Parameter Golf Research Garden

Executive Link-Densification Audit

4 min read

Status note

This audit captured a pre-integration snapshot of the graph. Some top-level routing fixes have since landed, especially around frontiers, challenge history, and local experiment history. Read the findings as an audit baseline rather than a claim that every issue below is still unresolved.

Scope

Audited the current knowledge garden structure under quartz-kb/content with emphasis on:

  • challenge framing
  • lanes
  • hypotheses
  • notes
  • papers
  • frontiers
  • ideas
  • experiments
  • challenge-history

Snapshot

  • 65 markdown nodes audited
  • 10 orphan nodes by inbound link count
  • 9 actionable orphans if the root home page is excluded
  • 3 hidden research clusters that are valuable but weakly surfaced:

Main finding

The garden’s highest-value link problem is buried synthesis, not missing trivia.

Several of the best cross-lane pages already exist, but the main graph does not route readers into them:

  1. Research Frontiers has no inbound links from the home page, graph page, lane pages, or hypothesis ledger.
  2. Challenge History is disconnected from the main challenge/* section even though it contains the stronger public-record structure.
  3. All five ideas/* pages are fully orphaned even though they are concrete, falsifiable research proposals.
  4. RWA Breadth Experiment is orphaned, so the architecture lane has almost no visible evidence trail.
  5. Unified Compression-Aware Architecture is orphaned despite being a potentially important synthesis node.

Priority diagnosis

Priority 1 — surface the existing bridge layer

The most valuable existing bridge pages are:

These are exactly the kinds of pages that should connect papers to hypotheses and future experiments, but they are mostly only linked from inside frontiers/ itself.

Priority 2 — connect challenge framing to public-record context

The garden currently splits public-history material across:

The challenge-history/* subtree points back into the main challenge pages, but the reverse routing is missing. The result is a one-way bridge: detailed public-record context exists, but readers entering from the main garden do not naturally discover it.

Priority 3 — connect hypotheses to evidence and next-step ideas

The garden has strong conceptual material but weak hypothesis-to-evidence routing:

This makes the graph intellectually rich but operationally thinner than it could be.

Weakly connected concepts that matter most

1. Shared depth needs a compression interface

The most important architecture/compression bridge is already written in Compression Interfaces for Shared Depth, but it is under-routed.

Why it matters:

  • it joins recurrence, normalization, phase specialization, and quantization robustness
  • it sharpens the best current research seam around shared-depth models
  • it has direct experiment implications

2. Tokenizer work is really head-and-budget work

The garden has the right ingredients:

But the frontier page and related idea pages are not surfaced from the main routes, so this seam reads as less central than it really is.

3. Evaluation-time refinement as a storage substitute

Refinement Loops as Decompression and Token-Adaptive Recurrent Refinement define a serious compute-for-bytes seam, but the graph currently understates it.

4. Byte allocation as a budget problem, not a bit-width problem

The quantization cluster is reasonably strong, but Byte Allocation Beats Average Bit-Width should be treated as a first-class bridge page because it reframes selective precision in challenge-native terms.

Dead-end or low-yield routing patterns

Research Atlas is a dead-end overview

The atlas is cited from challenge-facing pages, but it does not currently act like a strong router into the rest of the garden. It should be a high-value waypoint, not a cul-de-sac.

ideas/* behaves like a detached notebook

The idea pages are good research seeds, but without inbound links they are effectively private notes rather than garden nodes.

experiments/* lacks an evidence layer

The one current experiment page is disconnected enough that hypotheses do not visibly accumulate evidence.

  1. Surface Research Frontiers from major hubs
  2. Surface Challenge History from the main challenge entry points
  3. Backlink orphan experiment and orphan synthesis nodes into their parent hypotheses
  4. Route ideas/* from the frontier and hypothesis pages they instantiate
  5. Only after that, consider authoring truly new bridge pages

Bottom line

The garden already contains much of the missing intellectual structure. The next gain comes from routing the existing best synthesis pages into the main graph, especially around:

  • shared depth × compression interfaces
  • tokenizer × LM-head × byte budget co-design
  • refinement-as-decompression
  • public-record history and strategy families

Graph View

Referenced by

Recent notes