This commit is contained in:
2026-06-05 15:01:34 +03:00
parent 50180aaa14
commit 4cef6af041
27 changed files with 13547 additions and 892 deletions

View File

@@ -9,6 +9,8 @@ description: Reference manual for GRACE-Poly v2.6 — syntax formats, complexity
@RELATION DISPATCHES -> [Std.Semantics.Python]
@RELATION DISPATCHES -> [Std.Semantics.Svelte]
@RELATION DISPATCHES -> [Std.Semantics.Testing]
@RATIONALE GRACE-Poly exists because autoregressive Transformers suffer from four architectural defects that make them unreliable at scale: (1) KV-cache eviction — after ~8K tokens early context is lost, so decisions from file #1 are forgotten by file #4; (2) attention sink — in files >400 LOC attention weights diffuse, making nested structures invisible; (3) hallucination by design — when a dependency is missing the model confabulates a plausible one instead of signaling uncertainty; (4) copy-paste regression — similar code is duplicated including rejected patterns. The protocol's anchors, relations, and decision-memory tags form an external cognitive exoskeleton that survives context compression and provides structured navigation where raw prose fails.
@REJECTED Trusting natural language comments for navigation was rejected — they lack syntactic density and are the first to be evicted under CSA compression. Docstring-only contracts were rejected — they are invisible to the semantic index and cannot be verified structurally. Ad-hoc conventions per agent were rejected — 44% orphan rate in this project proves that without a dedicated curator, the semantic graph degenerates within 3-4 sessions.
## 0. SSOT DECLARATION
@@ -152,4 +154,84 @@ All agents use Axiom MCP for GRACE-semantic operations. This is the canonical to
- `skill({name="semantics-svelte"})` Svelte 5 (Runes), UX state machines, Tailwind
- `skill({name="semantics-testing"})` pytest/vitest test constraints, external ontology
## VIII. ATTENTION ARCHITECTURE & OPTIMIZATION RULES
The GRACE anchor format is not arbitrary it is optimized for the specific attention compression mechanisms in the underlying model (MLA CSA HCA DSA sliding window). Understanding these mechanisms is critical: a contract that violates these rules becomes invisible to the model after context compression, causing downstream hallucination.
### Attention Compression Pipeline
| Layer | Compression | Mechanism | What Survives | What Dies |
|-------|:----------:|-----------|---------------|-----------|
| **MLA** | 3.5× | KV vectors compressed to 576d latent codes. Information density per token is paramount. | Dense tokens (symbols, brackets, semantic tags). | Verbose prose, long descriptions. |
| **CSA** | 4× + topk sparse | Every ~4 tokens pooled into 1 KV record. Only topk records selected per query. | Contracts in 1-2 anchor lines. | Contracts spread across 15+ lines details lost in pooling. |
| **HCA** | 128× | Aggressive pooling over distant context. Dense attention computed on compressed records. | Statistical signatures: hierarchical IDs (`Core.Auth.Login`), repeated `@SEMANTICS` keywords. | Flat IDs (`LoginFunction`) become noise. One-off tag values. |
| **DSA** | Lightning Indexer | Fast linear scorer estimates relevance of each compressed record to query keywords. | Records whose `@SEMANTICS` match query keywords. | Records with different naming than the query. |
| **Sliding window** | None (preserved) | Small window of recent uncompressed tokens for local detail. | Contracts 150 lines fit entirely in the window. | Contracts >150 lines partially invisible. |
### ATTN_1 — FIRST-LINE DENSITY (CSA + MLA)
The opening anchor MUST pack maximum signal into one line:
```
#region Domain.Sub.ContractId [C:N] [TYPE TypeName] [SEMANTICS tag1,tag2,tag3]
```
- ID, complexity, type, and semantic tags on ONE line → survives CSA 4× pooling as a single KV record.
- `@BRIEF` on line 2 is secondary — it may be pooled separately.
- **NEVER** spread the anchor signature across multiple lines in a CSA-sensitive context.
### ATTN_2 — HIERARCHICAL IDS (HCA 128×)
Contract IDs MUST use dot-separated domain prefixes with 2-3 levels of hierarchy:
- `Core.Auth.Login` → after HCA 128×, `Core.Auth` survives as a statistical signature.
- `Core.Auth.Session` → same domain group; `Auth` signature reinforced.
- `users_login`**dies** at 128×, indistinguishable from noise.
**Rule:** Every non-C1 contract ID carries at least 2 levels: `Domain.Name`. C1 contracts (DTOs, constants) inside a hierarchical parent module may use single-level IDs — the parent provides the domain context.
**Good:** `Core.Auth.Login`, `Migration.RunTask`, `Users.ListModel`, `Tasks.TaskCard`, `Test.Migration.RunTask`
**Bad:** `login_handler`, `migrate`, `format_timestamp`, `UserListModel` (missing domain prefix)
**Stack disambiguation:** Use domain prefix, not stack prefix. The file path already encodes the stack (`backend/src/` vs `frontend/src/`):
- Backend: `Core.Auth.Login`, `Api.Dashboards.List`, `Plugin.Translate.Execute`
- Frontend: `Users.ListModel`, `Tasks.TaskCard`, `Dashboards.Hub`
- Tests: `Test.Core.Auth`, `Test.Users.ListModel`
### ATTN_3 — SEMANTIC GROUPING (DSA Lightning Indexer)
The DSA Indexer scores compressed records by keyword match against the query. `@SEMANTICS` keywords are your index keys:
- All contracts in the `auth` domain MUST share `[SEMANTICS auth, ...]`.
- `grep "@SEMANTICS.*auth"` → Indexer scores all auth records high.
- If one auth contract uses `@SEMANTICS login` and another `@SEMANTICS authentication`, the Indexer may fail to group them.
**Rule:** Identical domain = identical primary keyword in `@SEMANTICS`. Secondary keywords can vary.
### ATTN_4 — FRACTAL BOUNDARIES (Sliding Window)
The sliding window preserves recent tokens without compression. A contract ≤150 lines fits entirely in the window and is fully visible to the attention mechanism:
- Contract ≤150 lines → guaranteed full visibility.
- Module ≤400 lines → manageable in a few attention passes.
- INV_7 (Module < 400 lines, CC 10) is not just a style rule it ensures the model can physically see the entire contract structure.
### Grep Heuristics (Zombie Mode — when MCP tools are unavailable)
When Axiom MCP is down, these grep patterns exploit the DSA Indexer's keyword sensitivity:
```bash
# Find all contracts in a domain (Indexer matches @SEMANTICS keywords)
grep -r "@SEMANTICS.*<domain>" src/
# Find API type binding (cross-stack traceability)
grep -r "@DATA_CONTRACT.*<ModelName>" src/
# Extract full contract body (awk, respecting fractal boundaries)
awk '/#region <ContractID>/,/#endregion <ContractID>/' file.py
# Find all contracts BIND_TO a store
grep -r "BINDS_TO.*\[<StoreId>\]" src/
```
#endregion Std.Semantics.Core