Update README typo

…g for Muon optimizer

From PR openai#1440 + arxiv:2603.09697 "Mousse: Rectifying the Geometry of Muon with
Curvature-Aware Preconditioning" (Feb 2026).

Inserts ~5 lines of diagonal preconditioning before zeropower_via_newtonschulz5
in the Muon optimizer step. Normalizes momentum gradient by row/col norms before
spectral orthogonalization, trace-normalizing the matrix:

  G_pre = G / (||row||_2 * ||col||_2)

Gated by USE_MOUSSE=1, falls back to vanilla Muon when unset. Idempotent via
MOUSSE_MARKER. Anchored on the unique zeropower call which is invariant under
all existing 22 patches.

This is the FIRST shippable finding in 5 research fires that fits our
train_loss metric (optimizer-side change affects training directly, unlike
EMA/Tilt/GPTQ which only affect eval). Subagent recommended PASS due to
medium effort estimate; overrode after confirming PR openai#1440 ships only the
SIMPLIFIED diagonal preconditioning version (5 LOC, not 50-80).

4 MS experiments queued for validation:
  MS0_mousse_alone, MS1_mousse_plus_leaky_ng, MS2_mousse_seed42, MS3_mousse_plus_engram

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

… confirmation),

MR2 promising, PR openai#1430 MERGED at 0.39642 BPB

Subagent reports PR openai#1430 (Per-Sample SLOT + Causal Backoff N-gram Mixer + TTT)
has been MERGED at claimed 0.39642 BPB — 65% below public SOTA. If real, this
fundamentally changes the competitive landscape. Audit fires openai#1-3 all flagged
this PR as likely illegal under issue openai#677. Now MERGED.

NEXT RESEARCH FIRE PRIORITY: deep-dive PR openai#1430 to verify legality and extract
implementation. If real, port it. If leak-based, document it.

Patches 17 (Mousse) and 18 (MuonEq-R) confirmed as known PORTS, not novel-to-comp.
They were always documented as ports in research fires openai#9 and openai#10.

Patches 15/16/21 still uncontested in 120+ open + 10 closed PRs (4 audits in a row).

Pod healthy, ~$2.30/$36 spend. MR2_seed42 = 3.3004 (better than MS2 = 3.3358),
suggesting MuonEq-R may slightly beat Mousse at L5 stack. Falsification of
Patches 17 and 18 proceeding rapidly.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…util

After 5 emergency interventions in 2 hours, the speed fix is finally working:
  GPU Memory: 744 MB -> 3370 MB (4.5x)
  GPU Util: 34% -> 100% (3x, FULLY MAXED)
  Power: 149W -> 218W
  Total compute/step: 270 GFLOP -> 17 TFLOP (64x)
  Total tokens/experiment: 1.5M -> 24M (16x)

CHAMP_L5_seed42 currently running successfully:
  step:100 train_loss:3.6128 step_avg:861ms

The actual root cause was Patch 22 EngramLite init anchor mismatch.
The torch.compile crashes were a red herring — every experiment was
crashing with AttributeError on self._engram_lite_enabled because the
forward apply ran but the init didn't. getattr wrap fixed it.

All prior "neutrality plateau" verdicts are now CONFIRMED INVALID:
Mousse/MuonEq-R/NorMuon/Depth Recurrence/Coprime/EngramLite/QK_GAIN
were all measured on 0.75% of intended data volume. Need re-validation.

PR openai#1430 still OPEN, 24h no activity. Patches 15/16/20/21/25 still novel
(9th consecutive audit confirmation).

NEW finding: TMA Megakernel in 5 PRs (custom Triton kernel, hardware-side).
We have ZERO hardware-side patches. Highest-leverage missing technique.

Spend ~$6.33/$36 (17.6%). Far below $25 flag threshold.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- Cargo workspace with 3 crates + bin/tri-railway
- trios-railway-core: ProjectId/EnvironmentId/ServiceId/DeployId newtypes,
  RailwayHash::seal (R7 audit triplet), Client over Railway GraphQL v2
- trios-railway-audit: DriftCode D1..D7 + DriftEvent + verdict (Gate-2
  PASS criterion), idempotent Neon DDL (railway_projects, railway_services,
  railway_audit_runs, railway_audit_events, v_railway_drift_open)
- trios-railway-experience: append-only L7 writer to
  .trinity/experience/<YYYYMMDD>.trinity (L21-safe, no truncation)
- bin/tri-railway: clap CLI with 'version', 'audit migrate-sql',
  'experience append' (mutating verbs deferred to issues openai#4..openai#9)
- LICENSE Apache-2.0, README, AGENTS.md, TASK.md per crate
- CI: fmt --check + clippy -D warnings + build + test
- Neon DDL applied to neondb (5 objects verified)
- 16 unit tests passing (6 audit + 8 core + 2 experience)
- ascii-only sources, R1 (no .sh / no Python in scripts/)

Anchor: phi^2 + phi^-2 = 3.

Closes #1

Agent: GENERAL

openai#143) (openai#29)

* feat(audit): tri-railway audit run + verdict CLI (closes openai#9, refs openai#143)

Anchor: phi^2 + phi^-2 = 3.

Adds the online-audit subcommands that close L-R14 (Gate-2 verdict)
formally:

  tri-railway audit run     --project <UUID> --target <BPB> [--ledger PATH] [--json]
  tri-railway audit verdict --ledger PATH    --target <BPB>

Behaviour:
  - audit run lists Railway services for a project (Q::project_view),
    converts them to RealService (with seed parsed from name like
    'trios-train-seed-43' or 'igla-final-seed-44'), optionally loads
    a JSONL ledger, calls trios_railway_audit::detect to produce the
    full D1..D7 drift event set, runs verdict() to compute Gate2Pass /
    NotYet / Drift, prints a text summary, optionally JSON, then
    seals one R7 audit triplet to .trinity/experience via the
    existing experience writer. Exit codes:
        0 = GATE-2 PASS   (>= 3 services with bpb < target, no error drift)
        1 = DRIFT         (any error-severity event)
        2 = NOT YET       (no errors, target not yet met)
  - audit verdict is the offline form for cron/CI: takes a JSONL
    ledger snapshot already serialized from Neon, computes the same
    verdict against synthetic services, prints one line, exits with
    the same codes.

Auth fix in trios-railway-core: RAILWAY_TOKEN_AUTH env var allows
forcing 'team' (Bearer) vs 'project' (Project-Access-Token) when the
UUID-shape heuristic guesses wrong. Personal API tokens are also
UUID-shaped but require Bearer; without this override, authenticating
to backboard.railway.com returned 'Not Authorized'. Verified with
both curl variants against the live IGLA project.

R5-honest verification (logs in PR body):
  cargo build --bin tri-railway --locked        : OK
  cargo test  --workspace --locked              : 22 passed, 0 failed
  cargo clippy -D warnings                      : 0
  Live smoke against IGLA (e4fe33bb-...):
    18 services, 16 D1_ORPHAN warnings, NOT YET, exit=2,
    R7 triplet sealed at /tmp/audit-smoke/.trinity/experience/<date>.trinity
  Synthetic ledger smoke:
    3 seeds bpb<1.85 -> GATE-2 PASS, exit=0
    3 seeds bpb<1.85 vs target=1.50 -> NOT YET, exit=2
    1 seed bpb=3.5e38 -> DRIFT (D5_OVERFLOW), exit=1

Closes openai#9. Refs openai#143 (IGLA RACE Gate-2 / L-R14).

* style(audit): cargo fmt --all (CI format-check fix)

---------

Co-authored-by: Perplexity Computer <computer@perplexity.ai>

Triggered by grad_norm=0.07 mid-flight obs in iter 112+122 (well below
grad_clip=1.0 → headroom for deeper backward gradient signal). Backward
coverage at K=16 increases 12.5%→19%.

Iter 85 (H63) PROMOTED stochastic {2,3,4} earlier but at WD=0.30 +
K-jitter (4,6,10) era — different regime. This is fixed=3 retest under
WD=0.01 + K-jitter (16,24) + iter 112+122 (gram=0.1 + softcap=30) baseline.

Expected ~+10% wallclock cost from extra TBPTT step.

Also: Tier 1 reordering (user directive 2026-05-02) — iter 108 (deq_k 16→10)
deprioritized to END of Tier 1 (was launched 04:08, killed at ~step 10,
commit fa22310 reverted). Throughput-only iter with no expected val_bpb
improvement; lower priority than val_bpb-improving + simplification iters.

Files:
  train_gpt.py: Hyperparameters.deq_bptt_k 2 → 3
  CLAUDE.md §5: architecture table mirror
  experiments/hypotheses.md: Tier 1 reordered, iter 108 moved to position openai#9

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>