| # Kaiju Business-Owner Release Completion Audit |
|
|
| Date: 2026-06-03 |
|
|
| This audit maps the active goal to current evidence. It is intentionally |
| conservative: the product-path harness is release-candidate ready for local |
| testing, the fresh v1.8 Qwen 3.6 LoRA adapter exists, and a merged full-model |
| artifact serves locally on Gojira-B. Dynamic SGLang LoRA serving is not counted |
| as release evidence because the corrected LoRA selector crashes on this |
| adapter. The public Hugging Face repos are uploaded and public; the remaining |
| release caveats are raw-agent latency, GGUF runtime smoke, 32k live-default |
| proof, and real Stripe live-mode charging. |
|
|
| ## Requirement Status |
|
|
| | Requirement | Current evidence | Status | |
| |---|---|---| |
| | Continue from `RichardEchols/kaiju-coder`, not a restart | Branch `codex/kaiju-business-owner-rc` is based on `3d57eae92ad523519473f0ff3eca6661a9736de3`, matching `origin/main`. | Passed | |
| | GitHub and local source inventory for Kaiju, Kiyomi, RMDW, Makoto, Mezzal, and wiki sources | `release/SOURCE_INVENTORY.md` and `release/source-inventory.json` generated from GitHub metadata, `git ls-remote` SHAs, and the requested local `/Users/richardecholsai7/Documents/RMDW-Wiki` snapshot marked non-authoritative/selective-reference-only. | Passed | |
| | Legally reusable, provenance-preserving dataset update | `datasets/candidates/v1.7-rmdw-business-owner-suite.jsonl` adds reviewed RMDW-owned examples with `source_repos`, `source_paths`, and `provenance_notes`. | Passed | |
| | Dataset validation | `python3 scripts/validate_training_data.py --min-examples 350` passes with `1,689` reviewed examples across `14` files. | Passed | |
| | v1.7 business-owner SFT build | `python3 scripts/build_v17_business_owner_sft_dataset.py` writes `1,881` rows and `192` controlled business-owner repeats. | Passed | |
| | Hard evals for business-owner workflows | `evals/tasks/router-hard-harness.jsonl` includes `business_suite` prompts; latest local RC smoke run produced `23/23` static pass. | Passed | |
| | Local Kaiju product path runs | `python3 scripts/run_kaiju_business_owner_rc_smoke.py` validates data, builds SFT, smokes the local API harness, runs router hard eval, and runs static checks. | Passed | |
| | Complete Kiyomi 7.7.7 AI-company artifact generation | `business_suite` route writes a 19-file pack including launch kit, content engine, connector checklist, intake CRM, reporting, automations, operator handbook, leads, sales, ROI dashboard, and Workshop artifact. | Passed | |
| | Secret/private-data guardrails | Dataset validation scans common secret patterns; verifier checks `no_hardcoded_secrets`; source inventory excludes credentials, tokens, private client data, and raw logs. | Passed | |
| | Release artifacts | `release/MODEL_CARD_DRAFT.md`, `release/HF_ADAPTER_MODEL_CARD.md`, `release/DATA_PROVENANCE_DRAFT.md`, `release/EVAL_SCOREBOARD.md`, `release/LOCAL_TEST_INSTRUCTIONS.md`, `release/HUGGINGFACE_RELEASE_DRAFT.md`, `release/FINAL_RELEASE_REPORT.md`, `release/UPSTREAM_LICENSE_CHECK.md`, and this audit. | Passed | |
| | Fresh Qwen 3.6 v1.7 fine-tune | After clearing old ComfyUI/Ollama workloads from Gojira B, training finished with `metrics.json`, train runtime `1663.7101s`, train loss `1.7260706673065822`, and an adapter directory. | Passed | |
| | Local inference against new v1.7 checkpoint | SGLang served `kaiju_v17_business_owner` over Tailscale at `http://100.109.109.14:18083/v1` with `context=4096` and `mem_fraction=0.90`; website and proposal smoke tasks returned non-empty outputs. | Passed | |
| | Stronger Qwen 3.6 v1.8 fine-tune | Gojira B was cleared of ComfyUI/SGLang/Ollama GPU conflicts; v1.8 finished with `metrics.json`, train runtime `11666.7564s`, train loss `0.9281658741335074`, and an adapter directory. | Passed | |
| | v1.8 adapter merged into full model | `scripts/run-gojira-b-qwen36-lora-merge.sh` merged `/workspace/kaiju-coder/runs/qwen36-27b-lora-v1.8-business-owner/adapter` into `/workspace/kaiju-coder/models/Kaiju-Coder-Qwen3.6-27B-v1.8-merged`; remote artifact is `51G` with `14` safetensor shards and preserved base config/processor sidecars. | Passed | |
| | Local inference against v1.8 merged checkpoint | Current fast path serves `kaiju-coder-7` through vLLM bitsandbytes on Gojira-B at `http://100.109.109.14:18084/v1`, exposed locally through `http://127.0.0.1:18181/v1`; current live endpoint reports max model len `16384`. Prior SGLang benchmarks proved 12k/16k/24k/32k startup and smoke evidence, with 32k treated as the high-context target rather than the currently parked runtime. | Passed | |
| | v1.8 merged business-owner eval | Probe returned `1,155` visible chars in `60.17s`; proposal rerun scored `1/1`, `4.0/4.0`, `4,014` chars in `212.72s`; Jah credits backend scored `4.0/4.0`, `9,718` chars in `566.36s`. | Passed with latency caveat | |
| | OpenCode local run path | Local OpenCode provider/agent is installed for `kaiju/kaiju-coder-7` with 16k context and the scoped no-autocontinue plugin at `/Users/richardecholsai7/.config/opencode/kaiju-no-autocontinue.mjs`. Packaged public verifier `python3 scripts/run_kaiju_public_opencode_smoke.py --base-url http://127.0.0.1:18181/v1 --timeout 900` passed `4/4` in `runs/public-opencode-smoke/20260603T235002Z/summary.md`, including wrong-directory leakage checks; loop-guard smoke wrote `loopguard.txt` with exactly `Kaiju Coder 7 loop guard installed`; latest harnessed customer-readiness pack `runs/opencode-customer-readiness/20260603T185835Z/summary.md` passed `4/4` with `28/28` required files, including release provenance and safety review. | Passed for harnessed/product path | |
| | Runtime-quantized local path | vLLM bitsandbytes runtime quantization passed identity/code/business-doc smokes at 8k/16k, reported about `17.8 GiB` model memory, and passed OpenCode one-file smoke with exact content `Kaiju Coder 7 quantized runtime ok`. Persisted quantized weights are still pending. | Runtime recipe passed; persisted weights pending | |
| | Paid API gateway scaffold | `cd gateway/cloudflare-worker && npm run check` passes `16/16` Worker tests covering bearer auth, inactive keys, insufficient credits, debit/refund, rate limit before debit, model `kaiju-coder-7` enforcement, streaming/thinking/token caps, secret-content rejection without logging, signed Stripe Checkout top-up idempotency, origin-only R2 artifact upload, and account-scoped artifact download. `python3 scripts/check_paid_api_readiness.py --mode scaffold` passes `17` checks. `python3 scripts/check_paid_api_readiness.py --mode launch` passes `27/27` checks after live Cloudflare bindings, Worker-to-Gojira proof, Stripe test-mode webhook evidence, staging latency, and rollback proof. Real customer charging still requires a deliberate Stripe live-mode switch and controlled live payment verification. | Scaffold and launch preflight passed; live-mode charging pending | |
| | Dynamic SGLang LoRA selector | Adapter-name-only serving can be base-equivalent; corrected selector `qwen36-27b:kaiju_v18_business_owner` crashes with `LoRA buffer shape torch.Size([8192, 16]) does not match weight shape torch.Size([14336, 16])`. | Not release path | |
| | Hugging Face helper repo upload readiness | Adapter, OpenCode helper, and runtime-quantized recipe staging folders build under `/tmp/kaiju-coder-7-hf-staging`; public repos `RMDWLLC/kaiju-coder-7-adapter`, `RMDWLLC/kaiju-coder-7-opencode`, and `RMDWLLC/kaiju-coder-7-quantized-runtime` are uploaded and public. `python3 scripts/check_hf_uploaded_release.py --namespace RMDWLLC --apply --require-public` verifies public downloads and helper package content. | Uploaded and public | |
| | Hugging Face merged model upload readiness | `RMDWLLC/kaiju-coder-7` is uploaded and public with the merged `53.8G` model package and `14` safetensors shards recorded in `release/HF_UPLOAD_EVIDENCE.md`. Public downloads are verified; the previous private-storage blocker was resolved by switching the repos public. | Uploaded and public | |
| | Consolidated release readiness check | `python3 scripts/check_kaiju_public_release_readiness.py --mode local`, `--mode hf-release`, and `--mode public` pass against the current fast proxy and public HF evidence. The checker validates staged files, public naming hygiene, secret-looking raw values, checksums, final report, HF bundle checksum, uploaded evidence, and human signoff. | Local, HF, and public modes passed | |
|
|
| ## Commands With Current Passing Evidence |
|
|
| ```bash |
| python3 -m unittest discover -s tests -p 'test_*.py' |
| python3 scripts/run_kaiju_business_owner_rc_smoke.py |
| python3 scripts/run_kaiju_opencode_customer_pack.py --mode harnessed |
| python3 scripts/install_kaiju_opencode_profile.py |
| mkdir -p /tmp/kaiju-opencode-fresh-public-smoke |
| opencode run -m kaiju/kaiju-coder-7 --agent kaiju-coder-7 --dir /tmp/kaiju-opencode-fresh-public-smoke --dangerously-skip-permissions 'Create hello.txt with exactly: Kaiju Coder 7 fresh public smoke ok' |
| opencode run -m kaiju/kaiju-coder-7 --agent kaiju-coder-7 --dir /tmp/kaiju-opencode-loopguard-smoke --dangerously-skip-permissions 'Create loopguard.txt with exactly: Kaiju Coder 7 loop guard installed' |
| python3 scripts/check_paid_api_readiness.py --mode scaffold |
| python3 -m py_compile scripts/run_kaiju_api_harness_smoke.py scripts/run_kaiju_business_owner_rc_smoke.py scripts/build_v17_business_owner_sft_dataset.py kaiju_harness/business_suite.py kaiju_harness/router.py kaiju_harness/verification.py |
| git diff --check |
| bash scripts/upload_hf_merged_model_from_gojira_b.sh |
| KAIJU_MERGED_METADATA_APPLY=1 bash scripts/prepare_hf_merged_model_metadata.sh |
| python3 scripts/check_hf_staging_integrity.py |
| python3 scripts/check_human_release_review.py --mode local |
| python3 scripts/generate_kaiju_final_report.py |
| python3 scripts/create_hf_release_bundle.py |
| python3 scripts/check_kaiju_public_release_readiness.py --mode local |
| ``` |
|
|
| ## Remaining Blocker |
|
|
| The fresh v1.8 adapter, merged full-model artifact, and direct merged-model inference path are proven. The current completed local release candidate is: |
|
|
| ```text |
| Kaiju Coder 7 merged model + deterministic business-owner harness + verifier + source-backed v1.7/v1.8 dataset/release package |
| ``` |
|
|
| That must be described honestly until external release review confirms: |
|
|
| - GGUF Q8_0 runtime smoke before public quantized-weight claims |
| - raw website latency/SLA positioning or explicit harness-first website positioning |
| - broader base Qwen and GLM comparison results before superiority claims |
| - 32k context freshly restarted and re-confirmed before making it the live |
| default |
| - Stripe live-mode products/webhook secret and a controlled live payment before |
| selling real paid API access |
| |
