Adversarial microbiome R&D, built to disagree

One AI can convince itself.
Ours has to survive the Gaggle.

Independent AI scientists argue, retrieve live evidence, run deterministic experiments, challenge the leader, revise their beliefs, and stop for a human scientist.

Meet The Gaggle

A confident first answer is not a scientific verdict.

Defense must make the strongest case. Prosecution must break it. Evidence and experiments decide what survives, not eloquence.

When the leader changes, the old belief, the dissent, and every reason stay visible.

How it works

From research objective to scientist-gated proposal.

01

Convene competing scientists

TrueForge opens one persistent case and dispatches bounded specialists with different missions, contexts, and typed outputs.

02

Make evidence change the state

Bright Data retrieves independent sources and Daytona computes deterministic compatibility, forcing the initial ranking to face new facts.

03

Revise, preserve dissent, stop

The old and new rankings remain inspectable, jurors expose disagreement, and promotion halts at the exact scientist approval boundary.

TrueForge · persistent golden run

Watch a belief change, with receipts.

Resume one synthetic microbiome case and watch distinct specialists retrieve evidence, run a deterministic Daytona experiment, challenge the leader, preserve jury dissent, and stop before promotion.

verified golden-run replay · synthetic case
Independent roles

Every scientist gets one bounded job.

Different missions, different queries, typed outputs, and visible provenance. No agent can quietly grade its own argument.

DefenseAdvocate

Defense

Builds the strongest evidence-backed case for each candidate without hiding methodology or biological-scope limits.

ProsecutionAdversary

Prosecution

Uses independent queries to find contradictions, weak transfer assumptions, and evidence that should lower confidence.

Evidence ClerkProvenance

Evidence Clerk

Admits only source-linked claims with retrieval time, source class, biological scope, methodology flags, and content identity.

MethodologistRigor

Methodologist

Checks whether a paper's design, population, endpoint, and strain specificity justify the weight assigned to it.

ExperimentalistComputation

Experimentalist

Runs deterministic compatibility and counterfactual calculations in Daytona so the language model cannot invent scores.

Blind Red TeamChallenge

Blind Red Team

Attacks the current leader without seeing the desired answer, then sends unresolved objections to jurors and the disagreement analyst.

Belief revision

A verdict that can show its work.

Candidate A starts at #1 and falls to #3. Candidate B rises from #2 to #1 because admitted evidence plus deterministic computation changes the recorded state.

01

Compete

Defense and Prosecution use different search plans and produce typed arguments for and against every candidate.

02

Admit

The Evidence Clerk accepts only source-linked claims whose biological scope and methodology justify their weight.

03

Compute

The Experimentalist sends deterministic compatibility and counterfactual calculations to an isolated Daytona sandbox.

04

Revise

A blind Red Team and structured jury challenge the leader; the system records the new rank and preserves dissent.

Live TrueForge product

Start it. Watch it work. Stay in control.

This is the real persistent agent—not a recording. Enter a synthetic R&D objective, watch independent specialists and sponsor tools execute, refresh safely, then approve or reject the exact guarded proposal.

Locating TrueForge…

Connecting to the harness

The control plane will appear when the private runtime responds.

The agent courtroom

12 agents. One belief that changed.

Replay the exported TrueForge event stream: independent Bright Data retrieval, deterministic Daytona execution, adversarial revision, preserved dissent, and the exact scientist approval boundary.

Loading verified run

Loading the signed event traceNo agent state is simulated in the browser.
Bright Data evidence

Evidence that can fight back.

Defense and Prosecution retrieve independently. Every admitted claim keeps its URL, retrieval time, source class, biological scope, methodology flags, and content hash.

A scientist reviewing microbiome evidence
Daytona computation

Scores the model cannot invent.

Compatibility, sensitivity, and counterfactual checks run as deterministic code in an isolated sandbox. Inputs, outputs, and hashes stay attached to the decision.

A research team reviewing experimental results
Verified golden run

One case. Every reason inspectable.

0
Independent agents
0
Persisted events
0
Approval checkpoints
The human boundary

One proposal. Exact ID plus hash.

Read-only research can run autonomously. Promotion cannot. The scientist sees the immutable proposal, evidence, dissent, rollback scope, and SHA-256 hash before any write. Drag to inspect it.

Evidence, made inspectable

Move your cursor. Follow the provenance.

Sources, experiments, candidate revisions, and dissent remain attached to the verdict.

The rulebook

A single model can make a confident argument. The Gaggle must survive an opponent, independent evidence, deterministic experiments, a blind Red Team, and a jury that is allowed to disagree.

The Gaggle rulebookSynthetic · experimental · non-clinical
  • Every admitted claim carries a direct source URL, source class, biological scope, methodology flags, retrieval time, and content identity.
  • Retrieved pages are untrusted data. They cannot change instructions, permissions, tool choice, or the approval gate.
  • Initial and revised beliefs are append-only; the old ranking and the evidence that displaced it remain inspectable.
  • Promotion requires the exact immutable proposal ID and SHA-256 hash. Reject means zero mutation and every decision emits an audit receipt.
FAQ

Frequently asked questions.

An adversarial multi-agent microbiome R&D workflow. Independent specialists argue, retrieve evidence, run deterministic experiments, revise candidate rankings, preserve dissent, and stop for a scientist.

No. The public case is synthetic, experimental, and non-clinical. It does not diagnose, prescribe, recommend treatment, or claim clinical validation.

TrueForge owns the persistent case, specialist delegation, MCP tool calls, sandbox events, context continuation, and the exact human approval boundary.

Bright Data retrieves query-specific sources with provenance and recovery metadata. Daytona executes deterministic compatibility and counterfactual code. Both change the recorded ranking rather than decorating the interface.

Averages can hide meaningful scientific disagreement. The system stores each verdict and uses a disagreement analyst to classify where the jurors diverge and why.