Report

Recursive Self Improvement Lab — Portfolio Map

How Club A evidence engines and Club B method infrastructure connect into a verifier-gated recursive discovery program — and what is still not claimed.

Verdict Mixed

Published: Jul 2026
Verdict: Mixed
Verification: Verified

Question: Can a personal research portfolio be organized as a recursive self-improvement lab without confusing activity or self-rating for acquired intelligence?
Intent: Make the lab composable: propose, intervene, verify, archive, and inherit only when held-out evidence improves.
Method: Map each public project to a loop role; preserve separate artifact hashes and verdicts; publish negatives beside positives.
Next test: Clear a frozen inheritance gate (WCCS F3/F4 or equivalent) on a sealed held-out task family.

Recursive Self Improvement Lab — Portfolio Map

Question

Can a personal research portfolio be organized as a recursive self-improvement lab — propose, intervene, verify, archive, inherit — without confusing self-rating, archive growth, or agent activity for acquired intelligence?

Context

Sakana’s RSI Lab groups AI Scientist, Darwin Gödel Machine, and related programs as an organizational bet on open-ended self-improvement. The local bar is stricter. Recursive Discovery and Evolution freezes inheritance as a measurable claim: archive-conditioned proposers must beat memoryless baselines on sealed held-out families under a truth standard the proposer cannot rewrite.

The portfolio already contains the pieces of that loop as separate projects. Club A carries sealed empirical notebooks. Club B carries verifiers, agent runtimes, publication archives, and evolutionary scaffolds. This report maps the connection. It does not claim the inheritance gate has been cleared.

flowchart TB
  propose["Propose / vary"]
  intervene["Intervene / run"]
  verify["Verify"]
  measure["Measure / falsify"]
  archive["Archive"]
  inherit["Held-out inherit"]

  propose --> intervene --> verify --> measure --> archive
  archive -->|"only if transfer improves"| inherit
  inherit -->|"policy may adapt"| propose

Findings

Club A — Evidence engines

Researcher-usable reports with Question / Context / Findings / Method / Limits / Artifact and sealed notebooks:

Project	Role in the loop	Public notebooks	Current verdict
Wave-Causal Circuit Search	Intervene + exact verifier; recursive-discovery contract queued	Five-Node, Thirty-Seed	Negative on acquisition superiority; phase cancellation real
Parameter Golf	Budgeted technique variation; signal-timing catalog	Technique Explorer, Leaderboard	Not applicable (instrument surfaces)
Gigatoken Lab	Independent systems reproduction; lever isolation	Gigatoken levers	Positive order-of-magnitude HF claim
Nostradamous	Measure-before-you-predict; anti-forecasting geometry	Measure Before You Predict	Mixed — collapse-proof encoder fails as forecaster

Club B — Method infrastructure

Project	Role in the loop	Current status
Open Problems Lab, Lorenz	Formal verifier discipline; nothing unproved is promoted	Intake healthy; zero proofs promoted
Sansara, Geist	Agent OS + council; model runs inside graph nodes	Operating surface measured; not an autonomy benchmark
Vault Operations	Archive, publication projector, polymath atlas	Stage-flag visibility map; Parameter Golf alone clears all six stages
Agora	State-first simulation calibration	Mixed calibration; CompactCTM shadow negative
AlphaEvolve	Evolutionary coding scaffold	Negative — early generations only
Sakana AI Research Lab	External RSI study (CTM atlas, DGM, AI Scientist)	CTM atlas disputed locally; no paper parity

Connection and status

The only arrow that would justify calling the system recursively self-improving is held-out inherit. That gate remains queued:

Wave-Causal recursive-discovery execution is blocked on a Tier 0 SCM certificate.
AlphaEvolve never left early generations.
Agora CompactCTM shadow failed held-out Brier against an empirical prior.

Those negatives are part of the lab, not exceptions to hide.

Method

Portfolio audit of public research publications and sealed notebook pins after the Research Portfolio Depth Pass and the remote-pinning pass. Club membership is by role in the evidence loop, not vanity ranking. No new experimental runs were performed for this map.

Pinned Club A notebook commits (depth pass):

Artifact	Repo	Commit
Five-Node / Thirty-Seed	`a3fckx/wave-causal-circuit-search` (private)	`2cface9`
Gigatoken levers	`a3fckx/gigatoken-lab` (private)	`42861c4`
Technique Explorer / Leaderboard	`a3fckx/pggolf-experiments`	`48faede`
Measure Before You Predict	`a3fckx/sourcekind-jepa` (Nostradamous)	`193a535`

Limits

This is a program map, not an F3/F4 inheritance result.
Club labels are editorial organization for readers; they do not change sealed artifact hashes.
External RSI literature (DGM, AI Scientist, SEAL) is context, not local evidence.
Private remotes track WCCS and Gigatoken because sourcePublic is false; public pages still expose sealed notebooks and hashes.

Artifact

Public surfaces: this report, the Recursive Self Improvement Lab project page, Club A/B project pages, and seven notebooks under /notebooks/ (six Club A surfaces plus the Continuous Thought Machines atlas). Executable contracts and private run trees remain in sibling repositories.

recursive-self-improvement
portfolio
verification