> ## Documentation Index
> Fetch the complete documentation index at: https://docs.bensevern.dev/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Repo thesis weaknesses

> Where the codebase does not yet meet the governing architecture frame (one product, two engines, many surfaces) — the thesis-conformance weakness board and live-harvested drift signals, generated and gated in CI.

The [suite matrix](/docs/suite-matrix) renders cross-language *surface* parity. This page renders the codebase's standing against the **governing architecture frame** — *one product, two engines, many surfaces* ([frame](https://github.com/benseverndev-oss/goldenmatch/blob/main/context-network/architecture/one-product-two-engines.md), decision 0047) — scored by its five decision tests (tenets). Everything below the line is generated from the thesis-conformance scorecard (`parity/thesis_conformance.yaml` + the live static harvest in `scripts/check_thesis_conformance.py`, via `scripts/gen_thesis_weaknesses.py`) and verified in CI, so it can't drift from the gate.

## Standing

**3** open weakness(es) — 🟢 low 3. **8** resolved (archived below). Undeclared (a real divergence with no conformance level / contract): **0**.

A weakness is a place the codebase does not yet fully meet the frame; each maps to one of the five decision tests (tenets). *Declared* means the divergence has a conformance level and a test — an accepted, contained cost — not an accident. The `declared: false` class is the dangerous one and is flagged below. Resolved items are archived off this live board (conformance v2 — the live list is the real risk surface, not a museum of closed wins) but kept for the record.

## The five tenets

| Tenet  | Decision test                                                                                 | Open |
| ------ | --------------------------------------------------------------------------------------------- | ---- |
| **T1** | One authoritative semantic owner per capability (no second source of truth).                  | 2    |
| **T2** | Conformance defines correctness (every divergence has a declared level + a test).             | 0    |
| **T3** | Kernelize on measurement (deferrals classified; no silent regress to fallback).               | 0    |
| **T4** | Compute vs control stay architecturally distinct (explicit, versioned seam).                  | 1    |
| **T5** | Arrow at bulk boundaries, not the universal calling convention. \[DECLARED WON -- see header] | 0    |

## Weakness board

Ranked by severity. *Routing* applies the conformance-v2 default-routing test (0047 amendment): a shared owner the default caller path does not use is a latent second source of truth, even when a fixture proves parity.

| Severity | Tenet | Declared | Routing                   | Weakness                                                                                                                                                                     |
| -------- | ----- | -------- | ------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 🟢 low   | T4    | yes      | —                         | Compute/control shared frame-residency budget: now a contract field (was env-only)                                                                                           |
| 🟢 low   | T1    | yes      | opt-in (fallback default) | Semantic-layer structural certification (uniqueness-at-grain + fan-out) is single-sourced through key-integrity-core; Python-native + TS-wasm + SQL all bind that one kernel |
| 🟢 low   | T1    | yes      | opt-in (fallback default) | Clustering + goldenanalysis frame kernels: both families now share a kernel (cluster-wasm + analysis-wasm)                                                                   |

## Live harvest (auto-detected)

Static signals harvested directly from `parity/*.yaml` + each package's `_native_loader.py` (no toolchain) — so drift shows up here without editing the curated board above.

**Scorer-kernel coverage (goldenmatch).** 22 of 24 scorers are kernel-backed; 2 declared-deferred; uncovered (must be empty — coverage floor): none.

| Deferred scorer    | Reason                                                                                            |
| ------------------ | ------------------------------------------------------------------------------------------------- |
| `embedding`        | n/a -- model-backed (sentence-transformers / Vertex); belongs to goldenembed, not a string kernel |
| `record_embedding` | n/a -- model-backed multi-field embedding                                                         |

**Fallback-only kernels** (a `-core` symbol the host references but the default does not run — `_FALLBACK_ONLY`):

* `goldenflow`: `phone_validate`

**Cross-language surface gaps** (declared Python-only / TS-only per surface — the same partition the [suite matrix](/docs/suite-matrix) renders in full):

| Package       | Surface              | Python-only | TS-only |
| ------------- | -------------------- | ----------- | ------- |
| `goldencheck` | cli\_commands        | 5           | 0       |
| `goldencheck` | mcp\_tools           | 1           | 0       |
| `goldenflow`  | cli\_commands        | 4           | 0       |
| `goldenflow`  | transforms           | 0           | 1       |
| `goldenmatch` | a2a\_skills          | 15          | 13      |
| `goldenmatch` | blocking\_strategies | 2           | 0       |
| `goldenmatch` | cli\_commands        | 10          | 0       |
| `goldenmatch` | mcp\_tools           | 10          | 0       |
| `goldenpipe`  | cli\_commands        | 1           | 0       |

## Resolved (archived)

Weaknesses verified resolved-and-stable and moved off the live risk board (conformance v2, 0047 amendment #4). Kept for the record — the evidence + reasoning live in `parity/thesis_conformance.yaml`; a rotted premise un-archives.

| Tenet | Weakness                                                                                                                  |
| ----- | ------------------------------------------------------------------------------------------------------------------------- |
| T1    | goldenmatch a2a\_skills two-sided gap is now all intentional divergence (MDM trio reconciled)                             |
| T3    | Kernel deferrals state a reason but not measured wall-clock provenance                                                    |
| T1    | TS Fellegi-Sunter scoring runs the shared fs-core kernel by default (batteries entry); pure-TS is the classified fallback |
| T4    | Incremental resolution against a persisted identity index (C2/C4 landed)                                                  |
| T4    | Compute->control handoff is a versioned ResolutionBatch consumed by apply\_batch(store, batch)                            |
| T2    | Standardize divergence now has a runnable characterization test (was prose-only)                                          |
| T1    | Golden-record rollup single-sourced through core.golden (was a 2nd hand-rolled rule in resolve.py)                        |
| T2    | Cross-surface scorer/package divergences are now declared in the 0046 verdict table                                       |