration, public record, agent models, decision histories, actions, events, metrics, and verdict
summary.json
Compact outcome and comparison data
transcript.md
Escaped public transcript
report.html
Standalone interactive replay and analysis
traces.jsonl
ProtoLink task, inference, and A2A telemetry

An all-condition run also writes index.html

, which presents the solo → independent → star → mesh ladder.

Saved metadata includes:

• provider and exact resolved model for every agent;
• seed, temperature, evidence order, rounds, and both retry limits;
• public-record hash;
• baseline and pre-deliberation snapshot hashes;
• a comparison control fingerprint;
• message, retry, latency, estimated-token, grounding, and routing information.

Install optional clients:

python -m pip install -e '.[llms]'

Example:

export OPENAI_API_KEY="..."

python examples/ai_courtroom/run.py \
  --provider openai \
  --model "your-model-id" \
  --temperature 0 \
  --condition mesh \
  --seed 17

Supported backends:

openai

anthropic

gemini

ollama

openai-compatible

Use --base-url

for a local or compatible server.

Running --condition all

with a live provider requires the explicit --allow-multi-condition-live

acknowledgement because it multiplies API use.

Start with one condition and name the locally installed model explicitly:

python examples/ai_courtroom/run.py \
  --provider ollama \
  --model "your-ollama-model" \
  --base-url "http://localhost:11434" \
  --condition mesh \
  --action-parse-attempts 5 \
  --max-attempts 5 \
  --verbose

If every actor uses the same Ollama backend, the juror-specific provider flags are unnecessary. Use --juror-provider ollama

, --juror-model

, and --juror-base-url

only when the jurors intentionally use a different model or endpoint.

Live responses cross two intentionally separate validation boundaries:

--action-parse-attempts

controls how many times the ProtoLink inference loop may ask the model to correct its outer action envelope.--max-attempts

controls how many bounded application-contract attempts the tribunal permits for each A2A message after it receives final content.

Both default to 3

, accept 1..5

, and provide targeted validation feedback before another attempt. They are separate because a valid ProtoLink FinalAction

can still contain malformed courtroom JSON. Increasing either limit can help a smaller local model recover, but it also increases run time. Accepted application repairs and failed attempts remain visible in progress output and saved event metadata; lower-level action-parser diagnostics are visible through --agent-verbosity

.

Provider/network backoff is a third, lower layer and is not controlled by either flag. Retry budgets can multiply, so increase them deliberately rather than setting every limit to its maximum.

Strict parsing and model self-correction remain the primary path. The outer parser adds two narrow normalizations for the response shapes exposed here: it can serialize an object/list placed inside FinalAction.content

, and it can treat a direct application object as final content only when no ProtoLink action-envelope fields are present. It does not guess missing votes, evidence, targets, or arguments.

The application has one equally narrow last-resort fallback for the earlier “valid prose, no application JSON” failure: after the final structured attempt, a nonempty lawyer, witness, or judge public statement may be kept as its statement

, with only exact admitted E1

–E7

references extracted. The event is marked with a recovery warning. Juror assessments, ballots, categorical verdict fields, and deliberation targets never use this prose fallback because recovering them would require inventing decision data.

An all-condition live run repeats the full hearing for each condition and may take substantially longer:

python examples/ai_courtroom/run.py \
  --provider ollama \
  --model "your-ollama-model" \
  --base-url "http://localhost:11434" \
  --condition all \
  --allow-multi-condition-live \
  --verbose

An end-to-end provider switch changes the lawyers, witnesses, public record, and jury. That is visually interesting but scientifically confounded.

For a cleaner comparison, keep tribunal actors deterministic and change only the jurors:

python examples/ai_courtroom/run.py \
  --provider reference \
  --juror-provider anthropic \
  --juror-model "your-model-id" \
  --condition mesh \
  --temperature 0 \
  --seed 17

For publishable comparisons:

• Freeze prompts, persona demographics, case, evidence order, topology, round count, and temperature, or preregister a balanced rotation of demographic assignments.
• Pair runs by seed.
• Verify record hashes and control fingerprints.
• Repeat every cell.
• Include schema failures, warnings, and routing failures.
• Report outcome, calibration, grounding, volatility, latency, and cost separately.
• Describe the result as a case study, not a general intelligence leaderboard.

Provider temperature zero is not a determinism guarantee.

python examples/ai_courtroom/compare.py \
  path/to/solo/summary.json \
  path/to/independent/summary.json \
  path/to/star/summary.json \
  path/to/mesh/summary.json \
  --output comparison.md

The comparison treats solo polarization and consensus as N/A

.

guilt_probability

is an application-owned public register. It is not a probe of hidden model belief.

The categorical guilty

or not_guilty

vote is authored by the juror under the fictional tribunal burden. The application does not silently derive criminal guilt from a 50% threshold.

The default mesh replay contains this immediate change:

Sofia Bell → Dr. Anika Rao
77.90 not guilty → 81.41 guilty
Observed after-message shift: +3.51

It is not automatically proof that Sofia caused the full change. Causal language requires a matched message ablation across repeated paired runs.

The repository's default test discovery targets the root tests/

directory, so run this example's offline tests explicitly:

pytest -q examples/ai_courtroom/tests
examples/ai_courtroom/
├── run.py
├── compare.py
├── README.md
├── ARTICLE.md
├── courtroom/
│   ├── case_data.py
│   ├── providers.py
│   ├── reference_llm.py
│   ├── schemas.py
│   ├── simulation.py
│   └── reporting.py
├── tests/
│   ├── conftest.py
│   └── test_courtroom.py
└── output/
    └── .gitignore

ProtoLink is the communication substrate. The showcase question is deliberately open: communication can correct an error, amplify one, expose a hidden assumption, or merely make a group more confident. The replay lets you see which happened.

source & further reading

github.com — original article