Evidence, not vibes
Source hashes, span-level excerpts, claim graphs, freshness tracking, and stale-claim quarantine keep every conclusion tied to what was actually observed.
source → claim → hypothesis → run → decision
AUTONOMOUS RESEARCH INFRASTRUCTURE
Give your agents a problem worth staying with.
Evidra is a research harness for long-running investigation and experimentation. It turns questions into evidence, evidence into experiments, and experiments into decisions that hold up.
✓ TypeScript CLI · Codex-native · local, container, or Modal execution
CAMPAIGN / 0047 LIVE
robustness under
distribution shift
distribution shift
OBJECTIVE↑ validation / ↓ variance
PHASE03 / 05 · REPLICATION
evidence graph 14 nodesbudget 02:18:41
SOURCE → CLAIM verified
LANES 04 ↗
INSPECT✳GATHER EVIDENCE✳FORM HYPOTHESES✳RUN EXPERIMENTS✳REPLICATE✳PROMOTE✳
01 / THE DIFFERENCE
Not another agent wrapper.
Most agent tooling is built around a conversation or a single code change. Evidra is built around the work that happens after the first promising answer.
THE USUAL LOOPEVIDRA
Transient contextDurable research memory
One agent, one directionIndependent lanes + cross-pollination
“It seems to work”Replication and evidence gates
Retry the same ideaChange route after failure
Optimize one scoreMetrics, artifacts, proofs, behavior
Stop when the chat endsCheckpoint, resume, and keep going
02 / THE RESEARCH HARNESS
The machinery between
question and confidence.
Reasoning is only one part of research. Evidra owns the deterministic layer that makes autonomous work inspectable, reproducible, and harder to fool.
Search as a team
Focus-aware lanes explore different formulations, exchange compact evidence boards, and spend peer review where disagreement is useful.
Failure becomes memory
Falsification agendas and negative evidence stop the harness from rediscovering the same dead ends.
hypothesis_017 repeated unchanged
Built for the long run
Checkpoints, leases, heartbeats, crash recovery, budgets, and resume semantics keep a campaign alive across machines and sessions.
Trust the result
Evaluator-integrity protection, code-health checks, best-so-far ratchets, paired statistics, replication requirements, and verified-state gates separate progress from self-deception.
✓ evaluator protected✓ replication observed✓ provenance complete
Any measurable outcome
Declare a metric, artifact, proof, behavior, or system property. The harness adapts to the objective—not the other way around.
METRICPROOFARTIFACT
03 / THE LOOP
A system that knows
what to do next.
One durable loop from open question to validated result. Every phase has an internal goal, a budget, and a reason to stop.
01Orientinspect the world
02Discovergather evidence
03Proposemake it falsifiable
04Experimentchange one thing
05Validatereplicate the signal
04 / IN PRACTICE
It works where your work lives.
Start in the terminal. Give it a repository, a challenge, or a research question. Evidra coordinates the rest.
$ evidra
EVIDRA RESEARCH DIRECTOR
› /research start
goal: find a robust improvement without evaluator leakage
◆ Phase 01 · orientation complete
◆ Retrieved 12 sources · 4 claims grounded verified
◆ Allocated 3 independent research lanes
◌ Running experiment_004 · replication working
›
05 / THE TECHNIQUES BEHIND
Built from the ideas
that make research compound.
Evidra brings modern agentic research techniques into one accountable runtime. Each mechanism is bounded, observable, and connected to measured outcomes.
Evolutionary search
Independent islands, family diversity, migration, survivor selection, and evaluator-backed crossover proposals explore more than one path.
Falsification memory
Failed hypotheses become negative evidence. The next cycle must change the mechanism, test, or execution route.
Replay intelligence
Recorded discovery trees can be replayed under alternative policies to compare branch order, cost, parallelism, and Pareto outcomes.
Evidence provenance
Hashed sources, span-level excerpts, claim graphs, freshness checks, and quarantine keep conclusions traceable.
Adaptive allocation
UCB portfolios, information gain, uncertainty, diversity, budget, and evidence pressure determine where the next unit of work goes.
Critic + replication
Independent critics, paired statistics, multi-seed checks, sequential-look correction, and replication gates protect against false wins.
Harness evolution
Changes to Evidra itself are treated as experiments with component snapshots, predictions, matched tasks, and measured retain or revert decisions.
Recovery by design
Timeouts, rate limits, worker crashes, OOMs, and invalid outputs trigger classified, route-changing recovery instead of blind retries.
06 / START HERE
Put the harness
to work.
Evidra is open source and runs from your terminal. Install it globally, authenticate your provider, and start a campaign in minutes.
Read the user guide ↗INSTALL GLOBALLY NODE 22+
$ curl -fsSL https://raw.githubusercontent.com/
StarAtNyte/evidra/master/install.sh | shTHEN
evidra/login codex/research start