26. Self-Model
JGI maintains a model of its own abilities.
Capability fields: name, status, reliability, cost, latency, authority, known limits, dependencies, recent evidence.
| Status | Meaning |
|---|---|
| UNKNOWN | Capability not yet assessed |
| DISCOVERED | Capability observed but not tested |
| EXPERIMENTAL | Under test, not trusted |
| PROVEN | Verified on held-out cases |
| RELIABLE | Proven with sustained reliability |
| DEGRADED | Reliability falling |
| REVOKED | Authority withdrawn |
Never assume capabilities from documentation alone.
27. Capability Gap Detector
When missions repeatedly fail because of the same missing ability, create a CapabilityGap.
Fields: affected missions, frequency, cost, human burden, workaround, potential reusable solution.
System leverage:
Leverage = Frequency x Cost x Breadth x ReuseProbability
High-leverage capability gaps become build candidates.
28. Self-Improvement Engine
JGI may improve: prompts, skills, workflows, retrieval, tests, routing, memory compression, automation, context compilation, agent allocation, tool adapters, failure recovery.
It may NOT automatically expand: legal authority, financial authority, security privileges, protected consequences.
Self-improvement loop:
OBSERVE FRICTION -> QUANTIFY -> REUSE COURT -> PROPOSE CHANGE -> BUILD ISOLATED -> BENCHMARK -> A/B AGAINST OLD -> SENTINEL -> PROMOTE WINNER -> RETAIN ROLLBACK
29. Self-Improvement Requires Benchmarks
No system improvement is accepted because "the new prompt seems smarter."
Measure: success rate, cost, latency, founder attention, false closure, Sentinel rejects, recovery success, economic outcome.
Require measurable improvement.
30. Shadow Mode
Before granting a newly improved decision system authority, run it in SHADOW MODE.
It predicts what it would do. Existing system continues operating. Compare. Only promote after sufficient evidence.
This applies especially to: new scoring, new routing, new autonomous authority, new memory promotion, new models.
31. Canary Autonomy
Authority expansion occurs gradually.
Example:
read 1 repo -> read all repos -> create isolated worktree -> mutate isolated worktree -> prepare commit -> local commit -> PR -> bounded deployment -> broader deployment
Each promotion requires evidence.
32. Autonomy Credit Score
Maintain a reliability-based autonomy score for action classes.
Possible inputs: verified successes, Sentinel failures, recovery success, rollback success, authority incidents, false closure, calibration.
More reliable actions earn wider bounded authority.
One category's success does not grant unrelated authority.
The paper lists possible inputs but gives no formula and no thresholds. The calculator below uses an illustrative formula, stated next to the tool.
59. Procedural Skill Compiler
Repeated successful traces become procedural skills.
A skill candidate requires: multiple examples, independent verification, clear preconditions, defined outputs, failure conditions, rollback, authority boundary.
Test on held-out cases.
60. Skill Reliability
Use Beta posterior.
Initialize: alpha = 1, beta = 1 Verified success: alpha += 1 Verified failure: beta += 1 Reliability estimate: alpha / (alpha + beta)
Use conservative confidence bounds for important autonomous skills.
61. Skill Mutation
Skills may evolve. Never overwrite proven skill immediately. Create skill_v2. Benchmark against skill_v1. Promote winner. Keep rollback.
62. Innovation Forge
JGI continuously searches for things the founder did not think to request.
- What do we have that nobody realizes is valuable?
- What new market becomes possible if two existing capabilities are joined?
- What manual workflow can collapse by 90%?
- What information do we see earlier than others?
- What internal capability could become a product?
- What unused data could become intelligence?
- What tool could be generalized?
- What business could exist now that could not exist six months ago?
104. Agent Benchmark Lab
Maintain internal eval suites. Benchmark capabilities like: repo reconstruction, bug fixing, lead research, market analysis, data reconciliation, long-horizon recovery, planning, business evaluation.
Every major model, prompt, or skill upgrade runs against evals.
105. Generalization Tests
To claim increasing generality, test skills on domains not used during development.
Example: skill trained on real-estate research, test on company diligence, infrastructure diligence, historical provenance.
Cross-domain performance is stronger evidence of generality.
106. JGI Capability Ladder
These are internal engineering levels. They are not scientific AGI declarations.
| Level | Name | Meaning |
|---|---|---|
| L0 | REACTIVE | Responds to explicit prompts |
| L1 | TOOL-USING | Executes tools |
| L2 | PERSISTENT | Maintains missions |
| L3 | SELF-STARTING | Discovers tasks |
| L4 | ECONOMIC | Prioritizes value |
| L5 | MULTI-AGENT | Creates temporary organizations |
| L6 | SELF-RECOVERING | Survives interruption |
| L7 | LEARNING | Improves from outcomes |
| L8 | SELF-IMPROVING | Modifies own procedures under eval |
| L9 | CROSS-DOMAIN | Transfers capabilities |
| L10 | ORGANIZATIONAL | Creates businesses, workflows, teams autonomously |
| L11 | LONG-HORIZON | Pursues goals across days or weeks with low supervision |
| L12 | GENERALIST | Strong performance across broad task families |
107. General Intelligence Scorecard
Measure multiple dimensions: breadth, depth, transfer, planning horizon, tool competence, memory, recovery, calibration, novel problem solving, economic performance, self-improvement, human-attention independence.
No single benchmark proves general intelligence.
The paper lists the dimensions but gives no scoring bands. The scorecard tool below uses illustrative bands, stated next to the tool.
108. Founder Independence Metric
FounderInterventionRate = interventions / missions
Also track FounderMinutesPerVerifiedOutcome.
Goal: continually reduce both without increasing risk or false completion.
109. Autonomy Horizon
Measure: longest duration JGI can operate productively without founder intervention. Increase this gradually.
Minutes -> hours -> overnight -> days
Never fake autonomy by doing useless work.
115. Meta-Learning
Periodically analyze: which task types are failing? Which context packs work? Which agents perform? Which planning methods win? Which verification detects issues? Which tasks create real value?
Then improve the orchestration itself.
Five Levels of Learning
The eventual system should not merely learn tasks. It should learn at five different levels simultaneously.
| Level | Name | Question |
|---|---|---|
| 1 | FACT LEARNING | What is true? |
| 2 | PROCEDURAL LEARNING | How do we reliably do this? |
| 3 | STRATEGIC LEARNING | Which things create value? |
| 4 | ORGANIZATIONAL LEARNING | Which agents, tools, and workflows should solve which problems? |
| 5 | META-LEARNING | How should JGI itself reason, allocate resources, verify, remember and improve? |
Level 5 is where this becomes materially different.
Three flywheels:
Flywheel 1: MORE MISSIONS -> MORE OUTCOMES -> BETTER CALIBRATION -> BETTER TASK SELECTION -> BETTER WORKER ROUTING -> BETTER SKILLS -> LOWER FOUNDER ATTENTION -> LONGER AUTONOMY HORIZON -> MORE MISSIONS Flywheel 2: FAILURE -> EVIDENCE -> ROOT CAUSE -> CAPABILITY GAP -> IMPROVEMENT -> BENCHMARK -> SKILL OR SYSTEM UPGRADE -> FUTURE FAILURES REDUCED Flywheel 3: WEIRD ASSET + UNSOLVED PROBLEM + CROSS-DOMAIN ANALOGY -> INNOVATION -> CHEAP PROOF -> CUSTOMER OR MARKET OUTCOME -> NEW PRODUCT, VENTURE, OR CAPABILITY
JGI is not one model that is really smart. It is a persistent cognitive organism whose effective intelligence compounds because its state, evidence, tools, workers, economic feedback, procedures and self-improvement survive every individual model invocation.
Interactive Tools
All six tools compute real results from your input, using the paper's own terms. Where the paper gives no formula, the tool says so.
Tool 1. Autonomy Credit Score Calculator (Section 32)
The paper lists possible inputs but gives no formula and no thresholds. This calculator uses an illustrative formula: each verified success earns 2 points, each recovery success and each rollback success earns 1 point, each Sentinel failure costs 5, each authority incident costs 10, each false closure costs 8. The total is multiplied by calibration (0 to 1) and clamped to 0 to 100. Verdict bands are illustrative: 70 or above EXPAND, 40 to 69 HOLD, below 40 SHRINK. Credit is per action class. One category's success does not grant unrelated authority.
Tool 2. Shadow Mode and Canary Autonomy Simulator (Sections 30 and 31)
The paper requires "sufficient evidence" before promotion but gives no numbers. The gates below are illustrative thresholds: shadow to canary needs at least 20 shadow runs with at least 95 percent agreement with the existing system. Canary to full promotion needs at least 95 percent canary success and zero authority incidents. A canary success rate below 80 percent, or any authority incident, rejects the behavior back to shadow.
Tool 3. Skill Reliability Tracker (Sections 59, 60, 61)
Paper formula, exact: alpha starts at 1, beta starts at 1. Verified success adds 1 to alpha. Verified failure adds 1 to beta. Reliability = alpha / (alpha + beta). Skills never overwrite a proven skill immediately; a mutation becomes skill_v2 and is benchmarked against skill_v1.
Tool 4. Benchmark Lab Runner (Sections 29, 104, 105)
Capabilities from Section 104. A benchmark passes when the observed success rate meets its threshold and at least 5 runs were made. Overall verdict: READY when all 8 pass, PARTIAL when 5 to 7 pass, NOT READY when fewer than 5 pass. The generalization checkbox records the Section 105 cross-domain test; a READY verdict without it is flagged as single-domain only.
| Capability | Pass threshold (%) | Observed success (%) | Runs | Result |
|---|---|---|---|---|
| Repo reconstruction | PASS | |||
| Bug fixing | PASS | |||
| Lead research | PASS | |||
| Market analysis | PASS | |||
| Data reconciliation | PASS | |||
| Long-horizon recovery | PASS | |||
| Planning | PASS | |||
| Business evaluation | PASS |
Tool 5. General Intelligence Scorecard (Sections 107 and 108)
Dimensions from Section 107, each scored 0 to 10. Overall score is the mean. The paper gives no scoring bands; the bands here are illustrative: 8.0 or above Strong, 5.0 to 7.9 Developing, below 5.0 Early. Section 108 formulas are exact: FounderInterventionRate = interventions / missions, plus FounderMinutesPerVerifiedOutcome.
Tool 6. Five Levels Tracker (Five Levels addendum)
Add a learning loop, set the level it is actually at, and check the evidence that advances it. Advancement evidence is an illustrative implementation of the addendum: Level 1 to 2 needs repeated successful traces with verified preconditions; Level 2 to 3 needs outcome data showing value; Level 3 to 4 needs routing evidence across agents, tools, and workflows; Level 4 to 5 needs evidence that the orchestration itself improved. Level 5 has no further level.