# Phase 177 locally actionable completion evidence

Date: 2026-07-15

## Completion audit

Phase 177 now has seven unchecked entries. Each is an empirical training,
checkpoint investigation, or publication outcome that cannot truthfully be
completed with repository-local synthetic data:

| Task       | Required completion evidence                                                                                                                  | External dependency keeping it open                                                                |
| ---------- | --------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------- |
| 177.2.1.5  | Full every-layer Gemma-4-9B and Qwen3-8B SAE checkpoint matrix at x8/x16/x32/x64, with Gemma Scope-scale quality measurements                 | Model weights, activation corpus, large storage, and sustained distributed accelerators            |
| 177.2.1.6  | Trained Iris code and Kuan Yin safety-head SAEs plus published feature dashboards                                                             | Fine-tuned head checkpoints, domain activation corpora, accelerator run, and dashboard publication |
| 177.3.2.1  | Causal refusal-circuit findings on the served safety checkpoint and sovereign research note                                                   | Safety checkpoint, approved jailbreak/benign evaluation set, and empirical interventions           |
| 177.3.2.2  | Causal factual-recall findings across the three named African-knowledge domains                                                               | Served checkpoint, culturally reviewed evaluation corpus, and empirical interventions              |
| 177.3.2.3  | Iris code-generation circuit findings joined to real Metis bug patterns and SFT curation                                                      | Iris checkpoint, production linter findings, evaluation corpus, and empirical interventions        |
| 177.3.2.4  | Matched pre/post-RLHF deception and sycophancy circuit comparison                                                                             | Both checkpoint revisions, controlled evaluation corpus, and empirical interventions               |
| 177.10.2.1 | Published checkpoint-specific report containing the actual feature inventory, refusal circuit, steering vulnerabilities, and deception probes | Outputs of 177.2.1.5-6 and 177.3.2.1/4, reviewer approval, and publication channel                 |

The audit treated the executable seams inside those entries as locally
actionable and completed them without claiming the external outcomes:

1. `nous-sae` now expands model topology into the exact residual x16/x32/x64,
   MLP-post x8/x16, and attention-out matrix. Domain-qualified Iris/Kuan Yin
   runs receive collision-free artifact ids, resumable execution, quality
   evidence, and a dashboard-aware publication gate.
2. `nous-circuits` now defines all four named investigations with checkpoint,
   dataset, method, causal-component, consumer, and limitation provenance.
   Incomplete campaigns are visibly non-publishable; complete evidence renders a
   sovereign research note.
3. The AISI report builder requires every named report section, identifies UK
   AISI, US AISI, and Kuan Yin audiences, and stamps incomplete reports
   `DO NOT PUBLISH`.

## Completed local integrations

### 177.6.1.2 — Galatea VLA progressive columns

- `GalateaVLAProgressiveController` assigns each `(fleet, task class)` route a
  new column, exposes all prior columns as lateral sources, limits the optimizer
  to the active column, snapshots frozen predecessors, and rejects any mutation.
- Galatea's VLA deployment registry validates the ONNX/TensorRT artifact digest
  and the ordered digest chain for every frozen prior column before routing a
  fleet/task deployment.
- Tests prove prior VLA logits remain bit-for-bit identical after a later fleet
  learns and that tampered prior chains cannot deploy.

### 177.10.2.4 — SAE side-channel precognition

- Kuan Yin consumes only derived, ephemeral feature telemetry: model revision,
  hook, feature id, category, threshold crossing, and label fidelity. It does
  not ingest residual tensors or user content through this boundary.
- Thresholded features are fidelity-weighted, retry-deduplicated, grouped by
  category, and fused with behavioural risk using noisy-OR.
- Deception, coercion, and CSAM-adjacent signals can fire before behavioural
  manifestation. Neural telemetry alone may add protective friction but cannot
  hold a user; CSAM-adjacent evidence always requires human review and never
  constitutes an automatic accusation.
- Cross-package Kuan Yin integration coverage proves the pre-behaviour warning
  reaches the real precognition export.

## Requirement-to-evidence map

| Local scope                 | Authoritative repository evidence                                 | Automated proof                                                               |
| --------------------------- | ----------------------------------------------------------------- | ----------------------------------------------------------------------------- |
| SAE suite orchestration     | `libs/nous/sae/nous_sae/suite.py`                                 | Hook/scale matrix, domain isolation, evidence and dashboard publication gates |
| Targeted campaigns          | `libs/nous/circuits/nous_circuits/campaigns.py`                   | Four canonical specs, incomplete-evidence rejection, causal note rendering    |
| SAE safety report           | `libs/nous/circuits/nous_circuits/safety_report.py`               | Required-section gate and AISI audience rendering                             |
| Galatea training boundary   | `libs/nous/progressive-nets/nous_progressive_nets/galatea.py`     | Route uniqueness, lateral reuse, optimizer isolation, bit-exact retention     |
| Galatea deployment boundary | `libs/galatea/ai/src/vla-runtime/progressive-vla-deployment.ts`   | Artifact/prior-chain verification and fleet/task routing                      |
| Kuan Yin side-channel       | `libs/kuanyin/precognition/src/sae-feature-signals.ts`            | Early warning, deduplication, provenance rejection, review-only CSAM handling |
| Kuan Yin integration        | `libs/kuanyin/evaluation/src/__tests__/integration-flows.spec.ts` | Precognition package export and protective-friction flow                      |

## Verification

All 2,187 affected tests pass: 34 `nous-sae`, 23 `nous-circuits`, 19
`nous-progressive-nets`, 1,895 Kuan Yin precognition, 98 Kuan Yin integration,
and 118 Galatea AI tests. The three Python packages also pass Ruff, mypy, sdist,
and wheel builds. Kuan Yin precognition/evaluation and Galatea AI pass their
declared Nx lint, typecheck where declared, and build gates (including required
dependency builds).

Verification ran sequentially and targeted on a 15 GiB host with 31 GiB swap
because a development watcher and other agents were active; no full-workspace
sweep was run.

These changes are headless model-training, deployment-contract, safety-logic,
and documentation work. They do not alter web, PWA, mobile, Unreal, or rendered
UI, so Playwright, mobile, and Unreal automation are not applicable.

## Remaining external outcomes

The seven items in the audit table intentionally remain unchecked until their
specified real training, checkpoint, evaluation, review, and publication
evidence exists. There are no remaining open repository-locally actionable Phase
177 tasks.
