| name | leaderboard-snapshot |
| description | Build a manual IGLA-race leaderboard from all available sources (trainer-igla seed_results.jsonl, ALPHA comments in trios#143, Audit Watchdog history, Railway service list, optional Neon bpb_samples). Use when the operator says "/leaderboard", asks for "live leaderboard", "таблицу фаворитов", "snapshot гонки", "champion table", "leaderboard сейчас", or any variant requesting a current race standings table — especially when telemetry path is partially or fully broken (no Neon BPB rows) and the table must be reconstructed manually from primary artefacts. |
/leaderboard — IGLA race manual snapshot
Anchor: phi^2 + phi^-2 = 3 · TRINITY · NEVER STOP
When to load this skill
The operator wants the current race standings table right now, even when:
bpb_samples / gardener_runs Neon tables are missing (#62) — telemetry path broken- Some Railway accounts probe-failed (Not Authorized — #61 P0 not yet merged)
- Trainer writer patch is sitting in a local-only branch (CODEX heartbeat path)
- Audit Watchdog reports
Ledger rows = 0 for many ticks in a row
The skill assembles the table from primary artefacts that are observable, and is honest about the unobserved cells.
One-shot procedure
Run these 6 steps in order. Step 6 (chat output) MUST be first user-visible action.
Step 1 — Pull latest seed_results.jsonl
cd /home/user/workspace/repos/trios-trainer-igla && git pull origin main
tail -50 assertions/seed_results.jsonl
cat assertions/champion_lock.txt
ls assertions/.gate2_done 2>/dev/null
Each row carries: seed, steps, bpb, optional ema_bpb, optimizer, hidden, lr, attn_layers, sha, gate_status, optional agent. Rows with gate_status=new_champion are the architectural high-water marks.
Step 2 — Pull last ~24h of comments from trios#143
gh issue view 143 --repo gHashTag/trios --json comments > /tmp/trios143.json
python3 -c "
import json
with open('/tmp/trios143.json') as f: d=json.load(f)
for c in d['comments'][-25:]:
print('---', c['createdAt'])
print(c['body'][:2000]); print()
"
Look for:
- AGENT ALPHA / 3-SEED RESULTS comments — primary BPB signal (e.g. champion
2.1919 @ 81000 seed=43 sha=cd91c45)
- AGENT CODEX HEARTBEAT — local-only patches blocked on push (writer often blocked here)
- 🛰️ Audit Watchdog comments — periodic snapshot of ledger row count and drift events per cluster (Acc1 / Acc2)
- RISK-ENG notes — falsifications (e.g. L7 Muon NS-1: +0.11 BPB worse; L8 ByteFlow rejected)
Step 3 — Probe Railway via the trios-perplexity MCP
railway_service_list(project="e4fe33bb-3b09-4842-9782-7d2dea1abc9b")
railway_service_list(project="39d833c1-4cb6-4af9-b61b-c204b6733a98")
This lists services (id + name + created_at). It does not stream logs — there is no railway_service_logs tool in this MCP today. So service presence answers "is the seed deployed?", but not "what step / BPB is it at?". Honest gap: until bin/tri-railway service logs ships (#56) and writer pushes BPB rows to bpb_samples (#62), L1/L2/L4 cells stay tracking / warmup.
Step 4 — Optional Neon BPB cross-check (currently 42P01)
SELECT seed, lane, step, bpb, ts
FROM bpb_samples
WHERE ts > now() - interval '12 hours'
ORDER BY seed, ts DESC;
SELECT lane, seed, before_bpb, after_bpb, ts
FROM gardener_runs
WHERE ts > now() - interval '12 hours'
ORDER BY ts DESC;
If both are 42P01, drop the column quietly and rely on Step 1+2 only. Do NOT synthesise BPB numbers; mark unobserved cells as tracking and the trend column as ⏳.
Step 5 — Assemble the markdown table
Columns (order is fixed):
| Rank | Seed | Lane | Steps | BPB | Δ→Gate-2 | Trend | Status |
Rules:
- Rank by BPB ASC. Locked / champion seeds at the top.
tracking rows after the locked block, no rank.
Δ→Gate-2 = BPB - 1.85, sign always + (we are above target).Trend = ↓ improving, → plateau, ↑ worse, ⏳ no observation.Status cites the source comment with its timestamp, e.g. locked (ALPHA 16:38Z). Use Markdown links to the issue comment when the URL is known.- For seeds that are deployed but un-observed:
Lane = lane code + service id from Step 3, BPB = tracking, Steps = —, Status = warmup + linked CODEX/ALPHA comment that explains the block.
Step 6 — Output to chat first, then meta
Operator wants the table first, no preamble. After the table:
- Champion BPB sparkline — last 3 architecture points (avg per cohort), each rendered as a tiny block-row (
▁..▆) plus the BPB average and total cumulative delta from the worst cohort.
- Honest gaps section — bulleted list of why the unobserved rows are unobserved, each gap citing the gating issue (#56 logs subcommand, #61 multi-account, #62 bpb_samples DDL).
- ETA table — Rung-1 / Rung-2 / now / final / Gate-2 deadline, both wall-clock UTC and T+.
- Arithmetic check —
gap_to_gate2 = best_bpb - 1.85 vs lane_delta_sum; if the lane-delta sum ≥ gap, race is arithmetically alive; if not, raise the flag.
Constants (refresh per race)
RACE_START_UTC = 2026-04-27T18:00:00Z
GATE2_DEADLINE_UTC = 2026-04-30T23:59:00Z
GATE2_TARGET_BPB = 1.85
ARCH_FLOOR_BPB = 2.19 # ARCHITECTURAL_FLOOR_BPB in tri-gardener::ledger
RUNG1_OFFSET_HOURS = 12
RUNG2_OFFSET_HOURS = 18
RUNG3_OFFSET_HOURS = 28
FINAL_OFFSET_HOURS = 50
DEADLINE_OFFSET_H = 54
Hard rules
- R5 honesty: never invent BPB numbers. Unobserved is
tracking, not "estimated". Plateau is only declared with ≥5 ticks AND step ≥ 50K (matches tri-gardener::ledger::ARCHITECTURAL_FLOOR_BPB cull-safety rule).
- R1 Rust-only does not apply here — this skill is read-only on JSONL + GitHub + Railway MCP; nothing is compiled or shipped. Use
python3 for parsing freely.
- No emoji in headers; medal emojis (🥇🥈🥉) are allowed in the rank cells only.
- Cite each fact with the comment timestamp — every row that depends on a
trios#143 ALPHA comment must link to that comment with its timestamp visible in the anchor text.
- Never use the word scrape/crawl when describing the Railway probe — say "list" or "probe".
Refs
phi^2 + phi^-2 = 3 · TRINITY · NEVER STOP
