falsify 0.2.0__tar.gz → 0.3.0__tar.gz

{falsify-0.2.0 → falsify-0.3.0}/NOTICE

@@ -32,7 +32,7 @@ Teams deploying falsify in production as part of a commercial service
 are encouraged — but not required by the MIT License — to contact the
 author about support, SLAs, and enterprise features:
 
-    hello@studio-11.co
+    hello@falsify.dev
 
 See docs/COMMERCIAL.md for details.
 

{falsify-0.2.0/falsify.egg-info → falsify-0.3.0}/PKG-INFO

@@ -1,7 +1,7 @@
 Metadata-Version: 2.4
 Name: falsify
-Version: 0.2.0
-Summary: Pre-registration and CI for AI-agent claims — deterministic PASS or FAIL.
+Version: 0.3.0
+Summary: PRML reference CLI — pre-register an ML evaluation claim as a SHA-256 manifest; verify PASS/FAIL/TAMPERED.
 Author: Cüneyt Öztürk
 License: MIT
 Project-URL: Homepage, https://falsify.dev
@@ -34,21 +34,25 @@ Dynamic: license-file
 
 **ML evaluation claims should be locked before the experiment runs, not reported after.**
 
-`falsify` commits a claim — metric, threshold, dataset hash, seed — as a SHA-256 manifest. Run the eval. The hash either matches or it doesn't.
+PRML commits a claim — metric, threshold, dataset hash, seed — as a SHA-256 manifest. Run the eval. The hash either matches or it doesn't.
 
 ```bash
-$ falsify lock claim.yaml
-locked: sha256:a3f9...c821
+$ pip install falsify
+$ falsify lock claim.prml.yaml
+locked: claim.prml.yaml
+  sha256:          c30dba8e0f566d1beebf4f8d468e6e07c821f0c72562dfb64ddf6596796f7797
 
-$ falsify verdict claim.yaml
-PASS  accuracy 0.934 >= 0.90  (hash verified)
+$ falsify verify claim.prml.yaml --observed 0.934
+PASS  metric=accuracy  observed=0.934  >=  threshold=0.9
 
-# tampered:
-$ falsify verdict claim.yaml
-TAMPERED  sha256 mismatch — spec modified after locking  (exit 3)
+# spec edited after locking → hash no longer matches:
+$ falsify verify claim.prml.yaml --observed 0.934
+TAMPERED  (exit 3)
 ```
 
-4 reference implementations — Python, JavaScript, Go, Rust — byte-equivalent on the 12 v0.1 conformance vectors (8 v0.2 candidates ship alongside, full 20-vector parity targeted for v0.2 freeze 2026-05-22). Designed for ML eval rigor. Maps to EU AI Act Article 12 evidence as a side effect.
+No install? Verify any manifest in-browser at [registry.falsify.dev](https://registry.falsify.dev). Byte-equivalent reference CLIs also ship for JS (`npm i -g falsify-js`), Go, and Rust.
+
+4 reference implementations (Python, JavaScript, Go, Rust) byte-equivalent on all 20 conformance vectors (12 v0.1 stable + 8 v0.2). PRML v0.2 frozen 2026-05-22. The same day, Lock #2 (a public hypothesis on the spec's own distribution, target ≥3 external contributors in 14 days) resolved at 0/3. The mechanism worked, the post-mortem is at [falsify.dev/notes/lock-2-postmortem](https://falsify.dev/notes/lock-2-postmortem/). Designed for ML eval rigor. Maps to EU AI Act Article 12 evidence as a side effect.
 
 > **Pre-registration + CI for AI-agent claims.** Lock the claim and threshold with SHA-256 *before* running the experiment — or the result doesn't count.
 
@@ -81,10 +85,6 @@ TAMPERED  sha256 mismatch — spec modified after locking  (exit 3)
 
 ---
 
-> **Latest — 2026-05-14** · v0.1 published on Zenodo: citable DOI [10.5281/zenodo.20177839](https://doi.org/10.5281/zenodo.20177839). PRML JSON Schema [merged into SchemaStore](https://github.com/SchemaStore/schemastore/pull/5673) (2026-05-11) by Mads Kristensen (Microsoft) — `.prml.yaml` files now autocomplete in VS Code, JetBrains, Helix, Zed, and Cursor. [OECD.AI Catalogue submission](https://oecd.ai/en/catalogue/tools) filed, vetting in progress. NIST AI 800-2 late comment archived. JTC 21 routed via Dr. Sebastian Hallensleben. `registry.falsify.dev` live with README badges at `registry.falsify.dev/badge/<hash>.svg`. **v0.1.4 released** ([release notes](https://github.com/studio-11-co/falsify/releases/tag/v0.1.4) · `pip install falsify==0.1.4`). PRML v0.1 specification published with **four reference implementations** (Python · [JavaScript](impl/js/) · [Go](impl/go/) · [Rust](impl/rust/)) all reproducing the [12 v0.1 vectors](spec/test-vectors/v0.1/) and [8 v0.2 candidate vectors](spec/test-vectors/v0.2/) byte-for-byte (20 vectors total). [14-page arXiv preprint](spec/paper/) and [v0.2 RFC](https://spec.falsify.dev/v0.2-rfc) (freeze 2026-05-22) open for public review.
-
----
-
 ## The problem
 
 Your team claims the model hits **94% accuracy**. You ship it. Three weeks later a customer proves the real number is **71%**.
@@ -97,9 +97,9 @@ PRML does not prove an ML result is true. It proves that a specific evaluation c
 
 **Falsify fixes this with a single idea from science:** you must pre-register the claim *before* you run the experiment. If you change the spec after seeing the data, the hash changes, the audit trail breaks, and CI fails with exit code 3.
 
-    $ falsify lock accuracy_claim        # SHA-256 the spec
-    $ falsify run  accuracy_claim        # reproducible experiment
-    $ falsify verdict accuracy_claim     # exit 0 = PASS, 10 = FAIL, 3 = tampered
+    $ falsify-engine lock accuracy_claim        # SHA-256 the spec
+    $ falsify-engine run  accuracy_claim        # reproducible experiment
+    $ falsify-engine verdict accuracy_claim     # exit 0 = PASS, 10 = FAIL, 3 = tampered
 
 Deterministic exit codes are the API. CI gates on them. Humans read the audit trail. The claim either survives contact with the data or it doesn't.
 
@@ -128,14 +128,14 @@ See [docs/CASE_STUDIES.md](docs/CASE_STUDIES.md) for three concrete adoption sto
 
 ---
 
-**Current version:** 0.1.0 — run `python3 falsify.py --version`.
+**Current version:** falsify 0.3.0 (PRML CLI) · falsify-engine 0.2.0 — `falsify --version`.
 **Working with Claude Code?** See [CLAUDE.md](CLAUDE.md).
 
 ---
 
 ## Specification artifacts
 
-Falsify is the reference implementation of **PRML v0.1** — Pre-Registered ML Manifest Specification. The spec, conformance suite, and adjacent documents live under `spec/`:
+This repository is the home of **PRML v0.1** — Pre-Registered ML Manifest Specification. The spec, conformance suite, reference implementations (`impl/`, JS/Go/Rust + a Python reference target), and adjacent documents live under `spec/`:
 
 - **[`spec/PRML-v0.1.md`](spec/PRML-v0.1.md)** — the spec (RFC-style, CC BY 4.0)
 - **[`spec/test-vectors/v0.1/`](spec/test-vectors/v0.1/)** — 12 conformance vectors with locked SHA-256 digests
@@ -153,6 +153,15 @@ Falsify is the reference implementation of **PRML v0.1** — Pre-Registered ML M
 - **[NIST AI RMF 1.0 crosswalk](https://spec.falsify.dev/nist-ai-rmf/)** — GOVERN / MAP / MEASURE / MANAGE subcategory map (incl. AI 600-1 GenAI Profile)
 - **[ISO/IEC 42001:2023 crosswalk](https://spec.falsify.dev/iso-42001/)** — AIMS clause-by-clause evidence map (Clauses 7-9 + Annex A controls)
 
+**Long-form working notes** (2026-05-23, written for compliance leads, AI governance officers, and notified body assessors preparing for the 2 August 2026 deadline; CC BY 4.0):
+
+- **[EU AI Act readiness assessment](https://falsify.dev/eu-ai-act-readiness/)** — six binding articles, ten-question gap check, evidence shape per obligation
+- **[2 August 2026 deadline](https://falsify.dev/ai-act-deadline-august-2026/)** — three application dates, Article 99 penalty structure, ten-week plan
+- **[Article 12 logging checklist](https://falsify.dev/article-12-checklist/)** — ten closeable questions, six event categories, printable single-page summary
+- **[Notified body evidence](https://falsify.dev/notified-body-evidence/)** — Annex VI vs Annex VII conformity assessment, six artefact families
+- **[ISO/IEC 42001 readiness](https://falsify.dev/iso-42001-readiness/)** — seven clauses, EU AI Act Article 17 overlap, twelve-month certification path
+- **[Lock #2 post-mortem](https://falsify.dev/notes/lock-2-postmortem/)** — field report on running a falsifiable spec in public
+
 **Reference implementations** (four languages, 12 v0.1 + 8 v0.2 candidate vectors = 20 total; multi-lang CI runs all 20 byte-for-byte per push and daily at 04:00 UTC):
 
 - **Python:** [`falsify.py`](falsify.py) — original reference, uses PyYAML
@@ -160,7 +169,7 @@ Falsify is the reference implementation of **PRML v0.1** — Pre-Registered ML M
 - **Go:** [`impl/go/`](impl/go/) — third reference, ~450 LOC, hand-rolled, stdlib only
 - **Rust:** [`impl/rust/`](impl/rust/) — fourth reference, ~600 LOC, hand-rolled, two deps (`serde_json`, `sha2`)
 
-Hosted spec at [spec.falsify.dev/v0.1](https://spec.falsify.dev/v0.1). Public review thread at [GitHub Discussion #6](https://github.com/studio-11-co/falsify/discussions/6). Comments via `hello@studio-11.co`.
+Hosted spec at [spec.falsify.dev/v0.1](https://spec.falsify.dev/v0.1). Public review thread at [GitHub Discussion #6](https://github.com/studio-11-co/falsify/discussions/6). Comments via `hello@falsify.dev`.
 
 **Companion projects** (separate repos under `studio-11-co`, each MIT or CC0 licensed):
 
@@ -216,6 +225,8 @@ is at <https://falsify.dev>, and the project page is at
 
 Requires Python **3.11+**.
 
+> **Two commands, one install.** `falsify` is the **PRML manifest CLI** — `lock` / `verify` / `hash` / `init` / `test-vectors` on a `*.prml.yaml` manifest (shown at the top). `falsify-engine` is the separate **pre-registration workflow engine** — the `init` → `lock` → `run` → `verdict` / `guard` loop over `.falsify/<name>/` specs. The workflow sections further down use `falsify-engine`; substitute it for `falsify` there. No install needed to verify a manifest: paste it at [registry.falsify.dev](https://registry.falsify.dev).
+
 ### Development install (from the repo)
 
 ```bash
@@ -257,16 +268,25 @@ exported hooks and how this repo eats its own dog food.
 ## Quickstart
 
 ```bash
-./demo.sh   # auto-narrated: PASS → tamper → FAIL → guard block
+# The falsify PRML CLI — lock a manifest, run your eval, verify.
+falsify init accuracy.prml.yaml          # writes a skeleton manifest
+# edit accuracy.prml.yaml: metric, comparator, threshold, dataset.hash, seed, producer
+falsify lock accuracy.prml.yaml          # canonicalize + SHA-256 + write sidecar
+# ... run your eval, get the observed value ...
+falsify verify accuracy.prml.yaml --observed 0.934
+# PASS (exit 0) · FAIL below threshold (exit 10) · TAMPERED if the spec changed (exit 3)
+```
+
+The pre-registration **workflow engine** (claim/falsification specs, `init` → `lock` → `run` → `verdict` → `guard` over `.falsify/<name>/`) ships in the same install as the `falsify-engine` command:
 
-# Either form works — `falsify` is the installed entry point,
-# `python3 falsify.py` is the uninstalled fallback.
-falsify init my_claim
+```bash
+./demo.sh                                # auto-narrated engine demo (PASS → tamper → FAIL → guard)
+falsify-engine init my_claim
 # edit .falsify/my_claim/spec.yaml to fill in the template
-falsify lock my_claim
-falsify run my_claim
-falsify verdict my_claim
-falsify hook install      # enable the commit-msg guard
+falsify-engine lock my_claim
+falsify-engine run my_claim
+falsify-engine verdict my_claim
+falsify-engine hook install      # enable the commit-msg guard
 ```
 
 Exit code `0` on PASS, `10` on FAIL. Everything else is documented
@@ -279,8 +299,8 @@ New to pre-registration? Walk through [TUTORIAL.md](TUTORIAL.md) — 15 minutes,
 ```bash
 falsify init --template accuracy
 falsify lock accuracy
-falsify run accuracy
-falsify verdict accuracy
+falsify-engine run accuracy
+falsify-engine verdict accuracy
 ```
 
 Five templates ship with a runnable spec + metric + dataset:
@@ -302,7 +322,7 @@ or the directory with `--dir`.
 make install   # pip install pyyaml
 make test      # run unittest suite
 make smoke     # run tests/smoke_test.sh
-make demo      # JUJU end-to-end (lock → run → verdict)
+make demo      # calibration end-to-end (lock → run → verdict)
 ```
 
 See [Makefile](Makefile) for all targets (`make help`).
@@ -314,18 +334,18 @@ Feature matrix vs adjacent tools: [docs/COMPARISON.md](docs/COMPARISON.md).
 
 ### Explain any claim
 
-`falsify why <name>` is the human-friendly companion to `verdict`
+`falsify-engine why <name>` is the human-friendly companion to `verdict`
 — it always exits `0` and tells you exactly what the next honest
 move is:
 
 ```
-claim: juju
+claim: calibration
 state: STALE
 reasoning: the spec has been edited (sha256:1038219d75a8) but no run
   exists against this hash. Last run was against sha256:164f619d4860.
 locked: yes (sha256:164f619d4860, 2h ago)
 last run: 2026-04-22T02:10:17+00:00 (2h ago)
-next action: `falsify run <name>` to produce a fresh verdict against
+next action: `falsify-engine run <name>` to produce a fresh verdict against
   the current spec.
 ```
 
@@ -334,13 +354,13 @@ and the last five runs.
 
 ### Spot drift with a sparkline
 
-`falsify trend <name>` draws an ASCII sparkline of the metric
+`falsify-engine trend <name>` draws an ASCII sparkline of the metric
 across its recorded runs, marks the threshold line, and classifies
 the trajectory as **improving**, **degrading**, **flat**, or
 **mixed**.
 
 ```
-claim: juju
+claim: calibration
 threshold: 0.25 (direction: below)
 runs: 20 shown (of 20)
 
@@ -362,14 +382,14 @@ trend: degrading
 
 ### Measure the CLI itself
 
-`falsify bench` spawns each subcommand under a fresh temporary
+`falsify-engine bench` spawns each subcommand under a fresh temporary
 directory and records per-command latency (min / median / p95 /
 max / mean / stddev). Useful as a sanity check before a release
 or when investigating a suspected startup-time regression.
 
 ```bash
-falsify bench --runs 5 --commands "--help,list,stats,score"
-falsify bench --runs 5 --json     # machine-readable output
+falsify-engine bench --runs 5 --commands "--help,list,stats,score"
+falsify-engine bench --runs 5 --json     # machine-readable output
 ```
 
 `--runs <N>` sets the timed-iteration count (default 5, capped at
@@ -427,7 +447,7 @@ compose the skills and CLI.
 
 **CI** (`.github/workflows/falsify.yml`) — on every push and PR,
 the workflow runs the unittest suite, `tests/smoke_test.sh`, the
-JUJU end-to-end (`lock` → `run` → `verdict`), a guard self-check,
+calibration end-to-end (`lock` → `run` → `verdict`), a guard self-check,
 and a skill-lint pass over every SKILL.md and agent file.
 
 ## Demo
@@ -495,7 +515,7 @@ ln -sf "$(pwd)/hooks/commit-msg" .git/hooks/commit-msg
 - `hypothesis.schema.yaml` — spec schema (claim, falsification,
   experiment, environment, artifacts).
 - `examples/hello_claim/` — tiny smoke-test fixture.
-- `examples/juju_sample/` — anonymized 20-row prediction ledger
+- `examples/calibration_sample/` — anonymized 20-row prediction ledger
   for the Brier score demo.
 - `hooks/commit-msg` — the guard hook.
 - `tests/` — `unittest` suite plus `smoke_test.sh` end-to-end driver.
@@ -519,14 +539,16 @@ Run `make dogfood` to re-verify. CI runs these on every PR.
 
 See [CHANGELOG.md](CHANGELOG.md) for release history.
 
+> **Latest — 2026-05-23** · **PRML v0.2 frozen** with all 20 conformance vectors (12 v0.1 stable + 8 v0.2) passing byte-for-byte across the four reference implementations. Lock #2 (public hypothesis on spec's own distribution) resolved at **0/3 external contributors**, mechanism worked, [post-mortem published](https://falsify.dev/notes/lock-2-postmortem/). **`mlflow-falsify` v0.2.0** shipped with `MLFLOW_FALSIFY_TAG_SCOPE=experiment` for HPO sweeps; [MLflow community plugin showcase PR](https://github.com/mlflow/mlflow/pull/23569) is live and under review. Five long-form working notes published for EU AI Act readiness: [readiness assessment](https://falsify.dev/eu-ai-act-readiness/), [2 August 2026 deadline](https://falsify.dev/ai-act-deadline-august-2026/), [Article 12 ten-item checklist](https://falsify.dev/article-12-checklist/), [notified body evidence](https://falsify.dev/notified-body-evidence/), [ISO/IEC 42001 readiness](https://falsify.dev/iso-42001-readiness/). DOI [10.5281/zenodo.20177839](https://doi.org/10.5281/zenodo.20177839). PRML JSON Schema in [SchemaStore](https://github.com/SchemaStore/schemastore/pull/5673) (Mads Kristensen / Microsoft) — `.prml.yaml` files autocomplete in VS Code, JetBrains, Helix, Zed, and Cursor. `registry.falsify.dev` live with README badges at `registry.falsify.dev/badge/<hash>.svg`.
+
 ## Roadmap
 
 Two roadmaps run alongside each other:
 
-- **CLI tool roadmap:** [ROADMAP.md](ROADMAP.md) — `falsify` features, integrations, dependencies. CLI v0.2 targeted 2026-06-15.
-- **Specification roadmap:** [spec/v0.2/ROADMAP.md](spec/v0.2/ROADMAP.md) — PRML format evolution, canonicalization grammar, conformance. Spec v0.2 freeze 2026-05-22.
+- **CLI tool roadmap:** [ROADMAP.md](ROADMAP.md) — `falsify` features, integrations, dependencies. **CLI v0.2.0 shipped 2026-05-22.** v0.3 features tracked alongside the v0.3 spec backlog.
+- **Specification roadmap:** [spec/v0.2/ROADMAP.md](spec/v0.2/ROADMAP.md) — PRML format evolution, canonicalization grammar, conformance. **Spec v0.2 frozen 2026-05-22.** v0.3 design backlog open under [`spec/v0.3-backlog/`](spec/v0.3-backlog/) (claim trees, suite manifests, selective-disclosure resistance via `leaves_total`).
 
-The CLI is downstream of the spec: when spec v0.2 freezes, CLI v0.2 follows about three weeks later. CLI v0.3 is loosely scoped for Q4 2026.
+The CLI is downstream of the spec: spec v0.2 frozen 2026-05-22, CLI v0.2.0 shipped to PyPI the same week. CLI v0.3 is loosely scoped for Q4 2026, tracking the v0.3 spec backlog.
 
 ## Trust model
 

{falsify-0.2.0 → falsify-0.3.0}/README.md

@@ -2,21 +2,25 @@
 
 **ML evaluation claims should be locked before the experiment runs, not reported after.**
 
-`falsify` commits a claim — metric, threshold, dataset hash, seed — as a SHA-256 manifest. Run the eval. The hash either matches or it doesn't.
+PRML commits a claim — metric, threshold, dataset hash, seed — as a SHA-256 manifest. Run the eval. The hash either matches or it doesn't.
 
 ```bash
-$ falsify lock claim.yaml
-locked: sha256:a3f9...c821
+$ pip install falsify
+$ falsify lock claim.prml.yaml
+locked: claim.prml.yaml
+  sha256:          c30dba8e0f566d1beebf4f8d468e6e07c821f0c72562dfb64ddf6596796f7797
 
-$ falsify verdict claim.yaml
-PASS  accuracy 0.934 >= 0.90  (hash verified)
+$ falsify verify claim.prml.yaml --observed 0.934
+PASS  metric=accuracy  observed=0.934  >=  threshold=0.9
 
-# tampered:
-$ falsify verdict claim.yaml
-TAMPERED  sha256 mismatch — spec modified after locking  (exit 3)
+# spec edited after locking → hash no longer matches:
+$ falsify verify claim.prml.yaml --observed 0.934
+TAMPERED  (exit 3)
 ```
 
-4 reference implementations — Python, JavaScript, Go, Rust — byte-equivalent on the 12 v0.1 conformance vectors (8 v0.2 candidates ship alongside, full 20-vector parity targeted for v0.2 freeze 2026-05-22). Designed for ML eval rigor. Maps to EU AI Act Article 12 evidence as a side effect.
+No install? Verify any manifest in-browser at [registry.falsify.dev](https://registry.falsify.dev). Byte-equivalent reference CLIs also ship for JS (`npm i -g falsify-js`), Go, and Rust.
+
+4 reference implementations (Python, JavaScript, Go, Rust) byte-equivalent on all 20 conformance vectors (12 v0.1 stable + 8 v0.2). PRML v0.2 frozen 2026-05-22. The same day, Lock #2 (a public hypothesis on the spec's own distribution, target ≥3 external contributors in 14 days) resolved at 0/3. The mechanism worked, the post-mortem is at [falsify.dev/notes/lock-2-postmortem](https://falsify.dev/notes/lock-2-postmortem/). Designed for ML eval rigor. Maps to EU AI Act Article 12 evidence as a side effect.
 
 > **Pre-registration + CI for AI-agent claims.** Lock the claim and threshold with SHA-256 *before* running the experiment — or the result doesn't count.
 
@@ -49,10 +53,6 @@ TAMPERED  sha256 mismatch — spec modified after locking  (exit 3)
 
 ---
 
-> **Latest — 2026-05-14** · v0.1 published on Zenodo: citable DOI [10.5281/zenodo.20177839](https://doi.org/10.5281/zenodo.20177839). PRML JSON Schema [merged into SchemaStore](https://github.com/SchemaStore/schemastore/pull/5673) (2026-05-11) by Mads Kristensen (Microsoft) — `.prml.yaml` files now autocomplete in VS Code, JetBrains, Helix, Zed, and Cursor. [OECD.AI Catalogue submission](https://oecd.ai/en/catalogue/tools) filed, vetting in progress. NIST AI 800-2 late comment archived. JTC 21 routed via Dr. Sebastian Hallensleben. `registry.falsify.dev` live with README badges at `registry.falsify.dev/badge/<hash>.svg`. **v0.1.4 released** ([release notes](https://github.com/studio-11-co/falsify/releases/tag/v0.1.4) · `pip install falsify==0.1.4`). PRML v0.1 specification published with **four reference implementations** (Python · [JavaScript](impl/js/) · [Go](impl/go/) · [Rust](impl/rust/)) all reproducing the [12 v0.1 vectors](spec/test-vectors/v0.1/) and [8 v0.2 candidate vectors](spec/test-vectors/v0.2/) byte-for-byte (20 vectors total). [14-page arXiv preprint](spec/paper/) and [v0.2 RFC](https://spec.falsify.dev/v0.2-rfc) (freeze 2026-05-22) open for public review.
-
----
-
 ## The problem
 
 Your team claims the model hits **94% accuracy**. You ship it. Three weeks later a customer proves the real number is **71%**.
@@ -65,9 +65,9 @@ PRML does not prove an ML result is true. It proves that a specific evaluation c
 
 **Falsify fixes this with a single idea from science:** you must pre-register the claim *before* you run the experiment. If you change the spec after seeing the data, the hash changes, the audit trail breaks, and CI fails with exit code 3.
 
-    $ falsify lock accuracy_claim        # SHA-256 the spec
-    $ falsify run  accuracy_claim        # reproducible experiment
-    $ falsify verdict accuracy_claim     # exit 0 = PASS, 10 = FAIL, 3 = tampered
+    $ falsify-engine lock accuracy_claim        # SHA-256 the spec
+    $ falsify-engine run  accuracy_claim        # reproducible experiment
+    $ falsify-engine verdict accuracy_claim     # exit 0 = PASS, 10 = FAIL, 3 = tampered
 
 Deterministic exit codes are the API. CI gates on them. Humans read the audit trail. The claim either survives contact with the data or it doesn't.
 
@@ -96,14 +96,14 @@ See [docs/CASE_STUDIES.md](docs/CASE_STUDIES.md) for three concrete adoption sto
 
 ---
 
-**Current version:** 0.1.0 — run `python3 falsify.py --version`.
+**Current version:** falsify 0.3.0 (PRML CLI) · falsify-engine 0.2.0 — `falsify --version`.
 **Working with Claude Code?** See [CLAUDE.md](CLAUDE.md).
 
 ---
 
 ## Specification artifacts
 
-Falsify is the reference implementation of **PRML v0.1** — Pre-Registered ML Manifest Specification. The spec, conformance suite, and adjacent documents live under `spec/`:
+This repository is the home of **PRML v0.1** — Pre-Registered ML Manifest Specification. The spec, conformance suite, reference implementations (`impl/`, JS/Go/Rust + a Python reference target), and adjacent documents live under `spec/`:
 
 - **[`spec/PRML-v0.1.md`](spec/PRML-v0.1.md)** — the spec (RFC-style, CC BY 4.0)
 - **[`spec/test-vectors/v0.1/`](spec/test-vectors/v0.1/)** — 12 conformance vectors with locked SHA-256 digests
@@ -121,6 +121,15 @@ Falsify is the reference implementation of **PRML v0.1** — Pre-Registered ML M
 - **[NIST AI RMF 1.0 crosswalk](https://spec.falsify.dev/nist-ai-rmf/)** — GOVERN / MAP / MEASURE / MANAGE subcategory map (incl. AI 600-1 GenAI Profile)
 - **[ISO/IEC 42001:2023 crosswalk](https://spec.falsify.dev/iso-42001/)** — AIMS clause-by-clause evidence map (Clauses 7-9 + Annex A controls)
 
+**Long-form working notes** (2026-05-23, written for compliance leads, AI governance officers, and notified body assessors preparing for the 2 August 2026 deadline; CC BY 4.0):
+
+- **[EU AI Act readiness assessment](https://falsify.dev/eu-ai-act-readiness/)** — six binding articles, ten-question gap check, evidence shape per obligation
+- **[2 August 2026 deadline](https://falsify.dev/ai-act-deadline-august-2026/)** — three application dates, Article 99 penalty structure, ten-week plan
+- **[Article 12 logging checklist](https://falsify.dev/article-12-checklist/)** — ten closeable questions, six event categories, printable single-page summary
+- **[Notified body evidence](https://falsify.dev/notified-body-evidence/)** — Annex VI vs Annex VII conformity assessment, six artefact families
+- **[ISO/IEC 42001 readiness](https://falsify.dev/iso-42001-readiness/)** — seven clauses, EU AI Act Article 17 overlap, twelve-month certification path
+- **[Lock #2 post-mortem](https://falsify.dev/notes/lock-2-postmortem/)** — field report on running a falsifiable spec in public
+
 **Reference implementations** (four languages, 12 v0.1 + 8 v0.2 candidate vectors = 20 total; multi-lang CI runs all 20 byte-for-byte per push and daily at 04:00 UTC):
 
 - **Python:** [`falsify.py`](falsify.py) — original reference, uses PyYAML
@@ -128,7 +137,7 @@ Falsify is the reference implementation of **PRML v0.1** — Pre-Registered ML M
 - **Go:** [`impl/go/`](impl/go/) — third reference, ~450 LOC, hand-rolled, stdlib only
 - **Rust:** [`impl/rust/`](impl/rust/) — fourth reference, ~600 LOC, hand-rolled, two deps (`serde_json`, `sha2`)
 
-Hosted spec at [spec.falsify.dev/v0.1](https://spec.falsify.dev/v0.1). Public review thread at [GitHub Discussion #6](https://github.com/studio-11-co/falsify/discussions/6). Comments via `hello@studio-11.co`.
+Hosted spec at [spec.falsify.dev/v0.1](https://spec.falsify.dev/v0.1). Public review thread at [GitHub Discussion #6](https://github.com/studio-11-co/falsify/discussions/6). Comments via `hello@falsify.dev`.
 
 **Companion projects** (separate repos under `studio-11-co`, each MIT or CC0 licensed):
 
@@ -184,6 +193,8 @@ is at <https://falsify.dev>, and the project page is at
 
 Requires Python **3.11+**.
 
+> **Two commands, one install.** `falsify` is the **PRML manifest CLI** — `lock` / `verify` / `hash` / `init` / `test-vectors` on a `*.prml.yaml` manifest (shown at the top). `falsify-engine` is the separate **pre-registration workflow engine** — the `init` → `lock` → `run` → `verdict` / `guard` loop over `.falsify/<name>/` specs. The workflow sections further down use `falsify-engine`; substitute it for `falsify` there. No install needed to verify a manifest: paste it at [registry.falsify.dev](https://registry.falsify.dev).
+
 ### Development install (from the repo)
 
 ```bash
@@ -225,16 +236,25 @@ exported hooks and how this repo eats its own dog food.
 ## Quickstart
 
 ```bash
-./demo.sh   # auto-narrated: PASS → tamper → FAIL → guard block
+# The falsify PRML CLI — lock a manifest, run your eval, verify.
+falsify init accuracy.prml.yaml          # writes a skeleton manifest
+# edit accuracy.prml.yaml: metric, comparator, threshold, dataset.hash, seed, producer
+falsify lock accuracy.prml.yaml          # canonicalize + SHA-256 + write sidecar
+# ... run your eval, get the observed value ...
+falsify verify accuracy.prml.yaml --observed 0.934
+# PASS (exit 0) · FAIL below threshold (exit 10) · TAMPERED if the spec changed (exit 3)
+```
+
+The pre-registration **workflow engine** (claim/falsification specs, `init` → `lock` → `run` → `verdict` → `guard` over `.falsify/<name>/`) ships in the same install as the `falsify-engine` command:
 
-# Either form works — `falsify` is the installed entry point,
-# `python3 falsify.py` is the uninstalled fallback.
-falsify init my_claim
+```bash
+./demo.sh                                # auto-narrated engine demo (PASS → tamper → FAIL → guard)
+falsify-engine init my_claim
 # edit .falsify/my_claim/spec.yaml to fill in the template
-falsify lock my_claim
-falsify run my_claim
-falsify verdict my_claim
-falsify hook install      # enable the commit-msg guard
+falsify-engine lock my_claim
+falsify-engine run my_claim
+falsify-engine verdict my_claim
+falsify-engine hook install      # enable the commit-msg guard
 ```
 
 Exit code `0` on PASS, `10` on FAIL. Everything else is documented
@@ -247,8 +267,8 @@ New to pre-registration? Walk through [TUTORIAL.md](TUTORIAL.md) — 15 minutes,
 ```bash
 falsify init --template accuracy
 falsify lock accuracy
-falsify run accuracy
-falsify verdict accuracy
+falsify-engine run accuracy
+falsify-engine verdict accuracy
 ```
 
 Five templates ship with a runnable spec + metric + dataset:
@@ -270,7 +290,7 @@ or the directory with `--dir`.
 make install   # pip install pyyaml
 make test      # run unittest suite
 make smoke     # run tests/smoke_test.sh
-make demo      # JUJU end-to-end (lock → run → verdict)
+make demo      # calibration end-to-end (lock → run → verdict)
 ```
 
 See [Makefile](Makefile) for all targets (`make help`).
@@ -282,18 +302,18 @@ Feature matrix vs adjacent tools: [docs/COMPARISON.md](docs/COMPARISON.md).
 
 ### Explain any claim
 
-`falsify why <name>` is the human-friendly companion to `verdict`
+`falsify-engine why <name>` is the human-friendly companion to `verdict`
 — it always exits `0` and tells you exactly what the next honest
 move is:
 
 ```
-claim: juju
+claim: calibration
 state: STALE
 reasoning: the spec has been edited (sha256:1038219d75a8) but no run
   exists against this hash. Last run was against sha256:164f619d4860.
 locked: yes (sha256:164f619d4860, 2h ago)
 last run: 2026-04-22T02:10:17+00:00 (2h ago)
-next action: `falsify run <name>` to produce a fresh verdict against
+next action: `falsify-engine run <name>` to produce a fresh verdict against
   the current spec.
 ```
 
@@ -302,13 +322,13 @@ and the last five runs.
 
 ### Spot drift with a sparkline
 
-`falsify trend <name>` draws an ASCII sparkline of the metric
+`falsify-engine trend <name>` draws an ASCII sparkline of the metric
 across its recorded runs, marks the threshold line, and classifies
 the trajectory as **improving**, **degrading**, **flat**, or
 **mixed**.
 
 ```
-claim: juju
+claim: calibration
 threshold: 0.25 (direction: below)
 runs: 20 shown (of 20)
 
@@ -330,14 +350,14 @@ trend: degrading
 
 ### Measure the CLI itself
 
-`falsify bench` spawns each subcommand under a fresh temporary
+`falsify-engine bench` spawns each subcommand under a fresh temporary
 directory and records per-command latency (min / median / p95 /
 max / mean / stddev). Useful as a sanity check before a release
 or when investigating a suspected startup-time regression.
 
 ```bash
-falsify bench --runs 5 --commands "--help,list,stats,score"
-falsify bench --runs 5 --json     # machine-readable output
+falsify-engine bench --runs 5 --commands "--help,list,stats,score"
+falsify-engine bench --runs 5 --json     # machine-readable output
 ```
 
 `--runs <N>` sets the timed-iteration count (default 5, capped at
@@ -395,7 +415,7 @@ compose the skills and CLI.
 
 **CI** (`.github/workflows/falsify.yml`) — on every push and PR,
 the workflow runs the unittest suite, `tests/smoke_test.sh`, the
-JUJU end-to-end (`lock` → `run` → `verdict`), a guard self-check,
+calibration end-to-end (`lock` → `run` → `verdict`), a guard self-check,
 and a skill-lint pass over every SKILL.md and agent file.
 
 ## Demo
@@ -463,7 +483,7 @@ ln -sf "$(pwd)/hooks/commit-msg" .git/hooks/commit-msg
 - `hypothesis.schema.yaml` — spec schema (claim, falsification,
   experiment, environment, artifacts).
 - `examples/hello_claim/` — tiny smoke-test fixture.
-- `examples/juju_sample/` — anonymized 20-row prediction ledger
+- `examples/calibration_sample/` — anonymized 20-row prediction ledger
   for the Brier score demo.
 - `hooks/commit-msg` — the guard hook.
 - `tests/` — `unittest` suite plus `smoke_test.sh` end-to-end driver.
@@ -487,14 +507,16 @@ Run `make dogfood` to re-verify. CI runs these on every PR.
 
 See [CHANGELOG.md](CHANGELOG.md) for release history.
 
+> **Latest — 2026-05-23** · **PRML v0.2 frozen** with all 20 conformance vectors (12 v0.1 stable + 8 v0.2) passing byte-for-byte across the four reference implementations. Lock #2 (public hypothesis on spec's own distribution) resolved at **0/3 external contributors**, mechanism worked, [post-mortem published](https://falsify.dev/notes/lock-2-postmortem/). **`mlflow-falsify` v0.2.0** shipped with `MLFLOW_FALSIFY_TAG_SCOPE=experiment` for HPO sweeps; [MLflow community plugin showcase PR](https://github.com/mlflow/mlflow/pull/23569) is live and under review. Five long-form working notes published for EU AI Act readiness: [readiness assessment](https://falsify.dev/eu-ai-act-readiness/), [2 August 2026 deadline](https://falsify.dev/ai-act-deadline-august-2026/), [Article 12 ten-item checklist](https://falsify.dev/article-12-checklist/), [notified body evidence](https://falsify.dev/notified-body-evidence/), [ISO/IEC 42001 readiness](https://falsify.dev/iso-42001-readiness/). DOI [10.5281/zenodo.20177839](https://doi.org/10.5281/zenodo.20177839). PRML JSON Schema in [SchemaStore](https://github.com/SchemaStore/schemastore/pull/5673) (Mads Kristensen / Microsoft) — `.prml.yaml` files autocomplete in VS Code, JetBrains, Helix, Zed, and Cursor. `registry.falsify.dev` live with README badges at `registry.falsify.dev/badge/<hash>.svg`.
+
 ## Roadmap
 
 Two roadmaps run alongside each other:
 
-- **CLI tool roadmap:** [ROADMAP.md](ROADMAP.md) — `falsify` features, integrations, dependencies. CLI v0.2 targeted 2026-06-15.
-- **Specification roadmap:** [spec/v0.2/ROADMAP.md](spec/v0.2/ROADMAP.md) — PRML format evolution, canonicalization grammar, conformance. Spec v0.2 freeze 2026-05-22.
+- **CLI tool roadmap:** [ROADMAP.md](ROADMAP.md) — `falsify` features, integrations, dependencies. **CLI v0.2.0 shipped 2026-05-22.** v0.3 features tracked alongside the v0.3 spec backlog.
+- **Specification roadmap:** [spec/v0.2/ROADMAP.md](spec/v0.2/ROADMAP.md) — PRML format evolution, canonicalization grammar, conformance. **Spec v0.2 frozen 2026-05-22.** v0.3 design backlog open under [`spec/v0.3-backlog/`](spec/v0.3-backlog/) (claim trees, suite manifests, selective-disclosure resistance via `leaves_total`).
 
-The CLI is downstream of the spec: when spec v0.2 freezes, CLI v0.2 follows about three weeks later. CLI v0.3 is loosely scoped for Q4 2026.
+The CLI is downstream of the spec: spec v0.2 frozen 2026-05-22, CLI v0.2.0 shipped to PyPI the same week. CLI v0.3 is loosely scoped for Q4 2026, tracking the v0.3 spec backlog.
 
 ## Trust model