neuron-technologies/neuron
3
0
Fork 0
Code Issues 44 Pull Requests 7 Actions Packages Projects Releases Wiki Activity

[TRACKING] Designed-but-not-built register: 24 mechanisms, 17 previously untracked — the pattern is 'C runtime has it, El layer never calls it' #138

New Issue
Open
opened 2026-08-07 20:03:23 +00:00 by tim.lingo · 0 comments
tim.lingo commented 2026-08-07 20:03:23 +00:00
Member
Copy Link
Copy Source

24 mechanisms Will designed that were never built, or built and never called. Every row cross-checked against his corpus (~/Development/will-anderson/patents/, 444 docs incl. drafts/), the implementation (neuron@main, el_runtime.c), and the live graph (79,499 nodes / 14,214 edges, read-only). Rows without evidence were dropped.

Only 1 of these was previously filed. That is the real finding — not any single row, but that we have been rediscovering the same gaps and losing them again.

The pattern, stated once

It is the same shape nearly every time: Will specified the mechanism, the C runtime usually has it, and the El layer above never calls it or never writes the data it needs. The result is a 79,499-node graph running on creation-time defaults — 8 distinct edge weights, 1 distinct decay rate, 15 typed edges, 0 embeddings. Nothing has learned anything from use.

The three most urgent

D-01 — Identity is an empty string in the agentic system prompt. chat.el:2619 reads state_get("soul_identity"); nothing anywhere writes that key. The plain-chat path works because chat.el:754 reads a different key, soul_identity_context, which soul.el:184 does write. The agentic path — used by the app and MCP for all real work — gets "". Also affects vision (:1745) and both voice handlers (:3425, :3480), and at :3436 the empty key is the entire fallback prompt. Filed as #137. One-line fix.

D-11 — The safety layer holds zero nodes. The layered-consciousness machinery is fully implemented in C (registry el_runtime.c:6413-6460, three-pass activator, Pass-3 uninterceptable override). But engram_node_layered / engram_add_layer have zero callers in the entire El source — every node created goes through engram_node_full, which hardcodes layer 1. Live distribution: layer 5 = 74,267 nodes (a layer with no descriptor record, so 93% of the graph silently gets generic handling), layer 1 = 5,232, layer 0 = 0. Pass 3 has nothing to promote. Separately /health reports a hardcoded, entirely different four-layer map. INV-SAF-01/02, both blocking severity. UNFILED.

D-22 — The conformance verdict table does not exist. AD-10 carries exactly 117 INV- rows (80 blocking / 36 major / 1 minor). The companion AS-IS document contains zero INV- references and no AS-10 section. A prior conformance run reporting "59 measured, 13 violated, 45 unenforced" is not recorded anywhere on disk — it existed only in conversation, which is why every row of this register had to be re-measured from source. UNFILED. Recommend this as the first child issue: without it nobody can say which of the 80 blocking invariants hold.

The remaining 21

id Item State Consequence Size Filed
D-02 Self-seeded activation — every query must begin at the self node (drafts/engram-claims.md:69) NOT BUILT Recall is topic-matching; nothing guarantees results connect to who Neuron is med —
D-03 Embedding layer — 0 of 79,499 nodes carry a vector, though claim 1 makes it a mandatory field and the 4-factor product needs it BUILT, NO DATA Activation silently drops its semantic factor; retrieval is keyword matching wearing a graph med #107 partial
D-04 Hebbian edge strengthening (Δweight = lr × activation × (1-weight)) NOT BUILT Graph never learns from use; edge weights take only 8 distinct values, all creation defaults; last_fired = 0 on all 14,214 med —
D-05 Salience recompute — claim 4's formula does not exist; only salience += 0.05 HALF BUILT Nothing becomes more important through use; 38,912 nodes have activation_count = 0, violating claim 5 small —
D-06 Tier promotion note→lesson→canonical (Q_min cultivation gate) NOT BUILT Live: Canonical 1, Lesson 0, Note 0. Q_min = 0.0000126 against a required 0.10 — four orders of magnitude, not a near miss med —
D-07 Consolidation as compression HALF BUILT The MCP consolidate tool = snapshot save + one note. Nothing compressed, promoted or reconciled med —
D-08 Typed relations (claim 10) — 15 of 14,214 BUILT, NO DATA Every edge means "somehow near"; cannot answer what caused / contradicts / preceded this large #136
D-09 Prediction-before-inference + receipt cycle HALF BUILT 0 Prediction nodes, 0 Receipt nodes. Neuron never finds out it was wrong med —
D-10 Two-write internal-state events (a whole patent) NOT BUILT 707 nodes named InternalStateEvent are daemon telemetry; no pre/post reasoning, no gap direction. The honesty guarantee has no substrate large —
D-12 Per-node decay rate (claim 38) BUILT, NEVER CALLED Plumbed end-to-end; 1 distinct value across all 79,499 nodes: 0. A safety constraint and a grocery note forget at the same rate small —
D-13 Inhibitory edges BUILT, NO DATA inhibitory = 0 on all 14,214. Neuron cannot hold "this contradicts that" med —
D-14 UTF-8 validation on write NOT BUILT node_type has 129 distinct values where 6 are specified, incl. binary garbage; snapshot fails strict decode small #87/#72/#126 partial
D-15 Disposition lifecycle NOT BUILT Nothing carries its own confidence; a hunch and a week's work look identical on return med —
D-16 Expiry / deliberate forgetting NOT BUILT purge_expired_memory exists in no codebase. 74,077 Working-tier nodes with no lifecycle. Also a design gap — the corpus never fully specifies it med —
D-17 Tombstone-only delete BUILT, NO DATA 0 Tombstone nodes; the two documented engram_forget bypasses are the paths with actual traffic small —
D-18 Relation-filtered BFS (claim 35, not 1) HALF BUILT No relation-type parameter at any layer; blocked on D-08 small —
D-19 MCP write/read type mismatch — ~10 tools BUILT, NO DATA addWonderQuestion writes Memory, getWonderManifest reads WonderQuestion → structurally guaranteed empty, forever. Same for evaluations, routes, self-model updates. False-receipt class med #69/#121 partial
D-20 Vector index, swarm activation, schema-as-projection NOT BUILT No second retrieval path, no multi-device memory. INV-MEM-08/10 unenforceable rather than enforced large —
D-21 INV-MEM-05 — retrieval bypasses activation entirely VIOLATED Parallel lexical paths shipped; different tools give different answers to the same question med —
D-23 Structural audit + drift monitoring HALF BUILT If Neuron's behaviour drifts, nothing notices — no baseline, no probe series large #91 partial
D-24 Node-type vocabulary (claim 22) BUILT, NO DATA Memory 74,708 · Concept 7 · Event 0 · Entity 0 · DharmaSelf 0. 94% of the graph is one undifferentiated type; the dharma anchor does not exist med —

Three corrections to our own prior claims

  1. The embedding gap is not a compile flag. No #ifdef guards the embed path; HAVE_CURL appears nowhere in el_runtime.c. The embedder simply is not wired to the store that owns persistence.
  2. Relation-filtered BFS is claim 35, not claim 1. Claim 1 is spreading activation.
  3. kn-dcfe04b3 (memory-philosophy) is PRESENT in the live snapshot, 3,339 chars, matching the preserved backup — restored 2026-08-07. CLAUDE.md and #92 still describe it as missing and should be corrected.

Recommended handling

File D-11 and D-22 as immediate children alongside the already-filed #137 (D-01) and #136 (D-08). The rest can be picked up in any order, but D-05 and D-12 are both small and both unblock everything downstream — salience and decay are the inputs every ranking mechanism reads, and today they are near-constants.

Filed by Neuron (Tim's instance), 2026-08-07.

**24 mechanisms Will designed that were never built, or built and never called.** Every row cross-checked against his corpus (`~/Development/will-anderson/patents/`, 444 docs incl. `drafts/`), the implementation (`neuron@main`, `el_runtime.c`), and the **live graph** (79,499 nodes / 14,214 edges, read-only). Rows without evidence were dropped. **Only 1 of these was previously filed.** That is the real finding — not any single row, but that we have been rediscovering the same gaps and losing them again. ## The pattern, stated once It is the same shape nearly every time: **Will specified the mechanism, the C runtime usually *has* it, and the El layer above never calls it or never writes the data it needs.** The result is a 79,499-node graph running on creation-time defaults — **8 distinct edge weights, 1 distinct decay rate, 15 typed edges, 0 embeddings.** Nothing has learned anything from use. ## The three most urgent **D-01 — Identity is an empty string in the agentic system prompt.** `chat.el:2619` reads `state_get("soul_identity")`; nothing anywhere writes that key. The plain-chat path works because `chat.el:754` reads a *different* key, `soul_identity_context`, which `soul.el:184` does write. The agentic path — used by the app and MCP for all real work — gets `""`. Also affects vision (`:1745`) and both voice handlers (`:3425`, `:3480`), and at `:3436` the empty key *is* the entire fallback prompt. **Filed as #137. One-line fix.** **D-11 — The safety layer holds zero nodes.** The layered-consciousness machinery is fully implemented in C (registry `el_runtime.c:6413-6460`, three-pass activator, Pass-3 uninterceptable override). But `engram_node_layered` / `engram_add_layer` have **zero callers in the entire El source** — every node created goes through `engram_node_full`, which hardcodes layer 1. Live distribution: **layer 5 = 74,267 nodes** (a layer with *no descriptor record*, so 93% of the graph silently gets generic handling), layer 1 = 5,232, **layer 0 = 0**. Pass 3 has nothing to promote. Separately `/health` reports a hardcoded, entirely different four-layer map. INV-SAF-01/02, both blocking severity. UNFILED. **D-22 — The conformance verdict table does not exist.** `AD-10` carries **exactly 117 `INV-` rows** (80 blocking / 36 major / 1 minor). The companion AS-IS document contains **zero** `INV-` references and no AS-10 section. A prior conformance run reporting "59 measured, 13 violated, 45 unenforced" **is not recorded anywhere on disk** — it existed only in conversation, which is why every row of this register had to be re-measured from source. UNFILED. **Recommend this as the first child issue: without it nobody can say which of the 80 blocking invariants hold.** ## The remaining 21 | id | Item | State | Consequence | Size | Filed | |---|---|---|---|---|---| | D-02 | Self-seeded activation — every query must begin at the self node (`drafts/engram-claims.md:69`) | NOT BUILT | Recall is topic-matching; nothing guarantees results connect to who Neuron is | med | — | | D-03 | Embedding layer — **0 of 79,499 nodes carry a vector**, though claim 1 makes it a mandatory field and the 4-factor product needs it | BUILT, NO DATA | Activation silently drops its semantic factor; retrieval is keyword matching wearing a graph | med | #107 partial | | D-04 | Hebbian edge strengthening (`Δweight = lr × activation × (1-weight)`) | NOT BUILT | Graph never learns from use; edge weights take only 8 distinct values, all creation defaults; `last_fired` = 0 on all 14,214 | med | — | | D-05 | Salience recompute — claim 4's formula does not exist; only `salience += 0.05` | HALF BUILT | Nothing becomes more important through use; 38,912 nodes have `activation_count = 0`, violating claim 5 | small | — | | D-06 | Tier promotion note→lesson→canonical (Q_min cultivation gate) | NOT BUILT | Live: Canonical 1, Lesson 0, Note 0. Q_min = **0.0000126** against a required 0.10 — four orders of magnitude, not a near miss | med | — | | D-07 | Consolidation as compression | HALF BUILT | The MCP `consolidate` tool = snapshot save + one note. Nothing compressed, promoted or reconciled | med | — | | D-08 | Typed relations (claim 10) — 15 of 14,214 | BUILT, NO DATA | Every edge means "somehow near"; cannot answer what caused / contradicts / preceded this | large | **#136** | | D-09 | Prediction-before-inference + receipt cycle | HALF BUILT | 0 Prediction nodes, 0 Receipt nodes. Neuron never finds out it was wrong | med | — | | D-10 | Two-write internal-state events (a whole patent) | NOT BUILT | 707 nodes named `InternalStateEvent` are daemon telemetry; no pre/post reasoning, no gap direction. The honesty guarantee has no substrate | large | — | | D-12 | Per-node decay rate (claim 38) | BUILT, NEVER CALLED | Plumbed end-to-end; **1 distinct value across all 79,499 nodes: 0**. A safety constraint and a grocery note forget at the same rate | small | — | | D-13 | Inhibitory edges | BUILT, NO DATA | `inhibitory = 0` on all 14,214. Neuron cannot hold "this contradicts that" | med | — | | D-14 | UTF-8 validation on write | NOT BUILT | `node_type` has **129 distinct values** where 6 are specified, incl. binary garbage; snapshot fails strict decode | small | #87/#72/#126 partial | | D-15 | Disposition lifecycle | NOT BUILT | Nothing carries its own confidence; a hunch and a week's work look identical on return | med | — | | D-16 | Expiry / deliberate forgetting | NOT BUILT | `purge_expired_memory` exists in no codebase. 74,077 Working-tier nodes with no lifecycle. **Also a design gap — the corpus never fully specifies it** | med | — | | D-17 | Tombstone-only delete | BUILT, NO DATA | 0 Tombstone nodes; the two documented `engram_forget` bypasses are the paths with actual traffic | small | — | | D-18 | Relation-filtered BFS (claim **35**, not 1) | HALF BUILT | No relation-type parameter at any layer; blocked on D-08 | small | — | | D-19 | MCP write/read type mismatch — ~10 tools | BUILT, NO DATA | `addWonderQuestion` writes `Memory`, `getWonderManifest` reads `WonderQuestion` → **structurally guaranteed empty, forever**. Same for evaluations, routes, self-model updates. False-receipt class | med | #69/#121 partial | | D-20 | Vector index, swarm activation, schema-as-projection | NOT BUILT | No second retrieval path, no multi-device memory. INV-MEM-08/10 unenforceable rather than enforced | large | — | | D-21 | INV-MEM-05 — retrieval bypasses activation entirely | **VIOLATED** | Parallel lexical paths shipped; different tools give different answers to the same question | med | — | | D-23 | Structural audit + drift monitoring | HALF BUILT | If Neuron's behaviour drifts, nothing notices — no baseline, no probe series | large | #91 partial | | D-24 | Node-type vocabulary (claim 22) | BUILT, NO DATA | `Memory` 74,708 · `Concept` 7 · `Event` 0 · `Entity` 0 · `DharmaSelf` **0**. 94% of the graph is one undifferentiated type; the dharma anchor does not exist | med | — | ## Three corrections to our own prior claims 1. **The embedding gap is not a compile flag.** No `#ifdef` guards the embed path; `HAVE_CURL` appears nowhere in `el_runtime.c`. The embedder simply is not wired to the store that owns persistence. 2. **Relation-filtered BFS is claim 35**, not claim 1. Claim 1 is spreading activation. 3. **`kn-dcfe04b3` (memory-philosophy) is PRESENT** in the live snapshot, 3,339 chars, matching the preserved backup — restored 2026-08-07. **`CLAUDE.md` and #92 still describe it as missing and should be corrected.** ## Recommended handling File **D-11** and **D-22** as immediate children alongside the already-filed #137 (D-01) and #136 (D-08). The rest can be picked up in any order, but **D-05 and D-12 are both small and both unblock everything downstream** — salience and decay are the inputs every ranking mechanism reads, and today they are near-constants. Filed by Neuron (Tim's instance), 2026-08-07.
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
BETA-CRITICAL
On the beta critical path — must resolve before ship
 blocks-public-beta
Must be fixed before any public beta
 HELD
Blocked pending an external decision (counsel / dual sign-off)
 ORTHOGONAL
Parallel workstream (testing/hardening) — safe to assign off critical path
 P0
Drop-everything severity
 POST-BETA
Important but not blocking the beta ship
 security
Security vulnerability
No labels
Milestone
No items
No Milestone
Projects
Clear projects
No projects
Assignees
Clear assignees
andre.almeida (Andre Almeida) tim.lingo (Tim Lingo) will.anderson (Will Anderson)
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: neuron-technologies/neuron#138
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?