# v0.9 Release - Prepared: 2026-11-14 - Published: 2026-11-14 to `main`; tag `v0.9.0` - Publication status: **published**. The `v0.9.0` tag and its GitHub Release are the canonical publication record, and this note is not. Package version: **0.9.0** (`lakefile.toml`, `CITATION.cff`). Every count below was measured on the release tree on 2026-11-14, from the ledgers and by the gate scripts named in the text. The working history behind this release, with the dates each paragraph was written, is `docs/releases/unreleased.md` as it stood at commit `0698d27b`. ## One-paragraph truth This release widens the atlas and then spends most of its effort checking the new material against print. Five clusters are new: Sovereignty, Goodhart, Decision, LinearSystems and Composition. Wireheading grew from 8 to 20 module files, and two survey rows, BY-001 and BY-002, reached `EXACT` through the linear-systems work. The ledger grew with it. `registry.yaml` went from 91 to 144 results, and every one of the 53 new rows is an artifact row, so the 49 claim rows are the same 49. The statement-by-statement source audit went from 8 sources and 113 graded rows to 28 sources and 566 rows. Two debt counters are now computed by scripts: scope debt stands at 4 owed cells, each costed in writing, and witness debt at 31 theorems that no worked example applies, a ceiling the gate enforces. The maintainer signed 34 AI-system bridge declarations between 2026-10-03 and 2026-10-05; with the 3 reviewed before v0.8.0, all 37 are signed, 30 of them at `REVIEWED`. Against that, the root import grew from 81 to 251 imports, 107 of them `Examples/` modules, and the plan to shrink it before this release was not carried out. ## What changed since v0.8.0 **New mathematics.** Module-file counts include each cluster's facade file. - **Sovereignty** (50 module files): Pauly's coalition logic, Goranko, Jamroga and Turrini's playability, Peleg's effectivity functions, the concurrent game frame, Wooldridge and van der Hoek's action systems, and six governance bridges. Each bridge is witnessed in `Examples/Sovereignty/Governance.lean`, with the escape shown wherever one exists. - **Goodhart** (5 module files): Manheim and Garrabrant's regressional and extremal cases, Zhuang and Hadfield-Menell's overoptimization result, and Skalse et al.'s hackability definitions. - **Decision** (5 module files): a rewardless MDP carrier, a discounted value layer, and `Decision.BoundedValue`, which values a stationary policy on an arbitrary nonempty state type. The contraction argument behind it is Econlib's `Blackwell.lean`, ported under Apache-2.0 as `AISafetyAtlas.Analysis.Blackwell`. `vPiBdd_eq_vPi` proves the bounded and finite value functions agree wherever both are defined. - **LinearSystems** (8 module files): Klamka's observability and controllability, with `determinesStateOn_iff_isObservable` and `isCompletelyReachable_iff_isControllable` proving each algebraic criterion equivalent to the dynamical property it names. This moved BY-001 and BY-002 to `EXACT`. - **Wireheading** went from 8 to 20 module files: Ring and Orseau's delusion box, the self-modifying agent, value learning, a stochastic corrupt-reward MDP with `PMF` transitions and policies, and goal preservation run over whole trajectories. - **Fairness**: Kleinberg, Mullainathan and Raghavan's Theorem 1.2, the approximate trade-off, in `Fairness.ApproximateRiskAssignment` at print's own error function. The records are `RELATED` because several of ยง3's displays needed repair before a proof went through; the repairs are listed in `docs/provenance/kleinberg-fairness-tradeoffs.md`. - **Containment**: Alfonseca et al.'s containment theorem (JAIR 2021) as `Verification.Containment.harming_undecidable`, a `REVIEWED` bridge graded in section 28 of the source audit. It gives BY-025 a `RELATED` in-tree record. - **d-separation** is now in the tree. `Causal.DSep` consumes the graph layer of the external Causalean library, pinned by commit at `e298f649` after upstream replaced its history on 2026-10-06. - **`atlas-check`** grew from 5 to 14 kinds. The nine new kinds are `access`, `conformity`, `enforcement`, `fairness`, `goodhart`, `membership`, `refusal`, `shield` and `unlearning`. **Joins between notions of agent behaviour.** v0.8 named seven models with no transports between them. Six of those seven now have an adapter module that imports the `Decision.MDP` carrier: `Wireheading.CRMDP` directly, and `Wireheading.AgentEquations`, `Wireheading.GoalPreservation`, `Verification.Robot`, `Preference` and `Compositional.Hyperproperties` through `AgentHistory`, `GoalPreservationCarrier`, `RobotRun`, `Trajectory` and `TraceSystem`. `Wireheading.GoalPreservationSource` is the one without an adapter. The knowability kernel has 16 consuming modules outside itself and outside `Examples/`, up from 10 (`report_consumers.py --hub`). **Ledger and grading.** - **Source audit.** `docs/provenance/source-coverage-audit.md` grades 28 sources and 566 statements: 300 `Yes`, 17 `Partial`, 208 `No` and 41 `Beyond` (`check_coverage_audit.py`). At `v0.8.0` it held 8 sources and 113 statements. - **The fourteen unrowed modules are now accounted for.** The Sovereignty, Goodhart and Decision clusters were written on 2026-09-10 and merged in #72 (2026-10-05); on 2026-09-10 they held 14 modules and no registry row. Today Goodhart has no unrowed module out of 5, Sovereignty has 3 out of 50 and Decision has 2 out of 5. Across the whole root import, 144 library modules are imported directly, 101 host an `IN_TREE` registry row, and the other 43 are each listed in the audit's module ledger with a reason: 8 graded through a sibling module and 35 with no printed statement to grade. `check_coverage_audit.py` fails if a root-imported module has neither. - **Scope debt is computed.** Four cells graded `Narrower` or `Mixed` are owed, and each declares that it is open and costed (`check_scope_owed.py`): Everitt et al.'s Definition 17 and Theorem 18, and Turner and Tadepalli's A.12 and A.13. - **Witness debt is computed.** `check_witness_debt.py` reports 3 of 2,696 pinned theorems ungrounded. Two of the three cannot be witnessed, because their antecedent type is proved empty by `isEmpty_o24Solution`, and `docs/status/witness-vacuity.json` names that proof. 31 theorems reach no `Examples/` application, on 20 leaves. Both numbers are pinned in `agent_gate.sh`. - **Bridges.** All 37 `BRIDGE` declarations are signed: 30 `REVIEWED` and 7 `STATEMENT_REVIEWED`. The maintainer signed 31 on 2026-10-03 and 2026-10-04 and 3 on 2026-10-05; the other 3 (BY-004, BY-012, BY-033) were reviewed before v0.8.0. Before signing, each allowed claim was narrowed to what its Lean statement proves. The label is per declaration, so the row-level count of reviewed source-claim interpretations stays at 3, with 1 statement-reviewed. - **Triage.** `UNTRIAGED` rows went from 22 to 13 of the rows carrying no atlas Lean. - **Source metadata.** A machine audit of source metadata and rights, with per-finding human dispositions, and missing source locators added (#69, #70). **Process.** - **Elaboration drift.** The 29 statements whose elaborated type changed when SCM domains became types have a recorded verdict in `docs/status/elaboration-accepted.json`, pinned by digest so a later change to any of them fails again (#75). - **Kernel replay.** The nightly check replays each module once, in parallel, after every scheduled run from 2026-10-07 to 2026-10-12 timed out (#76). The drift dump is parallel too (#78). - **Public API pin.** Since #72 the pin excludes `Examples/`, which are worked witnesses and carry no compatibility promise. ## Release inventory The release tree records: - **379** Atlas Lean declarations in the registry, up from 265; - **144** results: **49** claim rows and **95** artifact rows; - **141** catalogued sources and **210** formalization records (**87** on claim rows); - **111** rows carrying atlas Lean at any grade (**24** of them claim rows); **16** claim rows reach `EXACT`/`EQUIVALENT` and **9** are `RELATED`-only; - **37** `BRIDGE` declarations: **30** `REVIEWED`, **7** `STATEMENT_REVIEWED`; - **3** open conjectures of **11** recorded, plus **6** determine-problem targets, unchanged from v0.8; - **638** tracked `.lean` files and **182,500** lines including `vendor/`, up from 412 files and 118,276 lines at `v0.8.0`; - **2,789** pinned public declarations. The v0.8 figure of 1,768 included `Examples/`, so the two are not comparable; - **11,585** entries in `docs/status/declaration-index.json`, up from 6,593. These counts record inventory and evidence. None of them is a completion target. ## Critical non-claims - **No MAIS problem is newly claimed solved.** The conjecture board is unchanged at 3 open of 11. MAIS-O70's verification remains conditional on frontier hypotheses that are stated and not proved. - **The two MAIS-O70 fidelity adjudications are still unsigned.** `docs/provenance/mais-o70-fidelity-adjudications.md` records both as drafts, and no gate script checks that file. - **Headline claim coverage moved by two rows.** BY-001 and BY-002 reached `EXACT`, and BY-025 gained a `RELATED` in-tree record. Claim rows carrying atlas Lean went from 21 to 24, and `RELATED`-only rows from 8 to 9. Every new row is an artifact row. 24 of the 49 claim rows have no formalization record in any system. - **A signed bridge licenses its stated conditional and nothing wider.** The 37 signatures are scoped to the narrowed claims in each review file. None is a claim about a deployed system. - **Sitting on one carrier is not a transport.** The six adapters project each model onto `Decision.MDP`. No theorem yet moves a result from one of the seven models to another. - **The atlas grades nothing inside Causalean.** `Causal.DSep` states Everitt et al.'s Definition 6 over Causalean's graph layer, and the audit grades that statement in its section 8. Causalean's SCM layer is not used, because its domains are measurable spaces where the atlas's are finite. - **The axiom audit does not grade a proof route**, for the reason v0.8 gave: `#print axioms` reads a declaration's type as well as its value. ## Known debts - **The root import was meant to shrink in v0.9 and did not.** `STATE.md` recorded the plan on 2026-10-05: a small root of primitives and facades, per-cluster entry points, and `Examples/` off the root. The root still has 251 imports, 107 of them `Examples/`, against 81 and 19 at `v0.8.0`. - **The module-ledger check reads only direct root imports.** 138 library modules are not imported directly by `AISafetyAtlas.lean`, and 100 of them host no `IN_TREE` registry row, 54 of those in `SingularLearning`. No script asks whether those modules need a row. - **Four scope cells are owed**, each costed. The maintainer ruled the null-context reading for Everitt's two on 2026-09-21, so what remains there is a measure-theoretic conditional expectation. The audit's own note argues against paying the cost of Turner and Tadepalli's two. - **18 leaves of witness debt need a model**, beyond the 2 that cannot be witnessed; they are listed by `check_witness_debt.py`. - **145 of 196 `Wider` and `Beyond` rows name no worked witness** (`check_scope_witnesses.py`, advisory). - **The `GoalPreservationSource` model has no adapter** to the shared carrier. ## Objective evidence Measured on this tree without a Lean build: - `scripts/agent_gate.sh`: every validator, generated-view check, prose check and path check passes, including `check_statement_freeze.py` (411 graded statements unchanged) and `check_module_graded.py` (307 library modules, 0 unreviewed). - `pytest`: 532 of 542 tests pass. The other 10 shell out to `lake env lean` and need a built library, which this checkout did not have. - No `sorry`, `admit`, `axiom` or `native_decide` token in the released Lean files (`validate_current_state.py`). The checks that need a full elaboration run in CI on the tag: `lake build`, the off-root targets in `scripts/lean_build_targets.txt`, `check_print_axioms.py`, `check_audit_coverage.py`, `axiom-audit`, `check_atlas_check.sh` and `generate_declaration_index.py --check`. ## Citation and publication boundary The published set is the repository's git tags and GitHub Releases. This note describes what the `v0.9.0` tag contains. The tag is the canonical record, and this file is not.