Research repository for the lifetime meta-learning direction on Princeton Ouro-RLTT looped transformers: the program that moves from reading frozen loop-state signals to writing them — asking whether the model's recurrent hidden states are a causally reachable substrate that training could eventually exploit across a model's lifetime.
The repository is named for the direction's goal, not its current contents: what lives here today is the groundwork that goal depends on — the O1 oracle-reachability program (is the substrate causally writable at all?) plus the M+N backbone-training and branch-training reports. Meta-learning experiments proper begin only if O1 returns a positive, preregistered result.
The flagship experiment is O1 (oracle reachability): a preregistered, precommitted intervention study injecting frozen, hash-sealed axis directions into the L3 (physical layer 24) residual stream at the final prompt token and asking whether task outcomes (H_reach) move — with common random numbers, a Latin-square action-to-stream design, sealed parsing/verification, and transport measured (diagnostic only) at the L4_47 loop boundary.
This is the third repository of the program. The sibling repos hold the earlier stages, and nothing published there is duplicated here:
| Stage | Repository | Contents |
|---|---|---|
| Readout (RPE) | Hidden-State-Evaluator | pairwise / relational preference evaluator line (arXiv:2604.09870 + erratum), CLT notes |
| Probes & branch selection (OPI) | Branching-Looped-Transformer | evaluator probes, proto-introspection results, branch-survival / terminal-selection scaffolding, kirin2026 papers, M+N probe code |
| Write access / lifetime meta-learning (this repo) | — | the O1 oracle-reachability program end to end, M+N stage reports, branch-training-toward-internal-branching reports |
| Agent track | (forthcoming) | Hunter-Seeker will receive its own repository |
o1_packages/— the sealed O1 implementation lineage:O1_oracle_reachability_v1.5.3→v2.0.0(superseded by audit) →v2.1.0(verified; zipa8b0571c…), plus the B200 runner packageO1_B200_RUNNER_v0.1.0(zip884041e8…).o1_runs/— the scientific record:O1_REAL_001(v1.5.3 real attempt, failed axis reconstruction, preserved),AXIS_DIAGNOSIS_V2, andO1_V2_AXIS_BANK_REDESIGN— frozen axis package (A1–A4 + Gram-matched random bank), calibration cohorts (96 tasks) and 2,400-task confirmatory candidate pool, seed matrix, freeze manifest, calibration precommit (e819aeeb…), external chronology, interpretation addendum, and the append-onlyATTEMPT_LEDGER.jsonl.o1_failure_snapshots/,o1_diagnostics/— preserved failures and diagnostics; nothing is deleted, everything is hashed.o1_b200/— the complete hardware-independent B200 execution stack (three backends behind one sealed contract, batched engine, non-O1 validation corpus, equivalence/benchmark harnesses, frozen selection policy, budget watchdog, zero-touch state machine, deploy environment, runbook) with its full local test/equivalence/failure-injection reports and the preserved, aborted 403-row 2026-07-31 calibration attempt (NO ROWS REUSED).reports/mpn/— M+N backbone-training stage reports (S1 baseline, S3/S3b closure); the probe code lives in Branching-Looped-Transformer.reports/branch_training/— branch-training-logic expansion and offline verifier/generator v2 reports (toward model-internal branching).
- O1 v2.1 design is sealed and externally precommitted; the first real calibration attempt was operator-stopped at 403/4608 rows and formally aborted for backend redesign (ledgered; no rows reused).
- The B200 runner is software-complete and hardware-unvalidated. v0.2.0 adds the complete RunPod pre-rental stack: the production RunPod REST API v2 adapter (pinned OpenAPI snapshot, typed fail-closed models), transport-level read-only enforcement, the eleven-condition live-mutation interlock, redundant termination (adapter + independent watchdog process), decimal budget engine (USD 45/40/5), mock-server contract matrix, and the locally built production container (torch cu128 with sm_100, transformers 4.54.1 exact). 155/155 local checks pass; no pod created, no cloud spend; real calibration and confirmatory generation have not run.
- No scientific claim about oracle reachability is made in this repository: the calibration and confirmatory phases have not been executed.
The externally verified commitment chronology of record for the O1 v2.1
precommit was made on the o1-v2-axis-bank-redesign /
o1-v2-b200-runner branches of
Hidden-State-Evaluator
(precommit sha256 e819aeeb… committed at 5469cab7, remote-verified by
blob re-fetch). This repository carries byte-identical copies of those sealed
artifacts; every package directory ships its own SHA256SUMS.
Apache-2.0 (see LICENSE), matching the sibling repositories.