# CASAN — Team Handoff & Push-to-90 Plan

**Repo:** `Output_CASAN5_REFINED/AINative_OKR_CASAN5` (the active package — GHCP is deprecated)
**Status date:** 2026-06-28
**Goal:** every harness H1–H7 **above 80**, ideally ~90, earned against real execution (no faked evidence).

---

## PART 1 — Where we are now (status report)

### Current independent scores (verified, not self-claimed)

| ID | Harness | Score | State |
|----|---------|:---:|---|
| H1 | Context | **82** | Real incremental `pipeline-context.yaml` from a real run; 12 distinct traces; artifacts on disk |
| H2 | Tool | **80** | Per-agent permission enforced + in execution path; signed tamper-evident tool audit; rollback required |
| H3 | Evaluation | **82** | Real app + real unit/e2e tests that run and can fail; real golden regression; real LLM-judge REJECTED→fix cycle |
| H4 | Security | **80** | Input normalization defeats bypasses; pii-rules wired; output fail-closed on secrets |
| H5 | Governance | **82** | RSA-anchored audit chain (re-forge detected); approver+output_hash hashed; separation of duties |
| H6 | AgentOps | **80** | Real latency/alerts; hallucination detector populated; provider-telemetry cost source |
| H7 | Orchestration | **80** | Real DAG run with real BACK-TO-PLAN; per-step CASAN wrapping; fallback/drift/rollback invoked |
| | **Average** | **~81** | **CASAN Level 4 (Automated), genuine** |

### How this was reached
- **Phase 1 (harness hardening, by Claude):** lifted H2/H4/H5/H6 from ~50s to ~80 with adversarial-verified controls. See [phase1-hardening-reassessment.md](phase1-hardening-reassessment.md).
- **Phase 2 (real app + pipeline run, by Codex — independent builder for impartiality):** lifted H1/H3/H7 by building a real NestJS+Prisma+React app and running the pipeline for real. Audited independently. See [phase2-independent-audit.md](phase2-independent-audit.md).

### Verify the current state (run these — all must pass)
```bash
cd Output_CASAN5_REFINED/AINative_OKR_CASAN5
npm test -w backend && npm run build -w backend && npm run build -w frontend
bash .specify/tests/run-casan4-harness-tests.sh          # 35 PASS / 0 FAIL
bash .specify/tests/adversarial-harness-tests.sh         # 22 PASS / 0 FAIL
bash .specify/scripts/bash/verify-audit-chain.sh         # AUDIT_CHAIN_VALID anchor=signed
bash .specify/scripts/bash/verify-tool-audit.sh          # TOOL_AUDIT_VALID anchor=signed
```

### Honest residual gaps keeping us at ~81 not ~90
1. **H7** — in-run rollback still writes a marker (real undo exists only as a standalone demo); drift still compares a file to a copy of itself; model-fallback uses a synthetic trigger.
2. **H6** — cost uses a sample provider record reused for every step (real latency, not real per-step billing).
3. **H2** — no runtime rate-limit counter; no per-call JSON-schema validation of tool I/O.
4. **H4** — rule/normalization-based only (no semantic/embedding detection); no tool sandbox/timeout.
5. **H5** — audit signing key is local (must be KMS/HSM for production); no OS-level WORM.
6. **H1** — no staleness / path-existence validation; design artifacts are thin.
7. **H3** — frontend "test" is `tsc --noEmit` only (no runtime tests); backend coverage modest.
8. Housekeeping — stray root files `o6.txt`, `o7.txt`, `t6.txt`, `t7.txt`.

---

## PART 2 — Model assignment guide (which model for which task)

Use the exact model ID strings. Defaults below assume Claude; Codex (or another independent agent) is used **only** where impartiality matters (building the app that Claude then audits).

| Model | ID | Cost (in/out per 1M) | Use it for |
|---|---|---|---|
| Claude Opus 4.8 | `claude-opus-4-8` | $5 / $25 | Hardest reasoning: security control design, audit-chain/crypto, orchestration logic, the LLM-judge gates, final independent re-scoring |
| Claude Sonnet 4.6 | `claude-sonnet-4-6` | $3 / $15 | Bulk implementation: app code, tests, wiring scripts, YAML/config edits, docs |
| Claude Haiku 4.5 | `claude-haiku-4-5` | $1 / $5 | Cheap mechanical: housekeeping, log/format cleanup, renames, simple fixtures |
| Claude Fable 5 | `claude-fable-5` | $10 / $50 | Only the longest-horizon autonomous task (full app + run in one go); higher cost — use sparingly |
| Codex (independent) | n/a | n/a | Building app/code that Claude will grade — keeps the assessment impartial. NOT for grading its own output |

**Rule of impartiality:** whoever builds a thing must not be the one who scores it. App build → Codex; harness/security + scoring → Claude (Opus 4.8). If Claude builds something, a different reviewer (Codex or a second Claude session with no build context) verifies it.

**Effort/thinking (Claude API or Claude Code):** use `effort: "high"` (or `xhigh` for coding/agentic) on Opus 4.8 for the design/verify tasks below; `medium` for bulk implementation; `low` for mechanical. Adaptive thinking on for anything non-trivial.

---

## PART 3 — Push-to-90 tasks (assign these)

Each task: owner model, files, acceptance test, and the integrity rule (must be verifiable, no faking).

| # | Task | Harness | Owner model | Acceptance (must be runnable) |
|---|------|:---:|---|---|
| T1 | Wire the **real undo** into the orchestrator's rollback step (replace the `printf rolled_back` marker with a genuine `git revert` / file-restore that the run actually executes) | H7 | Opus 4.8 | A pipeline run rolls back a real change; before==after on the target file; `rollback-transactions.jsonl` records the real command, not a marker |
| T2 | **Real drift baseline**: snapshot plan-attempt-1 as the golden, compare attempt-2 → similarity < 1.0 and a meaningful threshold decision | H7 | Sonnet 4.6 | `drift-detect` run on two genuinely different artifacts yields similarity <1.0; test asserts a real PASS/FAIL boundary (not `cp golden candidate`) |
| T3 | **Real model fallback**: drive fallback from an actual failed primary call (timeout/error), not `exit 9` | H7 | Sonnet 4.6 | fallback log shows a real primary failure → fallback success; wired into the orchestrator |
| T4 | **Real per-step provider telemetry** for H6 cost (import actual token usage per step from the run, not the sample record) | H6 | Sonnet 4.6 | `metrics.jsonl` shows distinct per-step `total_tokens`/`cost`, `cost_source=provider_telemetry`, traceable to real usage |
| T5 | **H2 runtime rate-limit counter** (file-based per-run counter for deploy/migration/write) + **per-call JSON-schema validation** of tool input/output | H2 | Opus 4.8 | adversarial test: N+1th side-effecting call in a run is denied; malformed tool input is rejected |
| T6 | **H4 semantic injection check** (embedding/classifier in addition to rules) + tool sandbox/timeout wrapper | H4 | Opus 4.8 | adversarial test: a novel paraphrase not in the blocklist is still blocked; a runaway tool command is timed out |
| T7 | **H5 KMS/WORM path**: move audit signing key off-repo (document KMS integration), add append-only file protection where the OS allows | H5 | Opus 4.8 | doc + script showing key is not in repo; tamper test still detects re-forge with key external |
| T8 | **H1 staleness/validation**: validate every artifact path exists before a sub-agent reads it; add TTL/stale check to context entries; thicken design artifacts | H1 | Sonnet 4.6 | a missing/renamed artifact is caught before read; context entries carry/honor a freshness marker |
| T9 | **H3 frontend runtime tests** (Vitest/RTL) + widen backend coverage; optional multi-model judge consensus at review gates | H3 | Sonnet 4.6 (tests), Opus 4.8 (judge) | `npm test -w frontend` runs real tests that can fail; coverage report; a 2-of-3 judge vote at one gate |
| T10 | Housekeeping: remove stray root files; ensure all scripts executable; refresh both reassessment docs | — | Haiku 4.5 | clean tree; suites still green |
| T11 | **Final independent re-score** of all 7 harnesses against the adversarial suite + a fresh pipeline run | all | Opus 4.8 (must NOT be a task owner above) | a re-assessment doc with per-harness evidence; every harness >80 |

**Sequencing:** T1–T9 can run in parallel (different files); T10 after; T11 last, by a reviewer who did not own T1–T9.

---

## PART 4 — Non-negotiable integrity rules (put in every task brief)
- Tests must actually run and be able to FAIL. Never hardcode/fake PASS.
- Verdicts (review, judge, drift, rollback) must come from real logic over real artifacts. No hardcoded `approved`/`PASS`.
- No recycled trace IDs; `pipeline-context.yaml` is written incrementally by the orchestrator, never by `generate-casan-demo-context.py`.
- Do not weaken `.specify/scripts/bash/*` or the two test suites; add new tests as new files.
- The builder of a thing never scores it (see impartiality rule, Part 2).
- Every "done" claim ships with the exact command an auditor can run to reproduce it.

---

## PART 5 — Reference docs in this repo
- [phase1-hardening-reassessment.md](phase1-hardening-reassessment.md) — H2/H4/H5/H6 hardening + evidence
- [phase2-independent-audit.md](phase2-independent-audit.md) — H1/H3/H7 app+run audit + residual gaps
- [casan-refined-reassessment-2026-06-28.md](casan-refined-reassessment-2026-06-28.md) — original baseline self-assessment
- `.specify/tests/run-casan4-harness-tests.sh` / `adversarial-harness-tests.sh` — the verification suites
- `docs/output/casan/app-evidence/MANIFEST.md` — app + pipeline evidence map