Dated observation note. Recorded 2026-08-09 at census tip 97,530, corpus complete (/v1/status → certParams.expected = parsed = 97,530, missing = 0, checked the same day). Sources: registry/PCCR.md (consensus events, per-row code evidence), OBS-003 (the in-window count), and the live census API (metrics.md#cert-census) queried on the recording date. No new computation was required; §2 states how each number was re-verified on the day of writing.

Terminology, per the label discipline in metrics.md: a "MoE certificate" is a canonical block whose certificate public data declares moe_e > 0 (the V2 grouped-GEMM tail). That is a declared and committed MoE-shaped job. Nothing in this note claims expert inference ran, was fresh, or served anyone.

1. The finding

A capability introduced by the chain's only PIP-documented fork, genuinely used in roughly one block in ten while legal, was withdrawn 41 days later by a softfork no document proposes, explains, or announces.

(Precision note: PCCR-0003 and roadmap §9.5 previously carried "≈44 days", a blocks × target-time estimate. The measured duration between the two activation block timestamps is 41.0 days; a dated precision line was added to PCCR-0003 today.)

2. How each number was verified on the recording date

• Consensus events and heights: PCCR-0002/0003 rows, code evidence node/chaincfg/params.go (MoEForkHeight, DenseOnlyForkHeight) and node/wire/certificate_v2.go @ v1.2.1 (SRC). Calendar dates from /v1/blocks/{71935,91630} block timestamps, fetched today.
• 1,929: OBS-003 measured it on 2026-08-06 (tip 96,397) from cert_public_params (moe_e > 0) over the window height range; the census window parameter cannot express height bounds, so the in-window figure is a research query, not a census call. Today's whole-corpus census at tip 97,530 returns moeBlocks = 1,929 (identical), so zero MoE certificates entered the corpus between the two measurements, and OBS-003's placement of all 1,929 inside the window carries over unchanged.
• 0 after 91,630: two independent lines. Rule: MoE certificates are consensus-invalid outside the window under PCCR-0003, with PCCR-0004 (v1.2.1) fixing the IsMoE predicate that feeds that validity check. Measurement: the whole-corpus count is unchanged at 1,929 (previous bullet), and the 24h / 7d windows measured today contain 0 of 462 and 0 of 2,857 blocks respectively.
• 9.8%: 1,929 / 19,695; denominator = canonical blocks at heights 71,935–91,629 inclusive. Census figures count canonical blocks only (orphaned certificates are retained in the corpus but excluded).

3. Why this matters

Pearl distributes a checkpoint its own consensus cannot validate work for. Qwen3-30B-A3B-Instruct-2507-pearl, one of the four checkpoints Pearl publishes, and the reference stack's end-to-end-tested MoE model, is mixture-of-experts (128 experts, top_k 8; config verified live 2026-08-05, pearl-notes.md F4). Since height 91,630 the network rejects any proof for that architecture.

The unmineable class is the class the ecosystem is moving toward. Much of the current open-weight frontier is mixture-of-experts; this note asserts the architecture only for Pearl's own published checkpoint, whose config we verified. For any named third-party model the claim is conditional: if it is MoE, then it is unmineable under PCCR-0003.

Dense models remain mineable, and that constraint is economic, not architectural. Roughly two-thirds of a dense model's linear arithmetic is INT7-mineable (PREREG-002 §3: down_proj is FP8 and excluded; only half the attention layers qualify).

The retraction also removed the chain's richest attestation surface. The V2 MoE tail is the only certificate region that ever committed to routing (hash_routing, expert index and routing offsets, pearl-notes.md §Certificate internals, SRC). Since 91,630 no Pearl certificate commits to any routing decision. Per the terminology above: committed routing data, not evidence that routing served anyone.

4. Governance reading (facts; no motive)

• Of the mainnet consensus changes in the register (PCCR-0002…0005 plus the release-gated PCCR-0004), exactly one has a PIP, and it is the one that was reversed.
• No public document explains the withdrawal. That absence is recorded as an absence. We do not assert a motive; if the retraction was a response to a defect, mechanism detail would fall under the roadmap §9.4 disclosure split, and nothing constructive is described here.
• The code, not the PIP repository, is Pearl's de-facto governance record; the independent ledger is registry/PCCR.md.

5. Boundaries

1. We measure that MoE proofs are rejected and that MoE blocks are zero. We do not know why the capability was withdrawn, and this note asserts no reason.
2. The 1,929 count is declared-and-committed MoE-shaped jobs (accepted by consensus under the window's rules). It is not a claim that expert inference occurred or served anyone.
3. Dense models remain mineable (§3); nothing here says Pearl "cannot mine models", only that it rejects proofs for this architecture class.
4. Named third-party models must not be asserted MoE or unmineable without checking their published configs; the defensible claim is conditional (§3).
5. PCCR-0004 (v1.2.1's IsMoE predicate change, PublicDataLen != 164 → > 164) alters which serialized certificates count as MoE and therefore interacts with PCCR-0003 validity; it is part of this note's evidence chain.

6. Reproduction

curl -s localhost:8081/v1/status | jq .certParams        # corpus completeness
curl -s  localhost:8081/v1/certs/census                | jq '{blocks, moeBlocks}'  # 97530 / 1929 (2026-08-09)
curl -s 'localhost:8081/v1/certs/census?window=7d'     | jq '{blocks, moeBlocks}'  # 2857 / 0
curl -s 'localhost:8081/v1/certs/census?window=24h'    | jq '{blocks, moeBlocks}'  # 462 / 0

Consensus events: registry/PCCR.md entries PCCR-0002, PCCR-0003, PCCR-0004 (each row carries file:line @ tag code evidence). In-window count: OBS-003 §2, from cert_public_params (moe_e > 0) over heights 71,935–91,629, a DB query, because the census window parameter cannot express height bounds. Every census figure in this note is gated by metrics.md#cert-census; the note introduces no new metric (confirmed against metrics.md on the recording date, per PLAN-001 Step 1A's gate).

7. Data

No new dataset. Census responses as quoted (computedAt 2026-08-09T12:59Z / 13:22Z); consensus evidence in PCCR; the in-window measurement is OBS-003's. Web presentation: the annotated timeline strip is specified as PLAN-002 §9.2 figure N2 and needs no chart library.