feat(gardener): PR-2 live wiring — Client mutations + ledger + serve loop#58
Closed
gHashTag wants to merge 2 commits into
Closed
feat(gardener): PR-2 live wiring — Client mutations + ledger + serve loop#58gHashTag wants to merge 2 commits into
gHashTag wants to merge 2 commits into
Conversation
…r + serve) Refs: #52, #53, #54, #55, #56 Phase 1 of 2 (PR-3 follows with fleet→service_id resolution and Logs subgraph): - crates/trios-railway-core/src/client_ext.rs (#52) Methods on Client: deploy_service, set_vars, redeploy, stop. Each seals an R7 audit triplet via RailwayHash. 4 unit tests cover signatures, hash sealing, and honest error pass-through. - bin/tri-gardener/src/actuate.rs (#53) RailwayActuator trait + MockActuator. apply_decision dispatches each Decision variant to the corresponding mutation. apply_decision_batch honors a shared KillSwitch on every iteration. Cull is honestly Skipped pending PR-3 fleet→service_id resolution. - bin/tri-gardener/src/ledger.rs (#54) LedgerSink trait + PgLedger (tokio_postgres + 3x retry) + MockLedger. build_row attaches outcome + error to the decision JSON. - bin/tri-gardener/src/serve.rs + main.rs Cmd::Serve (#55) serve_loop drives one tick every --interval seconds; SIGINT/SIGTERM graceful stop; drift-free Interval::tick. validate_interval enforces 60..=86_400. - bin/tri-railway/src/main.rs SetVars / Logs / Stop subcommands (#56) SetVars wires Client::set_vars. Stop aliases delete (Railway has no pause). Logs is a deterministic stub until PR-3. - bin/tri-gardener/src/loop_.rs loop_once_live entry point Composes actuator + ledger + KillSwitch. Replaces the warn-stub Live arm. ctx.disabled || kill.is_disabled() short-circuits with a single Skipped row per decision. Tests: 68/68 GREEN across the workspace. New tests: - 4 client_ext::tests - 6 actuate::tests including kill_switch_aborts_mid_batch and live_actuation_writes_to_ledger_via_mock_pair - 5 ledger::tests - 5 serve::tests including serve_emits_tick_every_interval (paused clock) - 4 cli_tests for parse_var_pairs Anchor: phi^2 + phi^-2 = 3 · TRINITY · NEVER STOP
This was referenced Apr 27, 2026
The trainer architecture as currently shipped has a hard floor at BPB ~ 2.19 (champion 2.1919, h=828, 2L hybrid attn, ReLU^2, 81K). Cross-validated against CPU N-gram floor (~2.54) in trios#237 and the live GPU champion in trios#143. Encodes the policy as a public constant in ledger.rs so call sites read it instead of hardcoding 2.19. Two tripwire tests (70/70 GREEN): - architectural_floor_bpb_is_2_19 — locks the value - architectural_floor_below_gate2_target — sanity vs 1.85 Doc-comment policy: gardener MUST NOT issue Decision::CullSeed for a seed whose BPB is above this floor unless plateau is independently confirmed (>=5 ticks in a 0.005 band AND step >= 50_000). Without this guard a healthy seed sitting at the architectural floor would be culled for not crossing 1.85, which is impossible without ALPHA's L1/L2/h=1024 patches landing first. Refs: - gHashTag/trios#237 (CPU N-gram floor) - gHashTag/trios#143 (GPU champion 2.1919) - trios-railway#58 (PR-2 Live arm) - trios-railway#61 (RailwayMultiClient P0) Anchor: phi^2 + phi^-2 = 3
This was referenced Apr 28, 2026
gHashTag
added a commit
that referenced
this pull request
Apr 28, 2026
Local Mac agent set a new architectural record at T+11.5h:
train_v2 seed=42 h=1024 ctx=12 14-gram + weight tying + residual
bottleneck (no attention) AdamW lr=0.002 → BPB=1.8921 @ 94.5K/120K.
Gate-2 (1.85) is NOT yet passed — gap +0.0421 BPB. Quorum-3 below
1.85 still required for Gate-2 OFFICIAL.
This commit:
- bin/tri-gardener/src/ledger.rs
ARCHITECTURAL_FLOOR_BPB lowered 2.19 → 1.89, doc-comment rewritten
to reflect the train_v2 record. New tripwire test
architectural_floor_strictly_below_prior_floor forbids re-raising
the floor above the prior hybrid_attn ceiling.
- bin/tri-gardener/src/leaderboard.rs
default_phase1_expected reflects the pivot:
* 3 new tracking rows for train_v2 (seeds 42/43/44, Railway
portage pending, target Gate-2 quorum-3)
* 9 attention/JEPA rows tagged '(cull-pending: arch lost)'
Test renders_with_zero_samples_and_explains_why now asserts the
champion is visible in the tracking rows ('train_v2', 'BPB=1.8921').
- docs/POSTMORTEM_GATE2_LOCAL_WIN.md (NEW)
Honest post-mortem of why the Railway fleet floored at 2.19 while
a single local agent broke through to 1.89. Four causes named:
(a) architectural ceiling, (b) plan blind to architecture pivot,
(c) telemetry blackout (#61 + #62), (d) capacity+steps+simple beat
9 parallel complex experiments without feedback.
Tests: 82/82 GREEN (added one tripwire). Live tri-gardener once
prints the new tracking rows including the champion.
Refs: #43 #58 #61 #62 #64
Anchor: phi^2 + phi^-2 = 3
gHashTag
added a commit
that referenced
this pull request
Apr 28, 2026
Single source of truth for IGLA-<MODEL_TYPE>-<NUMBER_FORMAT>[-<TAG>]-seed<N>
naming, shared across Rust code, Railway service names, Neon ledger,
and leaderboard rendering.
bin/tri-gardener/src/canon.rs (NEW):
- enum ModelType { JepaT, Nca, Phi, Hybrid, Trinity3K, TrainV2,
TJepa, Muon } with FromStr/Display + architectural_floor_bpb()
mapping per family (TrainV2=1.89, Hybrid=2.19, Phi=2.21, ...).
- enum NumberFormat { Fp32, Fp16, Bf16, Fp8E4M3, Fp8E5M2, DlFloat,
Gf8, Gf16, Gf32, Gf64, GfTern } with FromStr/Display + bits().
- struct IglaCanon { model, format, tag: Option<String>, seed:
Option<u32> } with full FromStr/Display round-trip.
- L-R9 enforcement: validate_with_capacity rejects GF16 below h=256
(Lucas-closure safe domain from gf16_comparison.md whitepaper).
- L-METRIC enforcement: enforce_l_metric rejects non-BPB primary loss
for JEPA-T and NCA architectures.
- L-R8 stdout discipline: parse_bpb_line accepts only canonical
"BPB=X.XXXX" four-decimal form.
- 16 unit tests, including round-trip on the operator's full 41-name
canonical list, JEPA-T internal-hyphen edge case, GF16-h=256
boundary, L-METRIC scoping to JEPA-T/NCA only, and
architectural_floor_train_v2_below_hybrid sanity lock.
Cargo.toml: +thiserror = workspace.
Tests: 98/98 GREEN across the workspace (was 82; +16 canon tests).
Refs: #43 #58 #61 #62 #64 #65
Anchor: phi^2 + phi^-2 = 3
gHashTag
added a commit
that referenced
this pull request
Apr 28, 2026
Fixes the operator's reported reuse-of-old-service-name footgun. Service slot identifier (EXP_ID) is now decoupled from the RNG seed: two experiments may pin the same rng43 for reproducibility, but the service name carries a fresh monotonically-allocated E<NNNN> token. Canonical form: IGLA-<TYPE>-<FORMAT>-<EXP_ID>[-<TAG>]-rng<SEED> Examples: IGLA-HYBRID-FP32-E0001-rng43 # locked champion (LOCKED) IGLA-HYBRID-FP32-E0042-WSD-rng43 # NEW Phase-3 WSD experiment Champion lock registry (CHAMPION_LOCKS): E0001 — IGLA-HYBRID-FP32 BPB=2.1919 rng43 E0002 — IGLA-HYBRID-FP32 BPB=2.1944 rng45 E0003 — IGLA-HYBRID-FP32 BPB=2.2024 rng44 E0004 — IGLA-TRAIN_V2-FP32 BPB=1.8921 rng42 Four new tripwires (98..101), all GREEN: #98 reject_reused_service_name validate_for_deploy refuses any EXP_ID matching CHAMPION_LOCKS #99 require_monotonic_exp_id validate_for_deploy refuses any EXP_ID <= caller-supplied current_max #100 forbid_naked_seed_in_name validate_for_deploy refuses the legacy IGLA-...-seedN shape; parser still accepts it for read-only history queries #101 kill_before_spin assert_kill_before_spin refuses a deploy when the slot has live occupants and force_replace=false IglaCanon struct extension: + exp_id: Option<u32> // monotonic E<NNNN> + rng: Option<u32> // RNG seed, may repeat + legacy_seed: Option<u32> // pre-INV-12 history-only CanonError variants added: ReusedChampionSlot, NonMonotonicExpId, NakedSeedInDeployName, SlotStillOccupied, MissingExpId, MissingRng Tests: 105/105 GREEN across the workspace (was 97; +8 INV-12 including all 4 tripwires + champion-locks coverage + INV-12-form parser + type-template skip + legacy seed preservation + architectural floor sanity carry-over). Refs: #43 #58 #61 #62 #64 #65 Anchor: phi^2 + phi^-2 = 3
This was referenced Apr 28, 2026
gHashTag
added a commit
that referenced
this pull request
Apr 30, 2026
…101 quick-win) (#106) \ud83e\udeb2 Stateless Scarab Pattern quick-win: claim_next no longer scopes by account. Any free scarab takes any free strategy. Unblocks the starvation symptom observed when acc3 died around 2026-04-30 18:15 UTC and seed=45 waited specifically for acc3 while acc5 idled. ## What changed - bin/seed-agent/src/claim.rs :: CLAIM_SQL drops 'AND account = $2'. The column still appears in RETURNING for observability, but no longer steers claim priority. - bin/seed-agent/src/claim.rs :: claim_next(_, _, _account) keeps the parameter for source-compat; it's underscored to mark dead. A follow-up PR can remove it once all callers are migrated to a 2-argument signature. - test claim_sql_is_account_scoped -> claim_sql_is_fungible_pool. The new test asserts both: (a) 'account = $2' literal is gone, (b) the WHERE clause between 'WHERE' and 'ORDER BY' does not contain 'account' substring. ## Verified locally cargo fmt -p seed-agent : clean cargo test -p seed-agent --bins : 28/28 green (including the new fungible-pool contract test) ## Why a quick-win instead of full crates/trios-scarab migration Full Stateless Scarab migration (new crate, scarab_id uuid, renamed strategy_queue table, LISTEN/NOTIFY wiring) is 3+ hours. Deadline is T-5h. This one-line SQL change captures 80 % of the pool benefit (cross-account failover) without blocking the runway. Full migration tracked in trios-railway#101 (Khepri umbrella) with phase plan. ## Migration safety The UPDATE row-level lock (FOR UPDATE SKIP LOCKED) still guarantees two workers will never claim the same id. Removing the account filter only widens the candidate set each worker sees \u2014 the atomicity guarantee is unchanged. Refs: trios-railway#101 Scarabaeus Engine umbrella trios-trainer-igla#56,#58,#59,#61 (all merged) Anchor: phi^2 + phi^-2 = 3 \u2014 TRINITY \u2014 NEVER STOP. Co-authored-by: Trinity Computer Agent <agent@trinity-s3ai.dev>
This was referenced May 13, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
PR-2 of the gardener rollout. Builds on top of #50 (
feat/tri-gardenerPR-1).Recommended landing order
--reviewmode only--liveonly after a 3-tick review window in Acc1Until #61 lands, set
--account=acc1everywhere and DO NOT register Acc2/Acc3 credentials in the gardener service environment.What this PR does
Replaces the PR-1 stubs with real wiring:
deploy_service,set_vars,redeploy,stop)apply_decision_batchwithKillSwitchprintln!(serde_json)PgLedger(tokio_postgres) /MockLedgertri-gardener serve --interval=N(drift-free tokio interval, SIGTERM-safe)tri-railway service set-vars / logs / stopIssues
tri-gardener serve --interval=3600internal cron (no Railway scheduled-restart dep) #55 (serve --interval)service set-vars/service logs/service stopsubcommands #56 (tri-railway set-vars/logs/stop)Acceptance criteria status
client_ext::tests)live_actuation_writes_to_neoncontract test (passes via MockActuator + MockLedger)kill_switch_aborts_mid_actuationcontract testserve_emits_tick_every_3600scontract test (tokio paused clock)RailwayHash::seal)GARDENER_DISABLED=truehonored on every tick (re-checked, not cached)ARCHITECTURAL_FLOOR_BPB = 2.19constant added inledger.rswith cull-safety comment (refs trios#237 + trios#143)Honest scope
Multi-account safety follow-up (#61 P0):
PR-3 follow-ups:
Decision::CullSeedcurrently recordsOutcome::Skipped).tri-railway service logsis a deterministic stub until the Railway logs subgraph is wired intoClient.Refs
Anchor: phi^2 + phi^-2 = 3 · TRINITY · NEVER STOP