Record: #1787 + Sparse Gate + Updated Frozen Carry — val_bpb 1.06287

.claude/scheduled_tasks.lock

@@ -0,0 +1 @@
+{"sessionId":"845a7b07-cc9e-4b32-9356-527eaa592da2","pid":3874006,"acquiredAt":1776804744293}

.claude/scripts/discord_post.sh

@@ -0,0 +1,26 @@
+#!/usr/bin/env bash
+# Post a message to Discord. Text from $1 or stdin.
+# Usage:
+#   discord_post.sh "one-line message"
+#   echo -e "multi\nline" | discord_post.sh
+set -euo pipefail
+
+SECRETS="${DISCORD_SECRETS:-/home/claude-user/.claude/skills/secrets.env}"
+set -a; source "$SECRETS"; set +a
+
+if [ $# -ge 1 ]; then
+  MSG="$1"
+else
+  MSG="$(cat)"
+fi
+
+# Discord hard-caps message content at 2000 chars.
+if [ "${#MSG}" -gt 1990 ]; then
+  MSG="${MSG:0:1980}…(truncated)"
+fi
+
+curl -sS -X POST \
+  -H "Authorization: Bot $DISCORD_BOT_TOKEN" \
+  -H "Content-Type: application/json" \
+  --data-raw "$(jq -n --arg c "$MSG" '{content:$c}')" \
+  "https://discord.com/api/v10/channels/$DISCORD_CHANNEL/messages" > /dev/null

.claude/scripts/discord_post_table.sh

@@ -0,0 +1,20 @@
+#!/usr/bin/env bash
+# Post a monospace code-block (table / multi-line) to Discord.
+# Usage:
+#   discord_post_table.sh < path/to/file
+#   cat table.txt | discord_post_table.sh
+#   discord_post_table.sh "line1\nline2\nline3"
+set -euo pipefail
+
+HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+
+if [ $# -ge 1 ]; then
+  BODY="$(printf '%b' "$1")"
+else
+  BODY="$(cat)"
+fi
+
+# Wrap in triple-backtick code block for fixed-width rendering.
+MSG="$(printf '```\n%s\n```' "$BODY")"
+
+printf '%s' "$MSG" | "$HERE/discord_post.sh"

.claude/skills/execution/SKILL.md

@@ -0,0 +1,37 @@
+---
+name: execution
+description: Activate execution role for the Parameter Golf repo. Invoke at the start of an execution session (pod is live or about to be). Loads execution protocol and reminds what execution does vs what research does.
+---
+
+# Execution role
+
+You are in **execution mode** for the Parameter Golf record-track push.
+
+## What execution does
+- Reads `CLAUDE.md` and `EXECUTION.md` at the top of this repo (authoritative).
+- Is handed one spec number at a time. Reads **only** `research/specs/NNN-slug.md`, plus the two top-level docs. Does not browse other specs.
+- Interviews the spec with the user before launching (see `EXECUTION.md` §"Spec interview protocol"). Surfaces ambiguities, resolves open questions.
+- Runs preflight checklist before every launch.
+- Follows the hardware ladder: 2×H100 mini → 8×H100 official, per the spec.
+- Writes artifacts to `runs/NNN-slug/` (or `runs/NNN-slug/seed_XX/` for multi-seed) with the exact shape in `EXECUTION.md`.
+- Stops the pod immediately after eval.
+- Hands back a one-paragraph summary to the user.
+
+## What execution does NOT do
+- Never modifies training logic code. Only environmental fixes (missing deps, path typos, `CUDA_VISIBLE_DEVICES`).
+- Never interprets results or decides promote/iterate/kill — that's research.
+- Never writes rows to `experiments.md`.
+- Never writes `research/evaluations/`.
+- Never launches without a completed spec interview + passed preflight.
+
+## First actions on session start
+1. Read `CLAUDE.md` and `EXECUTION.md` (especially the "Pod operations playbook" section — don't re-learn its lessons at $0.40/min).
+2. Ask the user which spec number to run.
+3. Open that spec and begin the interview.
+
+## Reminders
+- If a logic bug surfaces mid-run: stop pod, hand back to research, do not patch on the fly.
+- Stop pods immediately after every run (`runpodctl pod stop <id>`). Same-day → stop. End-of-day → `runpodctl pod delete <id>` to fully terminate.
+- **Rsync artifacts BEFORE stopping the pod**, not after. Starting a pod back up to rsync costs ~1 min of pod time and another SSH re-handshake.
+- `final.json` is the deliverable — if it's not written, the run is lost.
+- Checkpoints live on the NA-1 volume, not in git. Git gets `checkpoints.md` pointer files only.

.claude/skills/frontier-scan/HEURISTICS.md

@@ -0,0 +1,85 @@
+# Frontier-scan heuristics
+
+Used by the frontier-scan skill to classify each PR. Update this file when new
+rulings land or new patterns emerge.
+
+## Our baseline
+
+**PR #1736** (dexhunter) — val_bpb **1.06549** (3-seed mean).
+Stack: SP8192 + CaseOps tokenizer + SmearGate + AttnOutGate + QuantGate +
+Loop45 depth-recurrence + phased TTT (multi-phase global SGD + doc-indep LoRA).
+
+## Legitimacy categories
+
+### `clean`
+No open disputes. Training and eval procedures follow competition rules.
+
+### `tokenizer-disputed`
+Uses a non-standard tokenizer. Apply Issue #1604 lossless-roundtrip test:
+- **Likely legal:** lossless CaseOps — `decode(encode(s)) == s` holds; byte
+  sidecar recovers original bytes; exact byte count recoverable from token IDs.
+- **Likely illegal:** lossy casefold (`.lower()`) — destroys case information;
+  byte count cannot be reproduced from token IDs alone.
+
+### `prequant-ttt-disputed`
+Issue #1017 Condition 3: "you may not test-time train on validation data before
+the artifact is frozen." A pre-quant TTT pass runs AdamW/SGD updates on val
+tokens, then freezes weights — this is the disallowed pattern.
+- Physics ceiling check: bigbag's empirical corpus-level TTT ceiling is
+  **~0.0003 bpb**. Any submission claiming >0.005 bpb from TTT-related levers
+  deserves scrutiny. A claim of ~0.038 bpb (as in #1735) is ~100× over the
+  ceiling — assign 85-90% probability of illegal ruling.
+
+### `byte-bug-suspect`
+GatedDeltaNet / FLA cluster (Issue #1719): `build_sentencepiece_luts` double-
+counts the byte denominator by ~17.46%. Corrected bpb is ~18% higher than
+claimed. PRs: #1698, #1711, #1712, #1734 (three self-closed). Flag any PR
+using flash-linear-attention (`from fla.` import) for this check.
+
+### `broken`
+Artifact oversized, training crash, or claim definitively refuted.
+
+### `other`
+Non-record track, WIP, negative-result, or reproduction only. Skip unless it
+contains an extractable novel lever.
+
+## Absorption rule
+
+Training-time levers (architecture changes, optimizer changes) are at risk of
+being absorbed by #1736's phased TTT — TTT adapts the model at eval time and
+can compensate for a missing training-time improvement. Quant-side levers have
+been empirically shown absorbed (specs 009/010/010b). Eval-time levers
+(Tap-In, score-first TTT, hash embeddings trained only via TTT) are downstream
+of TTT and **cannot be absorbed**.
+
+## Already-in-#1736 (exclude from "novel lever" classification)
+
+SP8192 tokenizer, CaseOps, SmearGate, AttnOutGate, QuantGate, Loop45, phased
+TTT (multi-phase SGD + LoRA), VarLen attention + fused MLP (TMA), Muon
+matrix_lr=0.026, per-layer adaptive GPTQ clip (MLP/ATTN/MATRIX sigmas), int8
+embeddings, logit_softcap, XSA (`F.normalize(v, dim=-1)`), RMSNorm Q/K,
+`leaky_relu(x,0.5).square()` MLP activation, GPTQ mixed int6/int8.
+
+## Already-specced (don't re-flag as new)
+
+| Spec | Lever |
+|------|-------|
+| 011  | Tapered weight decay |
+| 012  | GradPower Muon p=0.9 |
+| 013  | xIELU + per-layer QK gain |
+| 014  | Tap-In min_match=1 |
+
+## Banned mechanisms (always illegal)
+
+- Lossy casefold (Issue #1604)
+- Pre-quant TTT (Issue #1017 Condition 3)
+- N-gram eval cache
+- Trinity/SLOT
+- GatedDeltaNet / FLA byte-bug cluster — treat any `from fla.` import as a
+  red flag requiring byte-accounting audit before accepting claimed bpb.
+
+## Actionable-delta threshold
+
+Flag a PR as **actionable** if its clean claimed bpb is below our baseline
+**1.06549**. Flag as **watch** if between 1.065 and 1.075. Above 1.075 is
+background noise unless it contains an isolated novel lever.