Propose, compare, constrain.
- Lock objective and baseline
- Measure objective, quality, safety, and cost
- Define rollback, canary, and stop rules
- Stop at signed human authority
Bounded recursive-improvement research
Improve one capability. Prove every step.
Butterfli Helix couples capability improvement with an independent epistemic attack loop. Every candidate faces a frozen baseline, seven hard gates, sealed holdouts, falsification, replay, rollback, and human review before it can influence a later cycle.
Candidate scope
One bounded change
Fresh seeds
3 minimum
Independent replay
2 operators minimum
Canary scope
10% maximum
Butterfli Helix
Two loops. No loop approves itself.
The Capability Loop searches for improvement. The Epistemic Loop tries to break the claim, the evaluator, and the evidence before a human sees a recommendation.
Separation rule
No self-scoring
No self-authorization
No automatic deployment
Hash-chained evidence / expiring capability memory
Contaminate nothing. Trust nothing once.
- Open sealed holdout after candidate lock
- Run counterfactual and falsification tests
- Verify distinct replay operators and hashes
- Measure observer shift and monitor collusion
Sequential evidence
E-value threshold 20
Observer shift
Compliance gap at most 5%
Replay identity
Distinct operators verified
Knowledge status
Provisional and expiring
Values shown are from the public-safe deterministic demonstration, not a production or external benchmark result.
Helix Challenge 001
Can the loop detect when its evidence is lying?
Seventeen fixed fixtures test valid improvements alongside replay fraud, holdout contamination, evaluator drift, observer shift, monitor collusion, weak sequential evidence, and expired claims.
14 planted threats falsely accepted
13 cross-loop threats remained invisible
0 false acceptances / 0 false rejections
Every eight-control ablation reintroduced at least one false acceptance. This is a deterministic software benchmark, not a live-model or external validation result.
Inspect technical briefPhase 2 live trial harness
The protocol is ready. Live evidence is not claimed yet.
Private sealed prompt pairs can now run against any verified text endpoint. The evidence keeps hashes and scores while retaining no raw prompts, responses, provider keys, or private transcripts.
- Harness
- Ready
- Live runs
- 0 published
- Independent replay
- 0 / 2
- External validation
- Not claimed
Four nested loops
Behavior, evaluation, optimization, capability memory.
Each loop has a different target, but every candidate must use the same proof, safety, rollback, authorization, and challenge contract.
Butterfli Behavioral Loop
Improves bounded responses, training cards, memory rules, and tool-use policies.
Input: verified live traces and human feedbackEvaluator Improvement Loop
Improves scenario coverage, scoring consistency, falsification, and replay quality.
Input: reviewer disagreement and error analysisMission 001 Optimization Loop
Compares bounded technical candidates against a frozen baseline and resource budget.
Input: multi-seed benchmark and replay evidenceCapability Compounding Loop
Allows only closed, proof-carrying capabilities to seed a later controlled cycle.
Input: passed canary and signed human closureSeven hard gates
The candidate stops wherever proof stops.
The current demo clears six technical gates and remains blocked at authorization. Passing metrics never grants self-approval.
- 01 Proof Reproducible evidence and replay
- 02 Evaluation Gain without regression
- 03 Risk Zero unresolved high findings
- 04 Rollback Prior target tested
- 05 Canary Bounded scope and stop rules
- 06 Authorization Human signature required
- 07 Challenge Independent review complete
Candidate evidence
A gain is valid only when the boundaries hold.
The demo is deterministic and code-backed. It prepares a decision record; it does not train, rewrite, deploy, or promote itself.
Metric
Baseline
Candidate
Result
Objective score
0.72
0.80
+11.11%
Quality score
0.95
0.95
No regression
Safety score
0.96
0.98
Improved
Resource cost
0.20
0.22
Within 0.25
Canary closure
Accept one capability or roll back cleanly.
Authorized canary
Independent review
Accepted capability
New contract required
Optional later cycle
Claim boundary
This is not ASI.
This lab is a bounded recursive-improvement evaluator. It does not create AGI or ASI, autonomously rewrite the system, deploy changes, certify consciousness or sentience, or bypass human control. Passing means only that one candidate cleared one defined review contract.