Canonical experiment:
agi_generality_reference_20260718_v3
Command:
cortex agi prove \
--experiment-id agi_generality_reference_20260718_v3 \
--cycles 96 \
--sessions 6
Result:
master status: ok
operational prerequisite suite: supported
AGI achieved: false
AGI claim supported: false
| Gate | Result | Primary evidence |
|---|---|---|
| Cross-domain transfer | ok |
0/3 baseline, 0/3 sham, 3/3 learned |
| Novel compositional invention | ok |
3/3 depth-four programs across 3 novel, pairwise-distinct sketches |
| Persistent autonomous curriculum | ok |
96 autonomous task decisions |
| Continual learning | ok |
final accuracy 1, average forgetting 0 |
| Bounded long-horizon autonomy | ok |
6 processes, 5 restarts, 0 human interventions |
Additional readback:
source/holdout domain-family overlap: 0
source/holdout primitive-ID overlap: 0
equal transfer search budget: true
depth-two invention controls: 0/3
verified depth-four inventions: 3/3
novel pairwise-distinct sketches: 3
interfering proposals rejected: 1 in sequential suite
campaign regression proposals rejected: 5
promoted regressions: 0
controlled transient faults: 3
recovered transient faults: 3
verified campaign skills retained: 93/93
campaign retained accuracy: 1
campaign ledger events: 108
campaign ledger valid: true
learner requests: 110
learner receipts: 110
independent verifications: 110
boundary audit findings: 0
evidence files hashed: 447
Evidence root digest:
c048d8e7235b9605b62492982586d19e93a59a6d7068f226fc56809d3ad0ae6d
The following validations passed on the packaged source tree:
npm run check
node --test \
test/agi_evidence_suite.test.js \
test/neuromorphic_learning.test.js \
test/agi_loop.test.js \
test/software_repair_benchmark.test.js
npm run build:all
Results:
syntax check: pass
targeted proof and regression tests: 22/22 pass
canonical full runtime build: pass
master report JSON Schema: pass
autonomous state JSON Schema: pass
all 447 evidence file hashes: pass
runtime source hashes in manifest: pass
The new AGI evidence test file contains seven positive and negative tests. It checks the transfer ablation, contaminated-request rejection, ledger-tamper detection, novel program synthesis, public-only curriculum selection, the full five-gate integration run, and the schema-enforced claim boundary.
The run demonstrates the five operational edges in a finite symbolic program-synthesis environment. It does not demonstrate open-world AGI, indefinite autonomy, or independent external replication. Those fields remain false in the master schema and report.