Framing
This audit distinguishes between two different problems:
- truly missing bridge pages — important conceptual connectors that do not yet seem to exist
- already-authored but effectively missing pages — bridge pages that exist but are so underlinked that they do not function as part of the garden
For this garden, the second problem is larger.
Already-authored but effectively missing bridge pages
These should be treated as the first repair targets.
Compression Interfaces for Shared Depth
Bridge role:
- recursive sharing ↔ normalization ↔ phase conditioning ↔ quantization robustness
Why it matters:
- this is one of the strongest architecture/compression synthesis nodes in the garden
- it directly reframes the recurrent-width lane into a more falsifiable thesis
- it should be reachable from both the recursion and quantization entry routes
Tokenizer-Head Co-Design Under a Hard Cap
Bridge role:
- tokenizer efficiency ↔ output-head cost ↔ training economics ↔ artifact budget
Why it matters:
- this is the missing conceptual center for the tokenizer lane
- it explains why tokenizer work is not just preprocessing
- it should be surfaced anywhere the garden discusses vocabulary, logits, or artifact bytes
Refinement Loops as Decompression
Bridge role:
- inference-time compute ↔ recurrent/shared depth ↔ storage substitution
Why it matters:
- it turns the inference-compute lane into a challenge-native thesis instead of a generic search discussion
- it should be discoverable from iterative-refinement and recurrent-architecture routes
Byte Allocation Beats Average Bit-Width
Bridge role:
- selective precision ↔ artifact accounting ↔ structured protection choices
Why it matters:
- it names the quantization problem in challenge-native budget terms
- it should help route readers from low-bit papers to sparse-outlier hypotheses
Challenge History
Bridge role:
- public runs ↔ archetypes ↔ likely future strategy families
Why it matters:
- this is the clearest public-record bridge layer
- it currently behaves like a private annex instead of a challenge-facing route
Truly missing bridge pages
These are the clearest genuinely absent bridge nodes.
1. ideas/index or equivalent idea-hub page
Why it is genuinely missing:
- every
ideas/*page is orphaned - the garden has no visible landing page that says “here are the next falsifiable proposals”
- without a hub, ideas are hard to route from frontiers and hypotheses
Research value:
- highest-value because it turns synthesis into next-step agenda
2. experiments/index or equivalent evidence hub
Why it is genuinely missing:
- experiment pages do not currently form a visible evidence layer
- hypotheses have no obvious route to concrete test records
- the graph risks becoming theory-heavy and evidence-light
Research value:
- high, because it would connect hypotheses to observed outcomes rather than leaving experiment notes isolated
3. A canonical bridge between challenge/* and challenge-history/*
Why it is genuinely missing:
- the garden currently splits public-record material across two parallel structures
- readers can arrive at History and Public Runs without ever discovering Challenge History
- the duplication creates ambiguity about where challenge history actually lives
Research value:
- high, because challenge-facing interpretation is part of research prioritization, not just editorial neatness
Bridge gaps that are more routing than authorship
These concepts probably do not need new pages first. They need stronger routing into the pages that already exist.
Shared depth × low-bit stability
Existing candidates:
- Compression Interfaces for Shared Depth
- Shared Depth Needs Cheap Specialization
- Phase-Conditioned Sharing
Tokenizer × head × byte budget
Existing candidates:
- Tokenizer-Head Co-Design Under a Hard Cap
- The LM Head Is Part of the Compression Problem
- Output-Head Compression
- Head-to-Depth Budget Swap
Compute-for-bytes refinement
Existing candidates:
- Refinement Loops as Decompression
- Iterative Refinement over Stored Depth
- Token-Adaptive Recurrent Refinement
Selective precision as byte ROI
Existing candidates:
Highest-value bridge conclusion
If only one bridge repair happens, it should be this:
promote
frontiers/*from a hidden side layer into the main research spine.
That single change would connect:
- papers to hypotheses
- lanes to next-step ideas
- architecture questions to compression questions
- tokenizer questions to byte-budget questions
- inference-time compute to recurrent model design
Only after that should the garden invest heavily in authoring net-new bridge pages.