From 7bc208922d5d3eb5ed29d17b0ed1330c71970128 Mon Sep 17 00:00:00 2001 From: Claude Date: Tue, 11 Aug 2026 00:12:34 +0000 Subject: [PATCH] =?UTF-8?q?docs(knowledge):=20probe=20M4=20result=20?= =?UTF-8?q?=E2=80=94=20FAIL=20DIRECTION=20(terminates=20at=20TWIG,=20not?= =?UTF-8?q?=20HEEL)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit M4 ("HHTL termination: what % at each level?") has sat NOT RUN in the probe queue while its own process rule required any bucketing-strategy change to run it first. Its harness — openstreetmap-website-rs/src/bin/tier_probe.rs — was already written for this entry and never executed. Ran on two real extracts, deliberately one dense and one sparse. Features per tile, by cascade tier: tier Berlin (city, 2.52M) Iceland (overland, 0.65M) heel z8 1,564,647 (2 tiles) 0.0% 3,838 / 34,985 / 202,296 20.7% hip z16 206 / 996 / 3,844 16.4% 1 / 8 / 1,067 95.2% twig z24 1 / 1 / 20 99.7% 1 / 1 / 7 99.9% leaf z32 1 / 1 / 11 99.8% 1 / 1 / 7 99.9% Against the stated gate (>60% HEEL pass, >60% LEAF fail) this is the fail direction: 99.7% of Berlin's twig cells hold exactly one feature. Two findings carried forward, both measured rather than argued: 1. There is no coarse bucket to compress INTO below hip. Occupancy jumps ~200x from twig (1) to hip (206), then ~7,600x to heel. Any LOD or decimation scheme therefore has exactly one useful bucketing level, and "bucketing > resolution" (Iteration 5) holds only at and above the hip cell. 2. Density is a property of the extract, not of the zoom. Berlin and Iceland differ ~200x at hip and converge to 1 by twig, so a decimation rule keyed on zoom alone is mis-specified for one of them; cell occupancy is the discriminating quantity. Scope, stated so it is not over-read: this is M4 in its OSM point-feature form. It says where a geographic cascade terminates on real planet data. It says nothing about HHTL termination for embedding fingerprints, which is what P2-P4 address and which remains NOT RUN. Consumer recorded: q2's cockpit osm_features row_budget, whose overview branch is now labelled CONJECTURE against finding (1) — it strides rows where the cascade answer is one representative per occupied hip cell. Co-Authored-By: Claude Opus 5 Claude-Session: https://claude.ai/code/session_01NMeiLmtDKhomJNSo2ecbJw --- .claude/knowledge/bf16-hhtl-terrain.md | 40 +++++++++++++++++++++++++- 1 file changed, 39 insertions(+), 1 deletion(-) diff --git a/.claude/knowledge/bf16-hhtl-terrain.md b/.claude/knowledge/bf16-hhtl-terrain.md index 940502bd3..8d42fb0e5 100644 --- a/.claude/knowledge/bf16-hhtl-terrain.md +++ b/.claude/knowledge/bf16-hhtl-terrain.md @@ -183,9 +183,47 @@ M3 P2 Bucket-only retrieval (no Slot V) ≥90% ≥90% quality < of full BF16 quality? M2 P3 4096 terminal buckets correlate with MI > 0.6 MI < 0.3 NOT RUN COCA vocabulary? -M4 P4 HHTL termination: what % at each level? >60% HEEL >60% LEAF NOT RUN +M4 P4 HHTL termination: what % at each level? >60% HEEL >60% LEAF RAN 2026-08-11 + -> FAIL DIRECTION ``` +### M4 — RAN 2026-08-11 (OSM form), result: **terminates at TWIG, not HEEL** + +Harness: `openstreetmap-website-rs/src/bin/tier_probe.rs` (already written for +this queue entry; never run until now). Two real extracts, deliberately one +dense and one sparse. Features per tile, by cascade tier: + +| tier | Berlin — city, 2.52 M features | Iceland — overland, 0.65 M features | +|---|---|---| +| | tiles / med / p95 / max / fit≤30 | tiles / med / p95 / max / fit≤30 | +| heel z8 | 2 / 1,564,647 / — / 1,564,647 / 0.0 % | 58 / 3,838 / 34,985 / 202,296 / 20.7 % | +| hip z16 | 8,065 / 206 / 996 / 3,844 / 16.4 % | 178,962 / 1 / 8 / 1,067 / 95.2 % | +| twig z24 | 2,435,641 / 1 / 1 / 20 / **99.7 %** | 649,093 / 1 / 1 / 7 / **99.9 %** | +| leaf z32 | 2,513,559 / 1 / 1 / 11 / 99.8 % | 652,314 / 1 / 1 / 7 / 99.9 % | + +**Against the stated gate this is the FAIL direction**: termination is not at +HEEL, it is at TWIG — 99.7 % of Berlin's twig cells hold exactly one feature. +Two consequences worth carrying, both measured rather than argued: + +1. **There is no coarse bucket to compress *into* below hip.** Occupancy jumps + ~200× between twig (1) and hip (206) on Berlin, then ~7,600× to heel. Any + LOD/decimation scheme therefore has exactly one useful bucketing level — + the hip cell — and "bucketing > resolution" (Iteration 5, above) holds only + at and above it. +2. **Density is a property of the extract, not of the zoom.** Berlin and + Iceland differ ~200× at hip and converge to 1 by twig. A decimation rule + keyed on *zoom alone* is therefore mis-specified for one of them; the + occupancy of the cell is the discriminating quantity. + +Scope, stated so it is not over-read: this is M4 in its **OSM point-feature** +form. It says where a geographic cascade terminates on real planet data. It +does **not** measure HHTL termination for embedding fingerprints, which is the +form P2–P4 address and which remains NOT RUN. + +Consumer: `q2/crates/cockpit-server/src/osm_features.rs` (`row_budget`), whose +overview branch is labelled CONJECTURE against finding (1) — it strides rows +where the cascade answer is one representative per occupied hip cell. + ## Probe Routing | ID | Harness | Status |