Restore exact user-visible invariance and start Cycle 8 carrier research

.github/workflows/ci.yml

@@ -55,10 +55,10 @@ jobs:
             export FUCKMARK_CLIPBOARD_CAPTURE="$capture"
             output="$(printf 'I do not agree and I cannot stay.\n' | "$command" --stdin --copy)"
             printf '%s\n' "$output"
-            test "$output" = "I don't agree and I can't stay."
-            test "$(cat "$capture")" = "I don't agree and I can't stay."
+            test "$output" = "I do not agree and I cannot stay."
+            test "$(cat "$capture")" = "I do not agree and I cannot stay."
           done
-          test "$(printf 'I do not agree.\n' | FuckMark)" = "I don't agree."
+          test "$(printf 'I do not agree.\n' | FuckMark)" = "I do not agree."
 
   calibration-readiness:
     name: Frozen calibration readiness

.github/workflows/cycle8-carrier-detector.yml

@@ -0,0 +1,56 @@
+name: Cycle8 Carrier Detector
+
+on:
+  workflow_dispatch:
+    inputs:
+      device:
+        description: Torch device
+        required: true
+        default: cpu
+        type: string
+      max_attempts:
+        description: Deterministic retry budget for ASCII 64-token unique pairs
+        required: true
+        default: "64"
+        type: string
+  pull_request:
+    paths:
+      - ".github/workflows/cycle8-carrier-detector.yml"
+      - "fuckmark/cycle8/**"
+      - "fuckmark/cycle8_hf.py"
+      - "fuckmark/product/**"
+      - "tools/cycle8_fixture_compare.py"
+      - "tests/test_cycle8_fixture_compare.py"
+      - "requirements-smoke.txt"
+
+permissions:
+  contents: read
+
+jobs:
+  exploratory-890000:
+    runs-on: ubuntu-24.04
+    timeout-minutes: 180
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+          cache: pip
+      - name: Install FuckMark
+        run: python -m pip install -e .
+      - name: Install exact CPU Torch runtime
+        run: python -m pip install "torch==2.13.0" --index-url https://download.pytorch.org/whl/cpu
+      - name: Install maintained Transformers SynthID dependencies
+        run: python -m pip install -r requirements-smoke.txt
+      - name: Run Cycle 8 exploratory detector compare on seed 890000
+        run: >-
+          python -m fuckmark.cycle8_hf
+          --device "${{ inputs.device || 'cpu' }}"
+          --max-attempts "${{ inputs.max_attempts || '64' }}"
+          --detector-json artifacts/cycle8-exploratory-890000/detector-compare.json
+      - uses: actions/upload-artifact@v4
+        if: always()
+        with:
+          name: cycle8-exploratory-890000
+          path: artifacts/cycle8-exploratory-890000/
+          if-no-files-found: error

CHANGELOG.md

@@ -1,5 +1,18 @@
 # Changelog
 
+## Unreleased — exact user-visible invariance
+
+- Restored Priority Zero: product output must keep the exact user-visible projection of the input.
+- Public CLI `release-cli-v4` fail-closes instead of applying contractions or other visible edits.
+- `release_transform_registry()` is the empty product-visible-invariance registry. Historical contraction/lexical/syntax/spacing catalogs remain replayable under explicit historical names.
+- Added `specs/fuckmark-user-visible-invariance-v1.contract.json` and Cycle 8 invisible-carrier research. No product-authorized carrier is promoted yet.
+- Reclassified Cycle 6 formal `NONZERO_RESIDUAL` 7/192 and Cycle 7 visible-edit stages as scientific evidence that is product-disqualified.
+- GitHub CI installed-CLI end-to-end now expects unchanged visible text, matching `release-cli-v4`.
+- Cycle 8 fixture compare reports visible-projection pass `20/20` for space-carrier arms.
+- Chromium `pre` screenshots: U+034F / U+FE00 pixel-equal; U+200C PNG bytes differ.
+- Detector-blind GPT-2 / SynthID development scoring on seeds `890000`, `900000`, and `910000` reduced watermarked detections for U+034F space-carrier x1 without unwatermarked inflation on these tiny 4-pair corpora. This is not confirmation and is not a product promotion.
+- Product contract loading uses the embedded v1 payload in installed wheels. Writers that replace pinned `specs/cycle8` tokenizer JSON refuse to run without a GPT-2 encoder; the detector harness still scores when tiktoken is absent. Chromium HTML evidence encodes less-than and greater-than as JSON unicode escapes so an ASCII script terminator cannot close the inline script.
+
 ## v0.2.0 — Exact-survival confirmation and release hardening
 
 ### Research

README.md

@@ -9,9 +9,15 @@
 
 # FuckMark
 
-FuckMark is a deterministic, detector-blind research harness for studying how small text transformations change statistical watermark observations while preserving the source text as much as possible.
+FuckMark is a deterministic, detector-blind research harness and CLI for disrupting statistical text watermarks.
 
-**Current release: v0.2.0**  
+**Product contract (Priority Zero):** the user-visible text must not change. Product-safe transforms may modify only rigorously validated non-rendering Unicode representation while preserving exact visible projection. Semantic equivalence, "minor" punctuation edits, and extra visible spaces are not product success.
+
+The public CLI currently fail-closes: no product-authorized invisible carrier has been promoted yet, so `FuckMark` returns the original text unchanged rather than applying contractions or other visible edits. That is intentional. Historical visible-edit research remains replayable through explicitly named historical registries.
+
+See [`docs/product-contract.md`](docs/product-contract.md) and `specs/fuckmark-user-visible-invariance-v1.contract.json`.
+
+**Current tagged release: v0.2.0**  
 Website: [mark.q1z.org](https://mark.q1z.org)  
 License: MIT
 
@@ -36,11 +42,11 @@ The expanded `content-region-destruction-v1` pool and pairwise-completion schedu
 
 ## Claim boundary
 
-The 97.40% result is narrow research evidence, not a universal watermark-removal guarantee.
+The 97.40% Cycle 4 result is narrow research evidence, not a universal watermark-removal guarantee, and it is not a product-contract success: those schedulers selected visible-edit candidates.
 
-It applies to the repository's frozen open GPT-2 / Hugging Face SynthID Weighted Mean confirmation configuration, fixed threshold, exact source protocol, and measured corpus. It does not establish proprietary-detector transfer, arbitrary-model transfer, unknown future watermark transfer, perfect removal, or semantic equivalence under every input.
+It applies to the repository's frozen open GPT-2 / Hugging Face SynthID Weighted Mean confirmation configuration, fixed threshold, exact source protocol, and measured corpus. It does not establish proprietary-detector transfer, arbitrary-model transfer, unknown future watermark transfer, or perfect removal.
 
-The public CLI also remains deliberately conservative. It uses only the release transform registry. Development schedulers, detector code, experimental search, and the quarantined U+200C diagnostic registry are not automatically selected by the command-line product path.
+The public CLI uses only `release_transform_registry()`, which is now the product-visible-invariance registry. That registry currently contains zero authorized carriers. Development schedulers, detector code, experimental search, historical contraction/lexical/syntax rules, Cycle 6 spacing, Cycle 7 durable visible edits, and the quarantined U+200C diagnostic registry are not automatically selected by the command-line product path.
 
 ## Install
 
@@ -99,6 +105,8 @@ Stream mode:
 printf 'I do not agree.\n' | FuckMark
 ```
 
+The current product CLI prints the same visible text. It does not emit `I don't agree.`
+
 File mode:
 
 ```text
@@ -136,11 +144,17 @@ The U+200C mechanism is not promoted into release behavior. It can strongly pert
 
 ## Cycle 6 development status
 
-Cycle 6 reached **0 of 16 detected** on the frozen seed-720000 development corpus with the unchanged `0.5570987654320988` threshold and the raw, NFKC, Cf-strip, and combined sanitizer arms. The timestamp-independent corpus content hash is `b114cf4d869c5a5d78ac52855a1a480b1f0e605137aee2cb269062880fcc22d3`; the stable scored-result hash is `73f44b173d7ec55ea80e7d2e9a46b7ea70d8b32682f465dc66643387d99aa8b6`.
+Cycle 6 reached **0 of 16 detected** on the frozen seed-720000 development corpus with the unchanged `0.5570987654320988` threshold and the raw, NFKC, Cf-strip, and combined sanitizer arms. Formal confirmation froze **NONZERO_RESIDUAL 7/192** transformed-WM detections (2/64 + 2/64 + 3/64) with 0/192 matched transformed-UW detections.
+
+This is detector-endpoint scientific evidence, **PRODUCT_DISQUALIFIED**. The Cycle 6 workhorse inserted repeated ordinary ASCII U+0020 spaces, which are visible spacing changes. Do not treat 7/192, 0/16, or the older 96%+ Cycle 4 figure as proof that FuckMark currently meets the product goal.
+
+## Cycle 7 status
+
+Cycle 7 Stage A/B/C researched collapse-resistant *visible* edits (contractions, hyphenation, punctuation, newlines). Those results remain historically valid as **INSUFFICIENT_EVIDENCE** and are **PRODUCT_DISQUALIFIED**. They must not enter `release_transform_registry()`.
 
-This is detector-endpoint development evidence, not geometry zero and not a release claim. Surviving original observations remain in every output. The attack also relies on visible repeated ASCII spaces: NFKC and Cf stripping preserve them, but repeated-space canonicalization removes the mechanism and human fidelity review is still pending.
+## Cycle 8 status
 
-The Cycle 6 formal-confirmation harness is preregistered for three new 64-sample corpora with matched transformed-unwatermarked controls. Sealed detector scoring remains blocked until the full all-16 blind B14 fidelity packet receives independent human adjudication and its audit hash is bound into the confirmation contract.
+Cycle 8 is the exact-visible-projection generation. See [`docs/cycle8/README.md`](docs/cycle8/README.md). U+200C remains a diagnostic baseline (stripped by Cf). Combining grapheme joiner U+034F is an early durable-track *hypothesis* under frozen sanitizers; it is not a release mechanism.
 
 ## Reproduce the research environment
 

docs/cli.md

@@ -1,10 +1,12 @@
 # FuckMark CLI
 
-FuckMark v0.2.0 exposes the same command under `FuckMark`, `Fuckmark`, and `fuckmark`. All aliases use `release-cli-v3` and the same content-addressed release transform registry.
+FuckMark v0.2.0 still exposes the same command under `FuckMark`, `Fuckmark`, and `fuckmark`. Current `main` uses `release-cli-v4` and the product visible-invariance registry.
 
-The public CLI is intentionally smaller than the research harness. It does not load detector code, watermark keys, model weights, network services, the Cycle 4 search schedulers, or quarantined U+200C diagnostics.
+The public CLI is intentionally smaller than the research harness. It does not load detector code, watermark keys, model weights, network services, Cycle 4/6/7 visible-edit schedulers, or quarantined U+200C diagnostics. It also does not apply contractions.
 
-Install instructions for Windows, macOS, and Linux are in [`install.md`](install.md). The official website is [mark.q1z.org](https://mark.q1z.org).
+**Priority Zero:** the CLI must not change user-visible text. Until a product-authorized invisible carrier exists, the CLI fail-closes and returns the input unchanged.
+
+Install instructions for Windows, macOS, and Linux are in [`install.md`](install.md). The official website is [mark.q1z.org](https://mark.q1z.org). The product contract is [`product-contract.md`](product-contract.md).
 
 ## Interactive use
 
@@ -16,7 +18,7 @@ FuckMark
 
 Paste the source text and finish with `:done` on its own line. The legacy `ok` terminator remains accepted.
 
-Interactive mode reports the exact project version, processes the text deterministically, reports the number of accepted release-safe changes, and copies the result to the platform clipboard. Color is emitted only when standard output is an interactive terminal. `--no-color` and the `NO_COLOR` environment variable disable ANSI output.
+Interactive mode reports the exact project version, processes the text deterministically, reports whether any product-authorized invisible transform was applied, and copies the result to the platform clipboard. Color is emitted only when standard output is an interactive terminal. `--no-color` and the `NO_COLOR` environment variable disable ANSI output.
 
 ## Stream use
 
@@ -26,6 +28,8 @@ Piped input automatically selects stream mode:
 printf 'I do not agree.\n' | FuckMark
 ```
 
+Current product output is the same visible line, not `I don't agree.`
+
 Use `--stdin` or `--non-interactive` to request the same mode explicitly. Standard output contains only transformed text. Clipboard access is disabled unless `--copy` is provided.
 
 ## File use
@@ -47,14 +51,14 @@ Without `--output`, file-mode output is written to standard output. `--output -`
 | `--version` | Print project, CLI, and release-registry identities. |
 | `--stdin`, `--non-interactive` | Read all standard input in explicit stream mode. |
 | `-o FILE`, `--output FILE` | Atomically write transformed UTF-8 text to a file. |
-| `--copy` | Also copy transformed text to the system clipboard. |
+| `--copy` | Also copy transformed text to the platform clipboard. |
 | `-q`, `--quiet` | Hide interactive processing and completion messages. |
 | `--no-color` | Disable ANSI terminal color. |
 
 ## Exit status
 
 | Status | Meaning |
-| ---: | --- |
+| ---: | ---: |
 | 0 | Transformation and requested outputs succeeded. |
 | 1 | Input, transformation, or output validation failed. |
 | 2 | Transformation succeeded, but clipboard transfer failed. The transformed text is still written to the requested non-clipboard destination or printed as fallback output. |
@@ -69,4 +73,4 @@ Errors are written to standard error. Empty input is rejected. Clipboard command
 
 ## Release boundary
 
-The CLI uses only `release_transform_registry()`. The confirmed Cycle 4 exact-survival result belongs to the research/evaluation path and is not a claim that the public CLI itself produces a 97.40% evasion rate. Development profiles, detector scoring, experimental search, and the U+200C diagnostic registry remain outside automatic release behavior.
+The CLI uses only `release_transform_registry()`, which means the exact visible-projection product contract, not "semantically conservative English edits." Historical contraction, Cycle 6 spacing, Cycle 7 durable visible edits, detector scoring, experimental search, and the U+200C diagnostic registry remain outside automatic release behavior.

docs/cycle7/README.md

@@ -1,8 +1,10 @@
-# Cycle 7 — durable transform development
+# Cycle 7 — durable visible-edit research (PRODUCT_DISQUALIFIED)
 
-Cycle 6 is **frozen**. Formal outcome: `NONZERO_RESIDUAL` **7/192**, not 0/192.
+Cycle 6 is **frozen**. Formal outcome: `NONZERO_RESIDUAL` **7/192**, not 0/192. That detector result remains historically valid and is **not** product-compliant FuckMark removal: Cycle 6 inserted visible ASCII spaces.
 
-This directory starts Cycle 7 development. It does not retune Cycle 6 and does not authorize a new formal confirmation.
+Cycle 7 researched collapse-resistant *visible* transforms. Those edits may preserve meaning. They still change user-visible text, so they are **PRODUCT_DISQUALIFIED** for the CLI/product path restored under `fuckmark-user-visible-invariance-v1`.
+
+Cycle 8 (`docs/cycle8/`) is the exact-visible-projection generation. Do not promote Cycle 7 catalogs into `release_transform_registry()`.
 
 ## Live state (verified 2026-08-25)
 

docs/cycle8/README.md

@@ -0,0 +1,30 @@
+# Cycle 8 — exact visible projection
+
+Cycle 7 is frozen as **visible-edit durability research**. It is **PRODUCT_DISQUALIFIED** even where detector snapshots looked interesting.
+
+Cycle 8 is a new scientific generation. The mechanism class is hidden Unicode representation that leaves `VISIBLE(original)` identical to `VISIBLE(transformed)`.
+
+## Priority order
+
+0. Exact user-visible text preservation
+1. Real detector evasion
+2. Sanitizer durability
+3. Independent replication
+4. Protected machine-content safety
+5. Low hidden payload
+6. Mechanistic explanation
+7. Proxy metrics
+
+## Seed ledger
+
+See `docs/cycle8/seed-ledger.md`. Seeds `890000`, `900000`, `910000`, and `920000` were assigned before Cycle 8 generation. Do not inspect `830000` / `840000` / `850000`. Do not inspect Cycle 7 reserved `880000`.
+
+## Carrier research
+
+See `docs/cycle8/carrier-research.md`.
+
+U+200C is the diagnostic baseline. It is not durable under Cf-strip.
+
+Early durable-track *hypothesis* (sanitizer and tokenizer screen, not a detector claim): U+034F COMBINING GRAPHEME JOINER and variation selectors U+FE00..U+FE0F. These are not product-authorized.
+
+Fixture compare `specs/cycle8/fixture-compare-v1.json` reports visible-projection pass `20/20` across identity, U+200C, U+034F x1, U+034F x8, and U+FE00 arms. Quote interiors remain protected spans. Seed `890000` exploratory detector scoring is development-only (`PROMISING_DEVELOPMENT`, not confirmation). U+034F is not product-authorized.

docs/cycle8/carrier-research.md

@@ -0,0 +1,41 @@
+# Cycle 8 carrier research
+
+Labels: `VERIFIED`, `HYPOTHESIS`, `REJECTED`, `PRODUCT_DISQUALIFIED`, `HISTORICAL_ONLY`, `UNKNOWN`.
+
+## H1. Normalization-stable non-Cf carrier
+
+`HYPOTHESIS` for product use. Sanitizer screen `VERIFIED` for U+034F and U+FE00..U+FE0F on the frozen arms:
+
+- category Mn, default-ignorable, combining class 0;
+- NFC/NFKC-stable in isolation;
+- survives Cf-strip (not Cf);
+- survives `whitespace-collapse-v1` when placed after an ASCII space.
+
+GPT-2 tokenizer screen `VERIFIED` on the Cycle 8 fixture (tiktoken `gpt2`): inserting U+034F after ASCII word spaces changed token IDs (`ids_equal=false`, token-count delta +40) with suffix realignment of only one token. An 8-copy space run produced delta +208. U+200C also disrupts GPT-2 (delta +28) but remains Cf-fragile.
+
+Not product-authorized. Fixture compare on four ASCII texts is `VERIFIED` for visible projection (`20/20`) and for U+034F / U+FE00 survival under Cf-strip, NFKC, and `whitespace-collapse-v1`. Chromium `pre` screenshots: U+034F and U+FE00 are `VERIFIED` pixel-equal to the original; U+200C is `REJECTED` (PNG bytes differ). Detector-blind GPT-2 / SynthID looks on seeds `890000`, `900000`, and `910000` are `PROMISING_DEVELOPMENT` / `HYPOTHESIS` only (four watermarked pairs per seed). Prefer U+034F x1: x8 overflowed GPT-2's 1024-token context on seed `900000`. The public CLI still authorizes zero carriers.
+
+## H2. Carrier runs at space boundaries
+
+`HYPOTHESIS`. Repeating an approved carrier after existing ASCII spaces increases hidden payload without splitting `[A-Za-z]+` words, so historical hard-invariant word matching still passes. This is the TransformRegistry-compatible high-density channel. Intra-word insertion is tokenizer-diagnostic only because it breaks the frozen hard-invariant word regex.
+
+## H3-H8
+
+Placement geometry, Cycle 6 scheduler reuse, and root-window correlation remain `HYPOTHESIS` for larger corpora. Tiny-corpus detector looks on seeds `890000`, `900000`, and `910000` are `PROMISING_DEVELOPMENT` only.
+
+## H9. Feasibility boundary
+
+Still `HYPOTHESIS`. If every tokenizer-disruptive invisible carrier is removed by ordinary sanitizers, Cycle 8 will freeze that boundary rather than weaken Priority Zero.
+
+## Rejected as product mechanisms
+
+| Mechanism | Label | Why |
+| --- | --- | --- |
+| Contractions / Cycle 7 durable edits | `PRODUCT_DISQUALIFIED` | visible words or punctuation change |
+| Cycle 6 U+0020 runs | `PRODUCT_DISQUALIFIED` | visible spacing change |
+| U+200C / other Cf | `REJECTED` as durable; diagnostic only | Cf-strip restores the original string |
+| NBSP / hair spaces / dashes / homoglyphs | `REJECTED` | visible or NFKC-mapped layout/glyph change |
+
+## Baseline
+
+U+200C after ASCII word spaces: visibility-aligned on the v1 projection, Cf-fragile, and Chromium-pixel-`REJECTED`. Keep as the diagnostic control. Do not promote to release.