Skip to content
Aletheonix

AletheonixResearch prototype v0.2

What if AI could discover what we haven’t thought to ask?

Aletheonix is a research prototype for self-directed discovery. It watches a system, decides which unexplained change deserves investigation, tests competing explanations in a sandbox, and accepts only what an independent verifier confirms.

Status
Research prototype. Not a product.
Evidence
A closed synthetic benchmark. No production results are claimed.
Scope
Structured question and hypothesis spaces. No language model in the loop.
Follow one investigationMethodology and limits
.00.25.50.751.00d0d4d8d12expected 0.670observed 0.491deploy crm_lookup 1.4→1.5change h175incidents → agent_betaincidents → crm_lookup
Fig. 1Benchmark S17, seed 0. Each dot is the evaluator score of one task handled by agent_beta on refunds (281 tasks over 14 days); the grey line is all other traffic. Nobody asked about this agent. Incident reports named both the agent (2) and the tool (7); they could not say which one caused the drop.Synthetic world · regenerated from the published scenario and seed · not production data

01One investigation · benchmark S17

From a signal nobody asked about to a change nobody may make alone.

In S17, incident reports pointed at an agent and at a tool. The system was given no question. It had to notice the drop, decide it mattered, and establish which component caused it. Every value below is recorded output from that run.

Benchmark S17 · seed 0 · autonomous condition · run_main @ 5c0ccc6Synthetic benchmark · documented output

01 · Observe

A metric moved, and nobody had asked about it.

Agent performance falls after a deployment: eval score for agent_beta on refunds tasks drops from 0.670 to 0.491.

14 incident reports in the same window: 2 attributed to agent_beta, 7 to crm_lookup, the rest to other agents. Attribution alone could not separate cause from symptom.

Observation record

signal · Fig. 1 above
metric
eval_score
entity
agent_beta · refunds
expected
0.670
observed
0.491
traffic
7.4% affected
context
a deploy aligned with the change point

Incident attribution · same window

14 reports
  • agent_beta · 2
  • crm_lookup · 7
  • other agents · 5

Attribution points at both. Only an experiment can separate cause from symptom.

Belief · tool version · crm_lookup44%

The belief is the posterior of the leading hypothesis. It stays at its prior until a verified verdict arrives.

02Benchmark results

Results from a closed world, with the limits beside them.

Without being given a question, the autonomous condition found the planted mechanism in 99% of non-null runs and made 0 false discoveries out of 112 claims. These are synthetic results. They show the architecture working under controlled conditions, not performance on production systems.

  • Autonomous
  • Human question
  • Threshold monitor

Planted mechanism found

non-null runs · higher is better

Planted mechanism found by condition
Autonomous99%
Human question96%
Threshold monitor43%

Top-ranked conclusion correct

all runs · higher is better

Top-ranked conclusion correct by condition
Autonomous92%
Human question94%
Threshold monitor36%

False discovery rate

share of claims that were wrong · lower is better

False discovery rate by condition
Autonomous0%
Human question0%
Threshold monitor42%
Fig. 2Three conditions on the same 30 hidden scenarios × 3 seeds. Autonomous: no question supplied. Human question: the same engine, told where to look. Threshold monitor: static alerts, no experiments.run_main · closed synthetic world · research repository @ 5c0ccc6
Per scenario and condition: planted mechanism found in seeds run, and false discoveries
Condition010203040506070809101112131415161718192021222324252627282930
AutonomousS01, Autonomous: found in 3 of 3S02, Autonomous: found in 3 of 3S03, Autonomous: found in 3 of 3S04, Autonomous: found in 3 of 3S05, Autonomous: found in 3 of 3S06, Autonomous: found in 3 of 3S07, Autonomous: found in 3 of 3S08, Autonomous: found in 3 of 3S09, Autonomous: found in 3 of 3S10, Autonomous: found in 3 of 3S11, Autonomous: found in 3 of 3S12, Autonomous: found in 3 of 3S13, Autonomous: found in 3 of 3S14, Autonomous: found in 3 of 3S15, Autonomous: found in 3 of 3S16, Autonomous: found in 3 of 3S17, Autonomous: found in 3 of 3S18, Autonomous: found in 3 of 3S19, Autonomous: found in 3 of 3S20, Autonomous: found in 3 of 3S21, Autonomous: null world, correctly quiet in 3 of 3S22, Autonomous: null world, correctly quiet in 3 of 3S23, Autonomous: found in 3 of 3S24, Autonomous: found in 3 of 3S25, Autonomous: found in 2 of 3S26, Autonomous: found in 3 of 3S27, Autonomous: found in 3 of 3S28, Autonomous: found in 3 of 3S29, Autonomous: found in 3 of 3S30, Autonomous: found in 3 of 3
Human questionS01, Human question: found in 3 of 3S02, Human question: found in 3 of 3S03, Human question: found in 3 of 3S04, Human question: found in 3 of 3S05, Human question: found in 3 of 3S06, Human question: found in 3 of 3S07, Human question: found in 3 of 3S08, Human question: found in 3 of 3S09, Human question: found in 3 of 3S10, Human question: found in 3 of 3S11, Human question: found in 3 of 3S12, Human question: found in 3 of 3S13, Human question: found in 1 of 3S14, Human question: found in 3 of 3S15, Human question: found in 3 of 3S16, Human question: found in 3 of 3S17, Human question: found in 3 of 3S18, Human question: found in 3 of 3S19, Human question: found in 3 of 3S20, Human question: found in 3 of 3S21, Human question: null world, correctly quiet in 3 of 3S22, Human question: null world, correctly quiet in 3 of 3S23, Human question: found in 3 of 3S24, Human question: found in 3 of 3S25, Human question: found in 2 of 3S26, Human question: found in 3 of 3S27, Human question: found in 3 of 3S28, Human question: found in 3 of 3S29, Human question: found in 3 of 3S30, Human question: found in 3 of 3
Threshold monitorS01, Threshold monitor: found in 3 of 3S02, Threshold monitor: found in 0 of 3S03, Threshold monitor: found in 3 of 3S04, Threshold monitor: found in 0 of 3S05, Threshold monitor: found in 3 of 3S06, Threshold monitor: found in 0 of 3S07, Threshold monitor: found in 0 of 3, 2 false discoveriesS08, Threshold monitor: found in 0 of 3, 3 false discoveriesS09, Threshold monitor: found in 0 of 3S10, Threshold monitor: found in 3 of 3S11, Threshold monitor: found in 3 of 3S12, Threshold monitor: found in 3 of 3, 3 false discoveriesS13, Threshold monitor: found in 0 of 3S14, Threshold monitor: found in 0 of 3S15, Threshold monitor: found in 0 of 3S16, Threshold monitor: found in 3 of 3S17, Threshold monitor: found in 3 of 3, 4 false discoveriesS18, Threshold monitor: found in 0 of 3S19, Threshold monitor: found in 3 of 3, 3 false discoveriesS20, Threshold monitor: found in 0 of 3, 6 false discoveriesS21, Threshold monitor: null world, correctly quiet in 3 of 3S22, Threshold monitor: null world, correctly quiet in 3 of 3S23, Threshold monitor: found in 0 of 3S24, Threshold monitor: found in 3 of 3, 3 false discoveriesS25, Threshold monitor: found in 0 of 3S26, Threshold monitor: found in 3 of 3, 4 false discoveriesS27, Threshold monitor: found in 0 of 3, 1 false discoveriesS28, Threshold monitor: found in 3 of 3, 1 false discoveriesS29, Threshold monitor: found in 0 of 3S30, Threshold monitor: found in 0 of 3
  • fill = share of seeds where the planted mechanism was found
  • null world: nothing planted; fill = correctly reported nothing
  • at least one false discovery
Fig. 3Every scenario, every condition. The autonomous condition fell short only on S25 (found in 2 of 3 seeds). The threshold monitor never found the mechanism in 16 of the 28 non-null scenarios and raised false alerts in 10.run_main · scenarios S01–S30 · hover a cell for counts
verdicts signed by the verifier
705verdicts signed by the verifier
held back as inconclusive, most often because a reproduction disagreed
128held back as inconclusive, most often because a reproduction disagreed
experiments per autonomous run: the cost of verification
7.8experiments per autonomous run: the cost of verification
false discovery during development (S27), recorded and fixed
1false discovery during development (S27), recorded and fixed

03Architecture

Six components. None grades its own work.

  1. 01Observerobserve
  2. 02Belief Graphexpect · predict
  3. 03Opportunity Scorerdetect · rank
  4. 04Hypothesis Generatorask · explain
  5. 05Experiment Enginedesign · run
  6. 06Verifiercheck · sign
  7. The Experiment Engine needs a governance grant and runs only in a sandbox. Only the Verifier’s signed verdicts flow back into the Belief Graph.
Fig. 4The loop as implemented in v0. Questions and hypotheses come from a typed taxonomy, not a language model. Governance and sandbox are interfaces to separate systems; in v0 a local policy and a synthetic world fork stand behind them.Architecture of research prototype v0.2 · full detail on the System page

  1. Generation is not verification.

    The component that proposes an explanation cannot sign the verdict on it.

  2. No discovery without evidence.

    Beliefs change only on signed verdicts from preregistered experiments.

  3. Goal generation does not imply authority.

    Deciding what to investigate grants no right to act on the answer.

Every component, its invariants and its limits

04Today and ahead

What exists, what is being tested, what is still research.

v0 chooses among questions inside a structured space: five detector families and a typed taxonomy of hypotheses and levers. It is not open-ended scientific intelligence. Choosing questions outside a fixed space is the research problem, not a result.

Implemented
  • Six-component loop with explicit invariants
  • Verifier-gated belief updates with signed verdicts
  • Preregistered, hashed experiment designs
  • Sample-size planning and power checks for claims of absence
  • Fresh-sample reproduction by the verifier
  • Confounder, measurement and multiplicity checks
  • Side-effect check on every proposed intervention
  • Memory of discoveries and null results across runs
  • Hash-chained audit log
  • Governance interface with a local policy; LucidRail adapter contract
  • Sandbox environment interface
  • OpenTelemetry importer and observe-only mode
  • Synthetic benchmark: 30 hidden scenarios, 3 conditions
Being tested
  • Whether findings hold up on real agent systems, scored with a sealed-list protocol
  • How much of the loop survives real traces, real noise and an imperfect sandbox
  • Replay and shadow-traffic sandboxes for real systems
Still research
  • Model-generated questions and hypotheses behind the same verifier
  • Open-ended hypotheses beyond a fixed lever catalogue
  • Valuing what is worth knowing, not only what is surprising
  • Calibrated predictions of intervention impact
  • Learning across systems
  • A read-only track record that governance can use to scale autonomy
  • Physical experimental environments

From “given a goal, solve it” to “given an environment, determine what is worth understanding.”

05Three independent systems

Discovery. Authority. Reality.

Separate systems with separate responsibilities, connected through defined interfaces. None is a subsidiary of another. Knowing something does not grant the right to act on it.

Aletheonix

Discovery

What should we understand?

Finds what is worth investigating, tests competing explanations and verifies the evidence.

Proposes experiments and interventions. Never grants itself permission.

You are here

LucidRail

Authority

What are we allowed to do?

Decides whether a consequential action may happen, under which grant, and records it.

Adapter contract defined. In v0 a local policy stands in behind the same interface.

lucidrail.ai ›

Terranoux

Reality

How do we test it in the physical world?

Bounded physical environments for experiments: laboratories, robotics, sensors, materials.

Sandbox interface defined. No physical environment is connected to Aletheonix today.

terranoux.com ›

Aletheonix discovers. LucidRail governs. Terranoux acts.

06Field-study partners

Bring a system worth understanding.

The benchmark cannot answer the research question. Only a real system can. We are looking for teams running AI-agent systems with at least two weeks of traces and, ideally, a replay or shadow environment.

The protocol is fixed in advance, and a negative result is reported as a negative result.

Become a field-study partnerRead the full protocol
  1. 01Seal the listBefore any run, you write down every problem you already know about. The list is hashed and stays closed.
  2. 02Run without a questionObserve-only first. With a sandbox, the full loop runs under the same budget and thresholds as the benchmark.
  3. 03Open the list and scoreKnown problems found, new and real, new and wrong, missed, cost. You see everything.