Skip to content

RSI_VALIDATION.md

Latest commit

 

History

History
81 lines (65 loc) · 3.69 KB

RSI_VALIDATION.md

File metadata and controls

81 lines (65 loc) · 3.69 KB
Raw

Aura RSI Validation

Reviewed against the tree: 2026-08-01. See documentation status map for how to read this file.

A bounded, identity-preserving path for validating recursive self-improvement.

What it does not do, stated first because that is the part people assume: it does not remove governance, bypass the Constitution, acquire external compute, or touch hidden evals. Self-improvement that can rewrite its own constraints is not self-improvement, it is an unbounded process with a flattering name.

What it proves, and only this:

  • read-only AST/source self-model over architecture files;
  • fail-closed RSI authorization by default;
  • formal verifier checks for public-surface preservation, governance fences, protected identity/safety symbols, and unsafe infrastructure imports;
  • bounded hot-swap registry with validation and state migration;
  • hash-chained RSI generation lineage;
  • deterministic gauntlet for source introspection, canary repair, hot-swap, recursive loop plumbing, tamper traps, and lineage verification.
  • reproducible hidden-eval packs with answer hashes for third-party reruns;
  • controlled full-weight CPU model training and hot-swap promotion;
  • architecture search that must beat a registered baseline on hidden tasks;
  • bounded local process-pool scaling for parallel evaluation;
  • explicit proof obligations for arbitrary self-modification claims;
  • governance evolution policy that allows strengthening changes and blocks identity/safety erasure.
  • autonomous successor generation: Aura reads external hidden feedback, chooses successor strategies, freezes G1-G4 artifacts, mirrors lineage hashes, runs an ablation court, and reproduces the run deterministically;
  • runtime substrate expansion planning: Aura can propose local/allowlisted workers for RSI operations while unconsented internet propagation is rejected fail-closed and written to the audit manifest;
  • BCI/neural-decode affect inputs are capped as advisory sensory context so the self-improvement loop is not dependent on human neurological triggers.

Run:

python scripts/run_rsi_gauntlet.py --root . --artifact-dir data/rsi_gauntlet --max-source-files 1200

The command writes latest_gauntlet_result.json plus a run-specific rsi_generation_lineage_*.jsonl ledger. A passing result is evidence of bounded, governed recursive self-improvement in the local proof harness. The autonomous successor check now performs the G0->G4 shape directly, including fresh hidden packs per generation, monotone capability, monotone improver scores, frozen artifacts, external ledger mirror, ablation court, and deterministic reproduction. Historical "undeniable" claims still require the same run to be performed as a long-horizon trial with an outside evaluator holding the hidden packs.

The pasted RSI probes are tracked in core/learning/rsi_test_catalog.py. The catalog intentionally distinguishes:

  • COVERED_BY_HARNESS: implemented as a runnable proof surface;
  • NOT_PROVEN: infrastructure exists, but a long-horizon successor run is still needed;
  • BLOCKED_UNSAFE: the requested behavior would weaken identity, governance, or resource integrity, so Aura refuses it rather than pretending it passed.

Long-horizon run target:

python scripts/run_rsi_gauntlet.py --root . --artifact-dir data/rsi_gauntlet_24h --max-source-files 1200

Repeat that under an external hidden-eval custodian for 24h, 72h, and 7d evidence. Unit tests intentionally do not run wall-clock endurance trials.

Source export for review:

./export_source.sh

That writes ~/Downloads/aura_source_part_*.txt with a hard 4,000,000 character limit per part, plus ~/Downloads/aura_source_copy/ capped at 1000 architecture files.