RSI MetaForge Core

A research runtime that lets a program propose, test, and adopt its own improvements — and only keeps the ones a sealed verifier can't reject.

The question this repository is built to answer is not "can a system look like it's improving?" but "is each improvement actually validated, or just plausible?" Everything here exists to make that distinction measurable.

---

How it works

Every change — a synthesized program, a new macro, a forged primitive — goes through the same loop. Nothing is kept unless it survives hidden, sealed gates the search never sees.

flowchart LR
    A[Propose<br/>candidate] --> B[Run in<br/>sandbox]
    B --> C{Sealed<br/>gates}
    C -->|pass| D[Adopt &<br/>record lineage]
    C -->|fail| E[Reject &<br/>roll back]
    D --> A
    E --> A

Loading

Guardrail	What it prevents
Hidden evaluation	Answers aren't on disk — the search can't peek.
Frozen baseline	Every gain is measured against a fixed reference, not itself.
Rollback on failure	Speculative changes that fail finalization are reverted.
Anti-cheat gates	Train-only fits and weakened verifiers are rejected, not counted.

---

Results at a glance

A staged research program (Phases 0–J) generalized these gated rules to the top-level searcher — and then to the instruction set itself — measuring every result on frozen instruments fixed before the runs. Each arm was reproduced twice, byte-for-byte.

Extended ISA (Phase J)  ████████████████████████████   28 / 33   ← T29 · T30 crossed their certificates
Adaptive (live)         ██████████████████████████     26 / 33   ← T15 · T27 · T28 solved together
Frozen baseline         ███████████████████            19 / 33
                        └────────────────────────────┘
                         designer tasks solved on the frozen Phase 0 instrument

A separate domain — real file tasks. The same gated discipline runs on CSV / log / config / repo files, scored by sealed hidden A/B on unseen eval seeds (the hidden expectations never touch the workspace):

File-world (csv_normalize · log_aggregate · config_migrate · repo_repair)
Adaptive         ████████████████████████████   1.000   ← mean over the four families
Frozen baseline  ██████                          0.204

Per-family breakdown in EVIDENCE.md.

• +7 tasks over the frozen baseline in Phase I, then +2 certified-infeasible tasks in Phase J — none lost, every arm deterministic (two byte-identical runs each).
• Open, and labeled honestly: T18 · T21 · T22 remain unsolved (missing vocabulary); T31 · T32 remain certified out of reach at the certified horizon even in the extended ISA; T29 · T30 crossed — each base-ISA impossibility certificate is committed beside its adopted solution.
• Predictions were registered before the final runs and scored after — misses included, in SEQUENCING_RESULT.md and CROSSING_RESULT.md.

Phase J: the certified boundary, crossed. The closure certificates define where the base ISA ends; Phase J extended it through the same gate discipline and converted two certified-infeasible tasks into gate-adopted, instrument-verified solutions. From the gap analysis (ISA_GAP_J.md), a two-primitive extension — BCAST (constant broadcast, lemma-justified) and ZGT (elementwise order test) — was frozen and user-approved in ISA_EXTENSION_SPEC.md; grants stay dormant by default (the incumbent configuration reproduces the committed Phase I artifact byte for byte) and became permanent only through the unchanged A/B + sealed-holdout discipline. Against pre-registered predictions (PREDICTIONS_J.md) and instruments frozen before any run, the crossing arm solved 28/33 (digest 1b36ff714b128546, two byte-identical runs): T29 fell at wave 1 with the certified-minimal 6-token program, and T30 fell at wave 2 by composing the granted ZGT with an exploration-origin macro mined from the frozen Track 2 archive — a route no designer witness used. T31/T32 remain honest nulls with quantified bands; each crossed task's base-ISA impossibility certificate and its adopted solution are committed side by side. Full record: CROSSING_RESULT.md.

---

Domains

The same gated loop is measured across several substrates — not only the integer benchmark:

• Stack VM (33 integer tasks): the core RSI-loop measurement baseline — synthesis, macros, and the extended ISA (--mode demo, --mode test).
• General-domain — list / string / grid / record (Section 11): an oracle-scarce debate pipeline with four strategies — S1 a process scorer (StepScorer), S2 committee debate with an adversarial distinguisher, S3 a learned world model (OpSemanticsModel), and S4 oracle-free meta-learning config selection with a held-out Goodhart audit (--mode directive-battery).
• File-world — CSV / JSON / log / repo: real file I/O scored by sealed hidden A/B; skill proposals mined from failure residues and adopted only through a validation gate (--mode file-battery).
• Self-forge → GD bridge: system-synthesized fold primitives, admitted through sealed gates, are registered as general-domain feature atoms that extend the vocabulary the debate pipeline searches (--mode forge-battery).
• Abstract strategic planning — HDC/VSA + Kuramoto oscillators (Section 14, "measured, not mystified"): an illustrative demo (--mode hdc-rsi) that encodes a symbolic action plan into hypervectors, performs analogical transfer via VSA bind/unbind, and selects steps by oscillator phase-locking; the measured interference-field battery is --mode hdc-battery.
• Self-curriculum — a closed task-generation loop (Phase SC): the system poses its own tasks (each admitted only with a machine-checked feasibility witness, constructed from invertible primitives), solves them against sealed held-outs under an MDL cap, and composes adopted solutions into harder tasks — zero human-authored tasks added. Fourteen anti-gaming invariants each ship with a dedicated in-suite test — twelve construct the attack outright, determinism is proven by two-run digest identity, budgets are literal-pinned and bound into the source pin — and an append-only hash-chained ledger keeps every failure in the denominator (--mode self-curriculum, --mode sc-battery). Claim boundary: this is a closed-world compounding measurement, not an open-ended autonomy claim — transfer to the frozen human instrument is the pre-registered external test (SELF_CURRICULUM_SPEC.md, PREDICTIONS_SC.md, SC_RESULT.md). Phase SC2 makes the fence itself an evolution target: the system invents new task kinds as pure data over a human-frozen five-form meta-checker algebra, taps hash-pinned external corpus material, and spans capsule products (--mode schema-forge, --mode sc2-battery) — with schema admission gated by mutation-kill soundness, discrimination, non-collusion, and novelty checks, and headline credit granted only on measured transfer (I15–I23, SELF_CURRICULUM_SPEC_V2.md, PREDICTIONS_SC2.md, SC2_RESULT.md). Claim boundary, verbatim from the directive: the check-form algebra is the new fence. This directive moves the boundary one level up and measures whether the move pays; it does not remove the boundary.
• Witness-sealed setter–solver — the oracle factory (Ascent Phase K): the system manufactures its own tasks as (task_spec, checker, witness) triples — the checker is a budget-capped VM program, the witness is vault-sealed at emission — and admits a task only through four hermetic certificates: sealed-witness feasibility, strategy-level non-triviality against a null battery and a seeded random-program sampler, descriptor-cell novelty, and frozen-frontier difficulty (the current solver snapshot must certifiably fail the task at the frozen instrument budget). The headline event is a crossing: an admitted task later solved at the same frozen budget with the task's own mined macros excluded — improvement attributable only to library growth from other tasks. The final record certifies one crossing, credited to its setter lineage through the learnability-band reward, with pose allocation visibly following the credit; nine families remain archived as frontier markers, witnesses still sealed (--mode ascent-k, --mode ascent-k-battery; ASCENT_K_SPEC.md, PREDICTIONS_K.md, K_RESULT.md). Claim boundary: one certified crossing at the frozen budgets is the measured claim; the flywheel's throughput ceiling and the credit-vs-supplier gap are recorded findings, not smoothed curves.
• The ascent line, closed end to end (Phases L–O): a single frozen MDL coding formula judges all vocabulary growth (2 adoptions, 203 exact-bit-delta rejections, and MORE certified crossings than raw adoption at identical budgets — L_RESULT.md); a proposal-ordering amortizer trained only on the system's own chain-verified ledgers triples certified passes per budget on sealed holdout families, with a failing retrain rolled back on record (M_RESULT.md); a frozen propositional proof kernel adopts 35 kernel-checked theorems with zero axiom adoptions and countermodel-archived refutations (N_RESULT.md); and the improver itself becomes a citizen — kernel-certified search transformations through one channel, counterfactual A/B adoptions (including an end-to-end ISA-extender protocol from impossibility pre-certificate to crossing evidence) through the other, with rejections and rollbacks on the same record (O_RESULT.md). Directive-level scoring, misses included: ASCENT_RESULT.md.
• The ISA-extension coupling loop — does an open closure revive or re-saturate? (Ascent Phase P): the mechanism a budget ladder cannot substitute for — a coupling loop that feeds K8-gated ISA extension back into K's task factory, so a crossed extension enlarges the closure the setter draws from (extended executor ak_run_tokens_ext, provably byte-equivalent to the untouched base executor at the base ISA; a mechanical request locator retargeted to K's own frontier markers; the Phase J speculative-grant discipline as the frozen admission gate K8; an append-only, hash-chained per-generation closure ledger K10). The pre-registered question — revive or re-saturate — is answered by measurement: from a base start under the frozen two-op catalog the trajectory re-saturates at exactly the frozen-K ceiling (coupling cum-crossed [0,3,3,…] = frozen 3, ISA never grows, at every mining budget), because K's setter never manufactures a task shaped like the capability it lacks — while a designer-stocked positive control certifies the same gate + extended executor DO cross genuine extension-requiring tasks (base 0 → admitted). The instrument budget is set above the extension-crossing search reach, so the null is a closure-shape result, not a budget artifact (--mode ascent-p, --mode ascent-p-battery; ASCENT_P_SPEC.md, PREDICTIONS_P.md, P_RESULT.md). Both outcomes were pre-declared wins; this is the definitive "even an open closure with a self-driven extender saturates" result. Claim boundary: exactly the gate-certified trajectory, bounded by the frozen catalog; no AGI / singularity / unbounded-capability claim.

---

What this is — and isn't

It is	It is not
Bounded, validation-gated self-modification	A general-intelligence claim
A reference harness for verifier discipline	Proof of unrestricted recursive self-improvement
Reproducible, gate-checked, artifact-backed	A drop-in production code-evolution system
Multi-domain verifier (list / string / grid / record / file)	A single-domain benchmark
Cross-substrate lesson transfer (VM → general-domain)	Domain-specific tuning

Full claim boundary and public validation record: EVIDENCE.md.

---

Quick start

Single-file runtime, standard library only — no install step.

# See it run
python rsi_levels_metaforge_unified.py --mode demo

# Reproduce the headline comparison
python rsi_levels_metaforge_unified.py --mode crossing-anchor   # extended ISA (Phase J)
python rsi_levels_metaforge_unified.py --mode transfer-anchor   # adaptive (live)
python rsi_levels_metaforge_unified.py --mode run-frozen        # frozen baseline

# Run the full test suite (270 tests)
python rsi_levels_metaforge_unified.py --mode test

More modes (evidence batteries, help)

python rsi_levels_metaforge_unified.py --mode file-battery      # hidden A/B on real file tasks (csv, log, config, repo)
python rsi_levels_metaforge_unified.py --mode forge-battery     # self-synthesized primitives lifted into the GD pipeline
python rsi_levels_metaforge_unified.py --mode horizon-scan      # closure certificates
python rsi_levels_metaforge_unified.py --mode cfs-battery       # continuous functional substrate search
python rsi_levels_metaforge_unified.py --mode expansion-battery # residue-driven vocabulary expansion
python rsi_levels_metaforge_unified.py --mode grammar-battery   # depth-1 grammar feature search
python rsi_levels_metaforge_unified.py --mode grammar2-battery  # depth-2 grammar feature search
python rsi_levels_metaforge_unified.py --mode self-curriculum   # closed self-curriculum loop (demo)
python rsi_levels_metaforge_unified.py --mode sc-battery        # self-curriculum evidence battery
python rsi_levels_metaforge_unified.py --mode schema-forge      # fence-expansion schema forge (demo)
python rsi_levels_metaforge_unified.py --mode sc2-battery       # fence-expansion evidence battery
python rsi_levels_metaforge_unified.py --mode ascent-k          # witness-sealed setter-solver (demo)
python rsi_levels_metaforge_unified.py --mode ascent-k-battery  # ascent phase K evidence battery
python rsi_levels_metaforge_unified.py --mode ascent-l-battery  # MDL master gate battery
python rsi_levels_metaforge_unified.py --mode ascent-m-battery  # ledger-trained amortizer battery
python rsi_levels_metaforge_unified.py --mode ascent-n-battery  # proof kernel & formal domain battery
python rsi_levels_metaforge_unified.py --mode ascent-o-battery  # meta-improvement (two channels) battery
python rsi_levels_metaforge_unified.py --mode ascent-p          # ISA-extension coupling loop (demo)
python rsi_levels_metaforge_unified.py --mode ascent-p-battery  # ISA-extension coupling loop battery
python rsi_levels_metaforge_unified.py --mode crossing-anchor   # Phase J: live arm + capability-grant channel
python rsi_levels_metaforge_unified.py --help                   # everything else

---

Continuous validation

Two workflows keep the claims honest on every change:

• Quick CI — compiles the runtime, checks the CLI, runs anti-cheat guards. Fast, on every push.
• Full Evidence — runs every evidence battery (including the self-curriculum, fence-expansion, and all five ascent batteries twice each, asserting byte-identical digests and artifact equality) plus the full 270-test suite, then uploads the logs and JSON artifacts. Green means the whole record reproduced on a clean machine.

---

Learn more

Explore interactively on DeepWiki, or read the reviewer docs:

Doc	Answers
Overview	Where do I start?
Architecture	How is it built?
RSI Loop	Where is the self-improvement loop?
Validation Gates	How is cheating prevented?
Evidence Logs	What was actually shown?
Limitations	Where does it stop?
General-Domain RSI	What domains beyond integer tasks are covered? (Section 11, line ~3384)

The Phases 0–I record — frozen instruments, registered predictions, per-phase reports, and final evaluations — lives under docs/, starting from SEQUENCING_RESULT.md. The Phase J record (certified-boundary crossing) runs from the gap analysis (ISA_GAP_J.md) and the frozen, user-approved extension spec (ISA_EXTENSION_SPEC.md) through the frozen evaluation instrument (frozen_holdout_extJ.json), the pre-registered predictions (PREDICTIONS_J.md), and the final result with its committed artifact (CROSSING_RESULT.md, final_live_phaseJ.json).

---

_{Maintained as a research artifact. Results, terminology, and implementation are experimental, bounded, and subject to revision.}