The mechanism board

Most chat AIs are built from a few standard parts: a way to look back at earlier words (attention), a block that processes each word (a feed-forward layer), big grids of multiplications, and a function that scores every possible next word. UOR-R4 is trying to replace each of those parts with something built from geometry, such as rotations, exact addresses and lookup tables. Each replacement is called a mechanism. This page lists all 27 of them and says where each one stands today. Status here means how far a mechanism has been built and tested. No mechanism has produced a useful conversation yet: the project is pre-alpha, and 0 of its 8 milestones are complete (STATUS.md).

The 22 mechanisms with a recorded parent or child, drawn as a family tree. Most of the live ones descend from earlier, parked attempts.

What the three statuses mean

There is no "retired" status. Owner decision D22 says a negative result closes one tested configuration, never the mechanism, and stopping a mechanism needs the owner (D22). This board replaces the 18 October labels (active, dormant, archived-negative), which did not map one to one onto these three and had no rows for ten mechanisms.

Every number in the tables is a measurement from a recorded run or test, linked in its row, unless the cell says otherwise. The "Next" column is a hypothesis about what would help, not a result. A passing test is not a trained capability, and NOT_RUN is not a pass. BPB means bits per byte of text (lower is better). "FIT" means a pre-registered test was reviewed and judged able to answer its question.

The family tree

An arrow means "came from" or "was replaced by". A solid arrow is a successor, a dotted arrow means the target is the integer form of the source, and a double arrow means the target is a specialisation of the source.

flowchart TD
  r4g1["R4G1 graph<br/>PARKED"]
  nat["Native geometric engine<br/>PARKED"]
  lut["D11 integer serving<br/>ACTIVE"]
  stack["Geometric stack<br/>PARKED"]
  disc["Discrete learning<br/>ACTIVE"]
  pz["Prime-zeta codebook<br/>ACTIVE"]
  rel["Relation read<br/>ACTIVE"]
  obj["Object store<br/>READY"]
  rank["Rank-margin hinge<br/>ACTIVE"]
  cc["Carrier credit<br/>PARKED"]
  lc["Legal construction<br/>PARKED"]
  sfbs["SpiralCore/FBS<br/>READY"]
  tx["Ternary/XOR<br/>PARKED"]
  spin["Spin-native weights<br/>PARKED"]
  ham["Hamilton gate<br/>READY"]
  rec["Recurrence gates<br/>READY"]
  flock["Flock read<br/>ACTIVE"]
  route["Routed read<br/>ACTIVE"]
  ptr["Pointer key<br/>PARKED"]
  cells["Address cells (A8)<br/>ACTIVE"]
  mem["Exact memory<br/>PARKED"]
  e8["E8 orbit memory<br/>PARKED"]

  r4g1 --> nat
  r4g1 --> lut
  nat --> stack
  nat --> disc
  nat --> pz
  nat --> rel
  nat --> obj
  nat --> rank
  cc -.-> disc
  lc --> disc
  sfbs -.-> disc
  tx --> disc
  spin --> sfbs
  stack --> ham
  stack --> rec
  stack -.-> flock
  route --> flock
  flock ==> rel
  ptr --> rel
  pz --> cells
  mem ==> e8

  classDef ACTIVE fill:#d7f0e0,stroke:#1a7f4b,color:#0b2e1a,stroke-width:2px
  classDef READY fill:#dbe8fb,stroke:#1d4ed8,color:#0b1f4d,stroke-width:2px,stroke-dasharray:6 3
  classDef PARKED fill:#e5e7eb,stroke:#4b5563,color:#111827,stroke-width:2px,stroke-dasharray:2 3
  class lut,disc,pz,rel,rank,flock,route,cells ACTIVE
  class obj,sfbs,ham,rec READY
  class r4g1,nat,stack,cc,lc,tx,spin,ptr,mem,e8 PARKED

Five mechanisms have no recorded parent or child, so they are not drawn: admission, the tokenizer, Hopf observation, read binding and VSA. Source for all edges: the family_tree field of the mechanism board (19 October 2026).

ACTIVE: on the current model path (10)

READY: built and tested, not yet trained in (4)

PARKED: tested forms failed or broke a rule, still open (13)

Which mechanism covers each part of a transformer

Each row names a part of a standard transformer, says what it does, and lists the mechanisms aimed at it (A = ACTIVE, R = READY, P = PARKED). Source: the slot_map of the mechanism board (19 October 2026).

The two gaps matter. The only position-related result is about recurrent state transport, not word order, and it was measured in a model that used banned parts. The live discrete model has no position mechanism, no normalization and no feed-forward block, and it scores below a unigram baseline (audit synthesis, section 1). Having a mechanism for a slot does not mean it works: a row can list four mechanisms and no trained result.

Where the audit questions a status

Statuses are left as the audit dossiers set them (docs/labs/repo-renormalization-2026-10-19/audit/mechanisms/). Two are open to challenge:

• D11 integer serving is ACTIVE as the end-of-pipeline contract, but no lawful model has ever been exported through it, so READY may be more accurate.
• Hopf observation is PARKED, but its kept-fibre map is built and unit-tested and its language test never ran. Only sector routing failed, so splitting the two forms may be more accurate.

How to read a mechanism page

Each page follows the same order:

1. In one paragraph: the idea and what it replaces, in plain words.
2. Figure: a picture of the mechanism at work.
3. How it works: the mechanism and its maths, labelled as definition, measurement or hypothesis.
4. Where it stands: its status, the evidence behind it with sources, and for a PARKED mechanism what would reopen it.
5. Go deeper: the code, the archived records and the nearest neighbours in the tree above.

Read the status first, then the "Where it stands" table. A claim there is a measurement only if it names a recorded run.

Go deeper

• Glossary for terms such as admission, FIT and export gap.
• The geometry page for pictures of the operators.
• Decisions log for D22 (nothing is closed for good), D37 (offline float training) and D11 (integer serving).
• Archived lab records for everything dated before 18 October 2026.