What is still open
The project has built and tested many of the parts of a language model. The few it has trained together (the discrete writer, the reads and the ranking loss) have not yet beaten a simple count table, and a full trunk of lawful parts has never been trained. The biggest open question is simple to state: can a model that uses no softmax learn to rank the right word above the wrong ones? Nobody knows yet. Beside it sits a list of tests that were never run, or that failed only because the test was set up badly. Most of them are cheap, and the cheapest decisive one is FIT-reviewed and queued, not yet released.
flowchart TD
W2["W2: read-code capacity capture<br/>(FIT-reviewed, queued, about $0.91)"] --> Q{"Which wall is it?"}
Q -- "representation (H_rep)" --> R["Fix the read and the address metric.<br/>Do not fund float training."]
Q -- "learning law (H_law)" --> W3["W3: BVG-K batched discrete moves<br/>(about $12)"]
W3 --> Q3{"Does integer top-1 beat<br/>both count tables?"}
Q3 -- yes --> NX["Add the missing trunk slots,<br/>then export"]
Q3 -- no --> W4["W4: one float-credit sheet<br/>(about $5 to $10 GPU)"]
The order of the next decisive tests in the sequence the owner approved (D39 item 4); each is FIT-reviewed, none is released. W1, the K16 run, has already finished with no KEEP.
Terms follow the glossary. Claims are labelled proof, measurement or hypothesis as in formal_vocabulary.md. Dollar figures are the chief architect's estimates, not recorded costs.
How the output head works, in two lines
The model first picks a shortlist of candidate cells, the nearest K of 256 by address. A row counts as admitted if the right token survives that cut. It then ranks the shortlist and takes the nearest token. So a model can fail at admission, at ranking, or at both. Source: docs/labs/repo-renormalization-2026-10-19/audit/synthesis.md section 1 and K16 outcome.
The central unknown
Can a softmax-free read learn ranking? This is a hypothesis, and the biggest one. Every float read that learned had softmax helping it somewhere in training. The best read without softmax, arm W, is a recorded REJECT. It was worse by 0.0172 BPB (bits per byte) and ran 15% slower, and its gradient came from a softmax-shaped surrogate. The starling read reached recall 0.329 where softmax reached 0.999. Both verdicts close those configurations only (decision D22), not the idea. See findings for the numbers and flock read for the discrete read that replaces them.
What we know today: the best trained discrete model scored top-1 of 65 and 463 of 14,287 development rows, below a unigram count table (1,226) and a context-free successor table (6,300) (measurement, K16 outcome). The cause is open. The record rules out one suspect: aliasing did not explain the ranking failure. It leaves ranking capacity, the read, the address key and the training schedule.
FIT-reviewed and queued
These are the closest to a decision. A FIT review is not an owner release: spending waits for the owner (charter §5.1).
- W2, the read-code capacity capture (about $0.91, already reviewed as FIT, same lease). It decides whether the wall is the representation (H_rep: the stored code cannot separate the answers) or the learning law (H_law: the code could, but the update rule does not find it). It also adds the count-table arms that the headline bars lacked. Source: architect review section 6,
docs/labs/course-correction-2026-10-18/audits/d37-reactivation-architect.md. - The admission-preserving rerun of A8. Pre-registered and reviewed FIT, never run. Its controller and a separate finite qualification producer are merged. The qualification is NOT_RUN. Source: admission-anchor controller.
Every step that trains also needs an owner release (charter §5.1), and training stays HELD until the owner releases it.
Never fairly tested, and cheap
| Test | What it would tell us | Source |
|---|---|---|
| Full-vector or interior-block address metric, at zero training | The architect's likely lever. The address key takes only 5 values to order up to 256 cells (proof, read from discrete.cu) | architect review §3.2 |
| A wider carrier (more roots or blocks) | Whether 16 bits per vector limits capacity. Needs new kernel code and a fresh CUDA qualification, so not free | discrete learning |
| Read-off against read-on for Route and Flock, on existing checkpoints | Whether the reads add anything at all. Route got 20 proposals in 16,404 batches | Route read; attention state |
| The prime-zeta code against random and quota-only injective nulls, at zero updates | Whether its structure matters or only its uniqueness | prime-zeta codebook |
| Whether a secant surrogate gives descent for the recurrence gates and SpiralCore/FBS | A small float check before any larger run | recurrence gates; SpiralCore/FBS |
| 4,096 against 16,384 tokenizer pieces | Whether vocabulary size is a lever. The trainer exists | tokenizer |
Never tested, moderate cost
- Ranking with admission intact. No run has had admission intact and then failed or succeeded at ranking (admission).
- A rank objective that keeps cells occupied. The registered route after the K16-R run collapsed to 19 or 20 occupied cells (rank margin).
- Full-epoch and frequent-token revisit schedules. One pass sees 43.3% of the training targets (proof, by arithmetic, model training).
- The finite relational head. K16-A reached its 16,404 batches and can satisfy the prerequisite for the next stage (architect review §5; relation read).
- Discrete Hamming binding (VSA) on the current writer (VSA).
Never tested, expensive or blocked
- A trunk built only from lawful parts. Train it with float in a new scanned module, then export. The 24-kernel extract has passed finite controls only (geometric stack). No exporter exists for it.
- The forced-copy oracle for the pointer. The cheapest decisive pointer test. It needs a lawful pointer first, because running it on the old softmax pointer would execute a banned model (pointer key).
Failed only on configuration
A negative closes the configuration, never the mechanism. These tests were flawed as set up, so they say little about the mechanism.
| Test | What was wrong with it |
|---|---|
| A2 pilot, Dose 1, K16-A and K16-R | All started from a seeded codebook with at most 16 distinct vectors |
| P8 (127 of 3,847) | Uniform placement at K = 8 |
| A8 first dose | Ran without the admission guard |
| Starling and rank-table reads | Used softmax-tuned recipes and softmax-surrogate gradients |
| E8 memory | Used a sigmoid gate and 2,000 CPU steps |
| Spin weights | Met its bars (1.5148 BPB, level with dense) but only through a dense contraction, which D25 bans; PARKED |
| M2 reply model, 437 of 512 then 0 of 128 | Panel fitting on rows already seen |
| Read binding | Measured on a softmax read |
| VSA | Used the banned CPU learner |
Source: docs/labs/repo-renormalization-2026-10-19/audit/synthesis.md section 2, which cites each mechanism's dossier in docs/labs/repo-renormalization-2026-10-19/audit/mechanisms/.
Not yet buildable
The map of lawful parts still has gaps that no test can fill until someone builds them (docs/labs/repo-renormalization-2026-10-19/audit/synthesis.md section 1).
- Position. No mechanism is assigned. The only related result is about recurrent state, and it is confounded (see findings).
- Normalization. None is designed.
- Address as the read. Never built, and nothing yet consumes the object store.
- A lawful exporter for any trained trunk, so the export gap is unmeasured.
- A faster writer. The writer scans the whole vocabulary, about 5 million comparisons per row, which breaks the per-token sparsity goal at larger vocabularies (proof, by count, architect review §3.5).
Open claims
| Claim | Kind | Source |
|---|---|---|
| Quaternion transport helps order in a lawful model | Hypothesis. The only result is confounded | findings |
| Float training of the table beats discrete moves | Hypothesis. W4 is conditional on W2 and W3 | architect review §6 |
| The export gap of a float-trained model is small | Hypothesis. The architect expects a structurally non-zero gap | architect review §3.3 |
| A norm-preserving Hamilton gate can suppress a signal | Hypothesis, a reading of the dossier only | Hamilton gate |
| Near-frontier capability needs more data than any trained stack used | Measurement for the past (#437, #1913), hypothesis for the future | findings |
What counts as success for the first milestone
The charter sets the M1 closing bar: integer top-1 through the D11 export beats both count tables on the same development panel, on each of two seeds, with tokens per second reported. A bits-per-byte bar returns later, once a lawful likelihood exists (charter §4.8). That settles the old open question about scoring a model with no likelihood, for M1 only. A lawful likelihood is still unbuilt. A conversation bar (M2) is further off, and a float student would need a new owner decision (docs/labs/repo-renormalization-2026-10-19/audit/synthesis.md section 4, stage 5).
Go deeper
- What the project knows and how it got here.
- The audit's shortest-path plan:
docs/labs/repo-renormalization-2026-10-19/audit/synthesis.mdsection 4. - Mechanism board and
NOW.mdfor the live plan.