ithrl 0.2.0__tar.gz → 0.3.0__tar.gz

ithrl-0.3.0/CHANGELOG.md

@@ -0,0 +1,70 @@
+# CHANGELOG
+
+## v0.3.0 (2026-11-10)
+
+### Features
+
+- Add finer final-response reduction ([#52](https://github.com/eberkut/ithrl/pull/52), [`562124a`](https://github.com/eberkut/ithrl/commit/562124ab071788a52e0d50dd538048ee67de464b))
+- Define source-bound text-section removal ([#51](https://github.com/eberkut/ithrl/pull/51), [`56c9e6c`](https://github.com/eberkut/ithrl/commit/56c9e6cfb24a83dc32573232b447e574cb0f616f))
+- Inspect retained context without provider calls ([#49](https://github.com/eberkut/ithrl/pull/49), [`5604688`](https://github.com/eberkut/ithrl/commit/5604688e269543305fe4b7b253a703dc404c1139))
+- **reduce**: Qualify paragraph removal on ordinary sessions ([#58](https://github.com/eberkut/ithrl/pull/58), [`5c6ae26`](https://github.com/eberkut/ithrl/commit/5c6ae26760ea7bd94b406ebf9ba3214e5c4ce64c))
+
+### Refactoring
+
+- Simplify experiment workflows and test fixtures ([#50](https://github.com/eberkut/ithrl/pull/50), [`782bfa2`](https://github.com/eberkut/ithrl/commit/782bfa2692a49a5bb2a4f555801abda23cde6fda))
+
+## v0.2.0 (2026-11-02)
+
+### Bug Fixes
+
+- **codex**: Bound and account for physical provider requests ([#33](https://github.com/eberkut/ithrl/pull/33), [`88529d6`](https://github.com/eberkut/ithrl/commit/88529d6f8b930dbcc546a0011f085c67d0bef1ac))
+- **copilot**: Recover catalogue connections and qualify fuzzing ([#37](https://github.com/eberkut/ithrl/pull/37), [`ff6cf63`](https://github.com/eberkut/ithrl/commit/ff6cf6324325f7f6dad8f8c9c80ea58bfe4e249a))
+- **copilot**: Restore native session prefixes through the CLI ([#34](https://github.com/eberkut/ithrl/pull/34), [`36845dd`](https://github.com/eberkut/ithrl/commit/36845ddb8239c0a488552297aaa70e6211979e59))
+- **fuzz**: Preserve accounting and validate campaign recovery ([#25](https://github.com/eberkut/ithrl/pull/25), [`f1f4793`](https://github.com/eberkut/ithrl/commit/f1f479318fe914f3033667b65381f7101bd3aaea))
+- **sampling**: Enforce calibration budgets and report durable progress ([#32](https://github.com/eberkut/ithrl/pull/32), [`bd40577`](https://github.com/eberkut/ithrl/commit/bd40577c734d70b36e59a9e6d3babafa1502bfd3))
+
+### Features
+
+- **extract**: Support literal predicates for native decisions ([#31](https://github.com/eberkut/ithrl/pull/31), [`3b18c2b`](https://github.com/eberkut/ithrl/commit/3b18c2bdd814c4e4c3bf79825a8cab77eb433661))
+- **fuzz**: Add mutation and sampled comparison services ([`56ee8b8`](https://github.com/eberkut/ithrl/commit/56ee8b8c5e1370b91db6c9c952eb8ceec19c95ae))
+- **fuzz**: Expose inspectable campaigns and saved results ([#24](https://github.com/eberkut/ithrl/pull/24), [`e60e472`](https://github.com/eberkut/ithrl/commit/e60e47230b99e50ff42de7cd8fb89b3769e6d6e2))
+- **fuzz**: Persist campaigns and recover interrupted execution ([#22](https://github.com/eberkut/ithrl/pull/22), [`806667f`](https://github.com/eberkut/ithrl/commit/806667f47e8bae1ccbf9b32fadc511ca559675d0))
+- **fuzz**: Qualify native Codex text appends ([#28](https://github.com/eberkut/ithrl/pull/28), [`a5b880b`](https://github.com/eberkut/ithrl/commit/a5b880b60ef9834bb57b2c4950e18cd17ff0b275))
+
+### Performance Improvements
+
+- Streamline local workflows and harden result publication ([#36](https://github.com/eberkut/ithrl/pull/36), [`43b9d81`](https://github.com/eberkut/ithrl/commit/43b9d81d0dcff85708da5c914986a68bbe8141db))
+
+## v0.1.0 (2026-10-23)
+
+### Bug Fixes
+
+- **adapters**: Detect saved session formats structurally ([`c8256b9`](https://github.com/eberkut/ithrl/commit/c8256b93ce79fb9c3f025e2b92a970f336497218))
+- **reduce**: Validate native candidates before sampled dispatch ([`90d4c0d`](https://github.com/eberkut/ithrl/commit/90d4c0d11a206389d44c6a59991df53c19802514))
+- **runner**: Clarify Codex isolation prerequisites ([`4a409c1`](https://github.com/eberkut/ithrl/commit/4a409c1f37caf301a00cfbd20d3a95cf15844eb9))
+
+### Features
+
+- Establish credential-free context reduction core ([`c517409`](https://github.com/eberkut/ithrl/commit/c517409c06e40c92b24a8b990685abd4acc9a9ba))
+- **adapters**: Parse Codex and Copilot session evidence ([`3b81c26`](https://github.com/eberkut/ithrl/commit/3b81c2607397d884bcfb1b19dbf995f66ccb0107))
+- **calibration**: Add bounded sampled runner calibration ([`f9db180`](https://github.com/eberkut/ithrl/commit/f9db18012565badc2cf2a3a4a6158487d4c5fb70))
+- **case**: Preserve immutable native source evidence ([`d108b96`](https://github.com/eberkut/ithrl/commit/d108b96487ccb69f6b36642ae331d36272038722))
+- **cli**: Add an installed example and focused public guides ([`1efc2ff`](https://github.com/eberkut/ithrl/commit/1efc2ffb03c734b666b1d1da9b5610ad62970267))
+- **cli**: Add guarded native calibration ([`06d2bfb`](https://github.com/eberkut/ithrl/commit/06d2bfb3f4d34f4dc845d9d6877257e4e79f9120))
+- **cli**: Integrate sampled reduction and exact resume ([`40be766`](https://github.com/eberkut/ithrl/commit/40be7664ba90ac5341d93fb5a725dbfa0efd48d3))
+- **extract**: Support ordinary Codex and Copilot sessions ([`3cd0c10`](https://github.com/eberkut/ithrl/commit/3cd0c1092aeb3669b86fdae07e887d28ba9d685f))
+- **reduce**: Add sampled contracts and durable progress ([`650252d`](https://github.com/eberkut/ithrl/commit/650252d76e8d7d72f3f5715ecaf488e752dc18b1))
+- **reduce**: Integrate sampled candidate search ([`72958d8`](https://github.com/eberkut/ithrl/commit/72958d8d6ae0307ca2d090a2a10d02606bafdea6))
+- **runner**: Add Codex final-response execution ([`ad30a9a`](https://github.com/eberkut/ithrl/commit/ad30a9a17b0e665056eca9b415cbf4e2770a7bf7))
+- **runner**: Add controlled Copilot execution ([`c9d1bb2`](https://github.com/eberkut/ithrl/commit/c9d1bb23bfdd98bb022fd0b62ca9b8a81021fe4b))
+- **runner**: Add isolated execution runtime ([`17d2224`](https://github.com/eberkut/ithrl/commit/17d22240414cdb32d4464f16e7b2621d99729ca0))
+- **runner**: Add shared execution foundation ([`89536e9`](https://github.com/eberkut/ithrl/commit/89536e919730fc0abf680a2928a89761c78339f5))
+- **runner**: Validate isolated native calibration ([`b0cd438`](https://github.com/eberkut/ithrl/commit/b0cd438d1fa280f0ddafa39d23653a7ac926da1b))
+- **sampling**: Collect bounded fresh-baseline blocks ([`16abf30`](https://github.com/eberkut/ithrl/commit/16abf30b964eae299177d1e1dade6f545ffbda31))
+- **sampling**: Support paced execution and safe overload replacement ([`5d70fbd`](https://github.com/eberkut/ithrl/commit/5d70fbd26901a0574883e07fd4055644c8515e80))
+- **statistics**: Add fixed-sample candidate classification ([`fb26fca`](https://github.com/eberkut/ithrl/commit/fb26fcaf7c064f49ac7d63b4e2ceb1ab3d6dcdb5))
+- **workflow**: Confirm sampled reductions with fresh evidence ([`ee291e4`](https://github.com/eberkut/ithrl/commit/ee291e4d906394452da61640989cef594f62caf7))
+
+### Refactoring
+
+- Simplify execution paths and avoid redundant context copies ([`34ab361`](https://github.com/eberkut/ithrl/commit/34ab36165fed08678a80679b3f1753e6f26f63f8))

{ithrl-0.2.0 → ithrl-0.3.0}/CONTRIBUTING.md

@@ -26,8 +26,9 @@ For task-specific instructions, see [documentation](#documentation),
 
 ## Development setup
 
-Use [uv](https://docs.astral.sh/uv/) and Python 3.12, 3.13 or 3.14. Python 3.14 is the primary
-development version. From the repository root:
+Use [uv](https://docs.astral.sh/uv/) at the version required by `tool.uv.required-version` in
+`pyproject.toml`, and Python 3.12, 3.13 or 3.14. CI reads the same uv pin through `setup-uv`'s
+`version-file` input. Python 3.14 is the primary development version. From the repository root:
 
 ```bash
 uv sync --locked --all-groups --all-extras
@@ -82,6 +83,8 @@ The required `Python tests` check summarizes the test matrix. It accepts a skipp
 after successful Markdown-only classification. Failed classification, missing decisions, test
 failures and unexpected skips fail this check.
 
+The logo sources, usage rules and regeneration commands are in [branding](docs/branding/README.md).
+
 ## Workflow and dependency checks
 
 For GitHub Actions or Dependabot changes, use the jactionlint release pinned in
@@ -128,13 +131,21 @@ Use `uv run rumdl fmt .` to format Markdown, then `uv run rumdl check .` to lint
 version and `pyproject.toml` configuration apply to pre-commit and CI. Prose wraps at 100
 characters; commands and tables retain their layout. MkDocs syntax is recognized within `docs/`. The
 generated CLI reference has narrow exceptions for Typer's prompts, option alignment and wrapping;
-its byte-for-byte generation check remains enforced. The generated changelog is checked under the
-shared rules.
+its byte-for-byte generation check remains enforced. The generated changelog keeps each entry on
+one source line, with a line-length exception.
 
 Keep installation and first-use instructions in README, and detailed user guides and reference in
 `docs/`. Support, security and conduct policies live at the repository root. Keep private sessions,
 cases and credentials out of documentation and generated output.
 
+Use [Diátaxis](https://diataxis.fr/) to match a section to the reader's need. A tutorial guides
+learning through a reproducible example with expected results. A how-to gives prerequisites,
+actions and a success check for a specific task. Reference describes commands and contracts;
+explanation covers rationale, evidence and limits. Link between these when the reader needs more
+detail. Apply the distinction within existing pages; add a page when it has a separate purpose.
+Use concrete instructions and claims supported by evidence. Remove filler, promotional language
+and repeated explanations.
+
 [ROADMAP.md](ROADMAP.md) describes future direction. GitHub issues own actionable scope,
 dependencies, acceptance criteria and status; release milestones group scope and the public
 [Project](https://github.com/users/eberkut/projects/1/views/4) provides views of those same issues.
@@ -300,8 +311,13 @@ justified and documented before release.
 Python Semantic Release, pinned in the `release` dependency group, calculates versions from
 Conventional Commits. `pyproject.toml` is the version source; installed CLI and Codex client
 metadata read that distribution version. The tool generates `CHANGELOG.md` from the full release
-history, omitting routine build, CI, documentation, style and test commits. Do not edit the
-generated changelog by hand; review its contents before approving a release.
+history, omitting routine build, CI, documentation, style and test commits. Its
+[template](.github/changelog/CHANGELOG.md.j2) keeps lists compact, with one entry per source line.
+Change the template to adjust formatting; do not edit generated entries by hand.
+
+A maintainer starts preparation; Semantic Release chooses the version and writes the notes.
+Review and merge the generated files through a PR so CI tests the prepared version before the
+publication workflow builds it. The publication workflow uses that committed version.
 
 Use `main` with complete history and all existing release tags available. Preview the next version
 without changing files, the index, commits, tags or remote state:
@@ -339,9 +355,9 @@ ownership and the following repository/index settings:
 
 - PyPI and TestPyPI Trusted Publishers bind `eberkut/ithrl`, `release.yml` and their respective
   `pypi` / `testpypi` environments. Publishing uses OIDC, without a stored PyPI API token.
-- Both environments admit only `main`. `pypi` requires the maintainer's approval, permits
-  self-review and disallows administrative approval bypass. Environment declarations in YAML
-  do not configure these protections; verify the actual settings before dispatch.
+- Both environments admit only `main`, require the maintainer's approval, permit self-review and
+  disallow administrative approval bypass. Environment declarations in YAML do not configure these
+  protections; verify the actual settings before dispatch.
 - Release-tag rules permit the maintainer's signed-tag creation and prevent unapproved tag
   replacement or deletion. The workflow never creates a tag.
 
@@ -353,10 +369,12 @@ gh workflow run release.yml --repo eberkut/ithrl --ref main \
   -f version=X.Y.Z -f source-sha=FULL_APPROVED_COMMIT_SHA
 ```
 
-1. Inspect that run's checks and successful **Publish TestPyPI rehearsal** job. Its summary records
+1. After source verification and validation pass, approve the `testpypi` deployment on the run's
+   GitHub page. This authorizes the public TestPyPI upload.
+2. Inspect the successful **Publish TestPyPI rehearsal** job. Its summary records
    the source, version, immutable artifact ID and checksum-manifest digest. The production job
    waits at the `pypi` environment. TestPyPI files, logs and artifacts are public.
-2. Download that run's `release-distributions` artifact into a new review directory. For example:
+3. Download that run's `release-distributions` artifact into a new review directory. For example:
 
    ```bash
    gh run download RUN_ID --repo eberkut/ithrl \
@@ -367,7 +385,7 @@ gh workflow run release.yml --repo eberkut/ithrl --ref main \
    from the review directory. Inspect the release notes and confirm TestPyPI serves the same
    distribution hashes. Complete a clean TestPyPI installation/example rehearsal before approval;
    resolve dependencies only from the normal PyPI index, without mixing indexes.
-3. After the rehearsal passes and tag creation is approved, create and verify the signed tag at
+4. After the rehearsal passes and tag creation is approved, create and verify the signed tag at
    the frozen source commit, then push only that tag through the permitted tag-rules path:
 
    ```bash
@@ -376,12 +394,12 @@ gh workflow run release.yml --repo eberkut/ithrl --ref main \
    git push origin refs/tags/vX.Y.Z
    ```
 
-4. Review and approve the `pypi` deployment on the run's GitHub page. The job verifies the existing
+5. Review and approve the `pypi` deployment on the run's GitHub page. The job verifies the existing
    annotated tag targets the frozen source, rechecks the files, and uploads to PyPI. The final job
    rechecks the tag/files and creates the GitHub Release with those distributions, notes and
    checksums. Signature verification belongs to the preceding maintainer tag check; the workflow
    checks the tag object and target commit.
-5. Verify the published versions, hashes, attestations and installed example. A successful upload
+6. Verify the published versions, hashes, attestations and installed example. A successful upload
    alone does not complete release acceptance.
 
 Complete review within 30 days: release artifacts are retained for 30 days, and GitHub limits the

{ithrl-0.2.0 → ithrl-0.3.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: ithrl
-Version: 0.2.0
+Version: 0.3.0
 Summary: Find smaller agent contexts that preserve a selected model behavior
 Project-URL: Source, https://github.com/eberkut/ithrl
 Project-URL: Issues, https://github.com/eberkut/ithrl/issues
@@ -31,7 +31,15 @@ Provides-Extra: fuzz
 Requires-Dist: scipy>=1.18; extra == 'fuzz'
 Description-Content-Type: text/markdown
 
-# ithrl
+<!-- rumdl-disable MD033 -->
+<h1>
+  <picture>
+    <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/eberkut/ithrl/main/docs/branding/ithrl-logo-dark.svg">
+    <source media="(prefers-color-scheme: light)" srcset="https://raw.githubusercontent.com/eberkut/ithrl/main/docs/branding/ithrl-logo-light.svg">
+    <img src="https://raw.githubusercontent.com/eberkut/ithrl/main/docs/branding/ithrl-logo.png" alt="ithrl" width="320" height="158">
+  </picture>
+</h1>
+<!-- rumdl-enable MD033 -->
 
 [![PyPI](https://img.shields.io/pypi/v/ithrl.svg)](https://pypi.org/project/ithrl/)
 [![CI](https://github.com/eberkut/ithrl/actions/workflows/ci.yml/badge.svg?branch=main)](https://github.com/eberkut/ithrl/actions/workflows/ci.yml)
@@ -88,6 +96,9 @@ Use a new directory to repeat the example; extraction refuses to overwrite an ex
 
 - `ithrl extract` browses saved sessions and creates a case from a selected decision.
 - `ithrl shrink CASE` searches for context removals that preserve the selected behavior.
+- `ithrl inspect retained CASE --result RESULT` explains retained units and recorded removal trials
+  from a saved reduction without model access. See
+  [results and privacy](https://github.com/eberkut/ithrl/blob/main/docs/results.md#inspect-retained-context).
 - `ithrl fuzz CASE --config PRIVATE_TOML` searches for and simplifies text appends that decrease how
   often the selected behavior occurs, with independent confirmation.
 
@@ -107,8 +118,8 @@ Returned actions are observed and denied. Cases contain copied session content;
 | Input | Extracted decisions | Reduction | Fuzzing |
 | --- | --- | --- | --- |
 | Bundled synthetic example | Synthetic tool request | Deterministic local runner | Local fake runner |
-| Codex CLI rollouts | Final responses | Experimental native runner | Experimental prior-user text appends |
-| Copilot CLI session events | Final responses and shell requests | Experimental final-response and denied `bash` runner | Experimental prior-user text appends for final-response targets |
+| Codex CLI rollouts | Final responses | Experimental native runner; [paragraph-removal measurements](https://github.com/eberkut/ithrl/blob/main/docs/measurements.md#ordinary-session-paragraph-removal) | Experimental prior-user text appends |
+| Copilot CLI session events | Final responses and shell requests | Experimental final-response paragraph removal and denied `bash` runner | Experimental prior-user text appends for final-response targets |
 
 Native Codex and Copilot final-response fuzzing supports appending text to eligible prior user
 input while preserving the fixed triggering input. The
@@ -116,9 +127,12 @@ input while preserving the fixed triggering input. The
 starts with a local example and covers native execution.
 
 [Saved-session measurements](https://github.com/eberkut/ithrl/blob/main/docs/measurements.md#saved-session-experiments)
-include a confirmed reduction for one Codex session and its chosen predicate: 50.17% lower mean
-input tokens per confirmation observation, using a 20-percentage-point preservation margin.
-The measurements report resource usage and unsuccessful outcomes alongside completed runs. Check
+include ordinary Codex and Copilot final-response experiments that removed paragraphs from retained
+prior-user text and passed fresh confirmation at a 20-percentage-point preservation margin. The
+[worked investigation](https://github.com/eberkut/ithrl/blob/main/docs/research/ordinary-session-investigation.md)
+explains what the smaller Codex context helped a developer inspect and what remains unresolved.
+The measurements distinguish private runtime qualification from integrated support and report
+resource usage and unsuccessful outcomes alongside completed runs. Check
 [native compatibility](https://github.com/eberkut/ithrl/blob/main/docs/native-runs.md#compatibility-and-measurements)
 for the installed harness's requirements.
 

{ithrl-0.2.0 → ithrl-0.3.0}/README.md

@@ -1,4 +1,12 @@
-# ithrl
+<!-- rumdl-disable MD033 -->
+<h1>
+  <picture>
+    <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/eberkut/ithrl/main/docs/branding/ithrl-logo-dark.svg">
+    <source media="(prefers-color-scheme: light)" srcset="https://raw.githubusercontent.com/eberkut/ithrl/main/docs/branding/ithrl-logo-light.svg">
+    <img src="https://raw.githubusercontent.com/eberkut/ithrl/main/docs/branding/ithrl-logo.png" alt="ithrl" width="320" height="158">
+  </picture>
+</h1>
+<!-- rumdl-enable MD033 -->
 
 [![PyPI](https://img.shields.io/pypi/v/ithrl.svg)](https://pypi.org/project/ithrl/)
 [![CI](https://github.com/eberkut/ithrl/actions/workflows/ci.yml/badge.svg?branch=main)](https://github.com/eberkut/ithrl/actions/workflows/ci.yml)
@@ -55,6 +63,9 @@ Use a new directory to repeat the example; extraction refuses to overwrite an ex
 
 - `ithrl extract` browses saved sessions and creates a case from a selected decision.
 - `ithrl shrink CASE` searches for context removals that preserve the selected behavior.
+- `ithrl inspect retained CASE --result RESULT` explains retained units and recorded removal trials
+  from a saved reduction without model access. See
+  [results and privacy](https://github.com/eberkut/ithrl/blob/main/docs/results.md#inspect-retained-context).
 - `ithrl fuzz CASE --config PRIVATE_TOML` searches for and simplifies text appends that decrease how
   often the selected behavior occurs, with independent confirmation.
 
@@ -74,8 +85,8 @@ Returned actions are observed and denied. Cases contain copied session content;
 | Input | Extracted decisions | Reduction | Fuzzing |
 | --- | --- | --- | --- |
 | Bundled synthetic example | Synthetic tool request | Deterministic local runner | Local fake runner |
-| Codex CLI rollouts | Final responses | Experimental native runner | Experimental prior-user text appends |
-| Copilot CLI session events | Final responses and shell requests | Experimental final-response and denied `bash` runner | Experimental prior-user text appends for final-response targets |
+| Codex CLI rollouts | Final responses | Experimental native runner; [paragraph-removal measurements](https://github.com/eberkut/ithrl/blob/main/docs/measurements.md#ordinary-session-paragraph-removal) | Experimental prior-user text appends |
+| Copilot CLI session events | Final responses and shell requests | Experimental final-response paragraph removal and denied `bash` runner | Experimental prior-user text appends for final-response targets |
 
 Native Codex and Copilot final-response fuzzing supports appending text to eligible prior user
 input while preserving the fixed triggering input. The
@@ -83,9 +94,12 @@ input while preserving the fixed triggering input. The
 starts with a local example and covers native execution.
 
 [Saved-session measurements](https://github.com/eberkut/ithrl/blob/main/docs/measurements.md#saved-session-experiments)
-include a confirmed reduction for one Codex session and its chosen predicate: 50.17% lower mean
-input tokens per confirmation observation, using a 20-percentage-point preservation margin.
-The measurements report resource usage and unsuccessful outcomes alongside completed runs. Check
+include ordinary Codex and Copilot final-response experiments that removed paragraphs from retained
+prior-user text and passed fresh confirmation at a 20-percentage-point preservation margin. The
+[worked investigation](https://github.com/eberkut/ithrl/blob/main/docs/research/ordinary-session-investigation.md)
+explains what the smaller Codex context helped a developer inspect and what remains unresolved.
+The measurements distinguish private runtime qualification from integrated support and report
+resource usage and unsuccessful outcomes alongside completed runs. Check
 [native compatibility](https://github.com/eberkut/ithrl/blob/main/docs/native-runs.md#compatibility-and-measurements)
 for the installed harness's requirements.
 

{ithrl-0.2.0 → ithrl-0.3.0}/ROADMAP.md

@@ -12,12 +12,12 @@ with scope and release dates decided from evidence.
 Codex and GitHub Copilot are the primary harness integrations. Closing feature gaps between them
 is the current priority.
 
-## Current work: context fuzzing
+## Completed: v0.2 context fuzzing
 
 The `ithrl fuzz` command explores bounded context changes, simplifies interventions and
 independently confirms their effects.
 
-The first fuzzing milestone supports **one confirmed finding per campaign**: a declared decrease
+The released fuzzing capability supports **one confirmed finding per campaign**: a declared decrease
 in the existing target's match probability among valid observations. A finding describes the
 selected experiment, including execution reliability and assumptions. It does not establish a
 model defect, a unique cause or semantic equivalence of the original and changed context.
@@ -36,27 +36,43 @@ See the [fuzzing capability issue](https://github.com/eberkut/ithrl/issues/5) an
 [`v0.2.0` milestone](https://github.com/eberkut/ithrl/milestone/1) for scope and acceptance
 criteria.
 
-[Release preparation and publication](https://github.com/eberkut/ithrl/issues/11) follow acceptance
-for both Codex and Copilot.
-They remain separate from implementing the capability. Broader mutation families and multiple-
-finding campaigns need their own evidence and prospective error allocation.
+[Version 0.2.0](https://github.com/eberkut/ithrl/releases/tag/v0.2.0) is published, with
+[release preparation and publication](https://github.com/eberkut/ithrl/issues/11) complete.
+Broader mutation families and multiple-finding campaigns need their own evidence and prospective
+error allocation.
+
+## Current work: v0.3 deeper reduction
+
+Provider-free retained-context inspection and finer removal within supported prior-user text are
+implemented, with Codex and Copilot final-response workflows. Ordinary saved-session qualification
+and a [worked investigation](docs/research/ordinary-session-investigation.md) are complete; the
+[measurements](docs/measurements.md#ordinary-session-paragraph-removal) record outcomes, costs and
+limits. Version 0.3 is not yet released.
+[Issue #12](https://github.com/eberkut/ithrl/issues/12) owns the milestone's acceptance and scope.
+
+Choose section boundaries, workload-specific usefulness criteria and execution budgets before
+provider-backed evaluation. Finer native removal needs its own reconstruction, safety and live-use
+evidence. Ablation, alternative reproducers, interaction analysis and statistically confirmed
+single-removal minimality are deferred; they are not v0.3 release requirements.
 
 ## Direction by version
 
 | Stage | Outcome | Candidate features |
 | --- | --- | --- |
 | **[v0.2: context fuzzing](https://github.com/eberkut/ithrl/issues/5)** | Find small context changes that alter a selected behavior | Bounded seeded mutations, discovery and mutation simplification, one fresh confirmed finding, exact recovery and explicit cost/reliability reporting |
-| **[v0.3: deeper reduction](https://github.com/eberkut/ithrl/issues/12)** | Examine how small and robust a reproducer is | Supported content-block/section removal, retained-context inspection, ablation, alternative reproducers, interactions and optional confirmed 1-minimality |
+| **[v0.3: deeper reduction](https://github.com/eberkut/ithrl/issues/12)** | Inspect retained context and reduce within supported prior-user text | Provider-free inspection, text-section removal and qualified Codex/Copilot final-response workflows |
 | **[v0.4: comparative debugging](https://github.com/eberkut/ithrl/issues/13)** | Isolate behavior-changing differences between compatible cases | `compare`, context-version delta debugging and another qualified harness integration |
 | **[v0.5: trace forensics](https://github.com/eberkut/ithrl/issues/15)** | Explain what information was acquired again | Passive reacquisition analysis around compaction, resume, forks and handoffs |
 
 ### Deeper reduction
 
 Reduction within a message must preserve its native structure and protocol dependencies. Retained
-context inspection should distinguish dependencies retained for validity, unremovable state and
-items whose removal changed the measured behavior.
+context inspection should distinguish protected or frozen material, dependency constraints,
+rejected removal trials, incomplete evidence and untested work. Trial explanations must identify
+the tested context; a rejected sampled screen does not establish behavioral necessity.
 
-Retained-item and full-original ablation answer different questions and must report their baseline.
+Deferred retained-item and full-original ablation answer different questions and must report their
+baseline.
 Alternative reproducer search can expose several preserving contexts and interactions; none is
 automatically a unique explanation. Deterministic single-removal minimality and statistically
 confirmed minimality are distinct claims. The latter requires additional complete comparisons,

{ithrl-0.2.0 → ithrl-0.3.0}/docs/architecture.md

@@ -8,7 +8,9 @@ runtime.
 
 The
 [README supported-inputs table](https://github.com/eberkut/ithrl/blob/main/README.md#supported-inputs)
-defines the public capability claim. The public commands are `extract`, `shrink` and `fuzz`. Fuzzing
+defines the public capability claim. The public commands are `extract`, `shrink`, `inspect retained`
+and `fuzz`. Retained-context inspection reads saved evidence without resolving runners or opening
+an observation store. Fuzzing
 has a credential-free local walkthrough and qualified Codex and Copilot append-text profiles for
 eligible prior user input. See [native runs](native-runs.md) for the operator
 procedure, [measurements](measurements.md#native-context-fuzzing) for qualification evidence and
@@ -79,6 +81,82 @@ run allowance leaves sample targets, freshness, effective policy and per-attempt
 Observation reuse requires compatible identities and roles; a matching candidate digest alone
 is insufficient.
 
+## Source-bound text sections
+
+The shared `text_sections.py` contract describes removal from one eligible prior-user text field.
+It is separate from native reconstruction and from search. Field discovery uses the existing
+adapter-owned editable-text surface: fixed decision input, instructions, protected nodes and opaque
+records remain outside it. A content block is a field; sections never cross block boundaries.
+
+`paragraph-v1` partitions the original string using physical LF, CRLF or CR line endings. A blank
+line contains only ASCII spaces or tabs before its ending. Each run of nonblank lines owns its
+following blank lines; leading blank lines form their own section. A whitespace-only field has one
+section, and an empty field has none. Other Unicode line separators are ordinary text. Offsets are
+half-open Python Unicode code-point positions; UTF-8 byte counts are reported separately. No
+trimming, newline conversion, separator insertion or Markdown interpretation occurs. Removing all
+sections leaves the existing field empty and preserves its enclosing record and content block.
+
+Catalogs contain metadata, not the original text. Section identities bind the immutable case,
+context node, content-block index, original-text digest, policy and exact range. A removal stores
+ordered section IDs rather than caller-supplied ranges. Validation re-derives the catalog from the
+immutable eligible field, refusing stale or foreign IDs, duplicates and noncanonical order. Fields
+are limited to 1 MiB of UTF-8 text and 1,024 sections; a field exceeding either limit is unavailable
+under this policy rather than partially catalogued.
+Case-level discovery omits fields beyond these limits while preserving complete catalogs for
+other eligible fields. Explicit validation still refuses an edit to an unavailable field.
+
+The candidate's existing mutation envelope distinguishes `append-text` from `remove-text-sections`.
+Empty and append-only candidates retain their existing serialization and identities. A section edit
+can accompany dependency-closed whole-record removal only when its parent survives; removing the
+same field through its parent or dependencies is refused. Appending still requires unchanged
+context. Exact-request materialization addresses the immutable original message before accounting
+for removed input positions and changes a defensive copy. A combined section edit refuses any
+dependency-closed removal intersecting the original last input, including a nested field, even
+when graph flags would allow coarse removal. The fixed input remains unchanged.
+Loading `Candidate` JSON validates its shape. Binding it to a `PreparedAttempt` invokes
+`prepare_candidate`, which re-derives its source bindings against immutable case evidence;
+well-formed JSON and digests alone do not authorize an edit. Project-owned schema tags remain `1`
+during pre-1.0 development.
+
+Retained-context inspection exposes metadata-only catalogs with a 4,096-unit report allowance;
+over-budget fields have no partial catalog. [Results](results.md#inspect-retained-context) explains
+availability, byte measures and interpretation. Original text remains behind the existing bounded
+opt-in.
+
+Codex and Copilot final-response materializers apply section removals to copied native input.
+Codex changes only the selected prior-user text block. Copilot changes the effective
+`transformedContent` when present, otherwise `content`, preserving the surrounding event and
+unedited prefix records. Copilot shell targets and fuzzing refuse section removal.
+
+Sampled reduction opts in through `reduction.section_search`. Its ordered field descriptors list
+eligible original section IDs; omitted sections remain protected during fine search. Separate coarse
+and fine comparison ceilings do not transfer unused capacity. The checkpoint freezes the exact
+coarse candidate and stopping state before independent per-field searches. Each fine candidate
+combines that same coarse removal with one field's cumulative section selection. It never combines
+edits across fields. The largest admitted editable-byte removal wins, with the first candidate
+retained on a tie; the usefulness threshold is reported separately from selection and confirmation.
+Fine search first tests removal of all declared eligible sections in a field. If that candidate
+is refused or rejected, search partitions that same selection into smaller removals. Protected
+sections remain untouched, and whole-record candidate order is unchanged.
+An optional frozen `require_criterion_for_confirmation` gate stops before durable selection and
+confirmation when the additional-byte criterion is unmet. Run-state validation also rejects a
+selection that violates that gate; disabling it preserves independent threshold reporting.
+The frozen `stop_when_no_retained_fields` setting additionally ends coarse search at a completed
+comparison boundary if all selected fields have been removed. It requires the confirmation gate,
+preserves the accepted coarse edit and makes no new sampling reservation.
+
+Candidate, experiment and algorithm identities bind the source, policy and traversal. Resume
+reconstructs the recorded coarse and fine decisions without executing attempts. One final durable
+selection precedes fresh confirmation against the full immutable original, including when either
+search stage has incomplete coverage. Results retain both stages and the selected mutation;
+inspection revalidates its source bindings without loading a runner or observation database.
+
+Provider-free tests cover these implementation contracts. The
+[ordinary-session measurements](measurements.md#ordinary-session-paragraph-removal) record completed
+qualification for the tested native profiles and their integration limits. The
+[workload criteria](research/reduction-design.md#text-section-workload-preparation) preserve the
+original preparation proposal separately from those measurements.
+
 ## Measurement, search and confirmation
 
 `measurement.py` invokes one prepared attempt, validates its returned evidence, evaluates the
@@ -141,13 +219,19 @@ before sampling resources are committed. An original reconstruction refusal is t
 candidate-local refusal during search consumes no provider calls. Admission starts no process or
 provider request; execution still validates the reconstructed candidate.
 
-Codex reconstructs copied history through the installed CLI's app-server protocol. A bounded
-Responses relay forwards to the fixed configured destination and records physical request counts.
-Native request and stream retries are disabled; a completed turn can still require several
-Responses requests. The caller selects a per-attempt cap from 1 to 16 (default 16). Capability,
-plan and saved-run identities retain this administrative limit. The relay rejects requests beyond
-the cap and records a `provider-call-limit` error. A completed HTTP response alone does not
-establish model acceptance or a completed native turn.
+Codex reconstructs copied history through the installed CLI's app-server protocol. A tool-output
+checkpoint includes the original output record and its captured call identity in the injected
+history, then continues with empty new input. User-message triggers remain new user input. The
+`codex-app-server-inject-checkpoint-v1` method binds this reconstruction contract; older
+reconstruction identities cannot resume under it. Compacted source prefixes remain unsupported. An
+explicit configuration-bound tool-output limit and conservative text-output admission prevent
+reliance on a model-specific truncation default. Non-text tool-output content remains unsupported. A
+bounded Responses relay forwards to the fixed configured destination and records physical request
+counts. Native request and stream retries are disabled; a completed turn can still require several
+Responses requests. The caller selects a per-attempt cap from 1 to 16 (default 16). Capability, plan
+and saved-run identities retain this administrative limit. The relay rejects requests beyond the cap
+and records a `provider-call-limit` error. A completed HTTP response alone does not establish model
+acceptance or a completed native turn.
 
 Copilot's Python SDK drives the installed CLI over stdio. It resumes the copied native prefix at
 the selected model-turn boundary without injecting a user message, including decisions after tool
@@ -212,6 +296,7 @@ Python module names are not public API commitments.
 | `fuzz/configuration.py` | Private operator policy and command limits, bound to an immutable original |
 | `fuzz/progress.py`, `fuzz/campaign.py` | Frozen mutation order, authoritative traversal recovery, durable selection and terminal campaign evidence over shared sampled state |
 | `workflow.py`, `results/` | Application phase composition and validated result models, summaries and snapshots |
+| `inspection.py`, `results/inspection.py`, `cli/inspect.py` | Offline result/case binding, retained-context interpretation and CLI presentation |
 | `cli/` | Public command selection, native setup, progress, continuation and presentation |
 | `path_safety.py`, `timestamps.py` | Shared private-path/lock primitives and timestamp validation |
 

ithrl-0.3.0/docs/branding/branding.css

@@ -0,0 +1,39 @@
+.ithrl-mark {
+  width: 2rem;
+  height: 2rem;
+  flex-shrink: 0;
+}
+
+.ithrl-mark .ink {
+  fill: #202427;
+}
+
+.ithrl-mark .accent {
+  fill: #00a58f;
+}
+
+.dark .ithrl-mark .ink {
+  fill: #fff;
+}
+
+.dark .ithrl-mark .accent {
+  fill: #24cdb5;
+}
+
+.ithrl-wordmark {
+  margin: 1rem 0 1.5rem;
+}
+
+.ithrl-wordmark img {
+  max-width: 100%;
+  height: auto;
+}
+
+.ithrl-logo-dark,
+.dark .ithrl-logo-light {
+  display: none;
+}
+
+.dark .ithrl-logo-dark {
+  display: block;
+}

ithrl-0.3.0/docs/branding/ithrl-favicon.svg

@@ -0,0 +1,16 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<svg xmlns="http://www.w3.org/2000/svg" width="512" height="512" viewBox="0 0 512 512" role="img" aria-labelledby="title description">
+<title id="title">ithrl</title>
+<desc id="description">ithrl spear; adapts to light and dark themes.</desc>
+<metadata>ithrl logo 1.0. Hand-refined spear geometry; Geist Bold outlined wordmark with ss08 i and ss03 l. See source/OFL-Geist.txt for font provenance.</metadata>
+<style>
+.ink{fill:#202427}.accent{fill:#00A58F}
+@media (prefers-color-scheme:dark){.ink{fill:#FFFFFF}.accent{fill:#24CDB5}}
+</style>
+<path id="context-upper-left" class="ink" fill="#202427" d="M136 184H286L225.62 256H75.62Z"/>
+<path id="context-upper-right" class="ink" fill="#202427" d="M356.06 214H470L412.94 282H299Z"/>
+<path id="context-lower-left-and-tail" class="ink" fill="#202427" d="M119 290H220L240 310L72.23 487L152 362H58.58Z"/>
+<path id="context-lower-right" class="ink" fill="#202427" d="M264 312H414L361.98 374H211.98Z"/>
+<path id="spear-tip" class="ink" fill="#202427" d="M266.60 227.32L294.18 250.46L456 29.65Z"/>
+<path id="revealed-diamond" class="accent" fill="#00A58F" d="M256 237L287 268L256 299L225 268Z"/>
+</svg>

ithrl-0.3.0/docs/branding/ithrl-icon.svg

@@ -0,0 +1,12 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<svg xmlns="http://www.w3.org/2000/svg" width="32" height="32" class="ithrl-mark size-8" viewBox="0 0 512 512" aria-hidden="true" focusable="false">
+<title id="title">ithrl</title>
+<desc id="description">ithrl spear.</desc>
+<metadata>ithrl logo 1.0. Hand-refined spear geometry; Geist Bold outlined wordmark with ss08 i and ss03 l. See source/OFL-Geist.txt for font provenance.</metadata>
+<path id="context-upper-left" class="ink" fill="#202427" d="M136 184H286L225.62 256H75.62Z"/>
+<path id="context-upper-right" class="ink" fill="#202427" d="M356.06 214H470L412.94 282H299Z"/>
+<path id="context-lower-left-and-tail" class="ink" fill="#202427" d="M119 290H220L240 310L72.23 487L152 362H58.58Z"/>
+<path id="context-lower-right" class="ink" fill="#202427" d="M264 312H414L361.98 374H211.98Z"/>
+<path id="spear-tip" class="ink" fill="#202427" d="M266.60 227.32L294.18 250.46L456 29.65Z"/>
+<path id="revealed-diamond" class="accent" fill="#00A58F" d="M256 237L287 268L256 299L225 268Z"/>
+</svg>

ithrl-0.3.0/docs/branding/ithrl-logo-dark.svg

@@ -0,0 +1,21 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<svg xmlns="http://www.w3.org/2000/svg" width="1040" height="512" viewBox="0 0 1040 512" role="img" aria-labelledby="title description">
+<title id="title">ithrl</title>
+<desc id="description">ithrl spear and lowercase wordmark for dark backgrounds.</desc>
+<metadata>ithrl logo 1.0. Hand-refined spear geometry; Geist Bold outlined wordmark with ss08 i and ss03 l. See source/OFL-Geist.txt for font provenance.</metadata>
+<g id="emblem">
+<path id="context-upper-left" fill="#FFFFFF" d="M136 184H286L225.62 256H75.62Z"/>
+<path id="context-upper-right" fill="#FFFFFF" d="M356.06 214H470L412.94 282H299Z"/>
+<path id="context-lower-left-and-tail" fill="#FFFFFF" d="M119 290H220L240 310L72.23 487L152 362H58.58Z"/>
+<path id="context-lower-right" fill="#FFFFFF" d="M264 312H414L361.98 374H211.98Z"/>
+<path id="spear-tip" fill="#FFFFFF" d="M266.60 227.32L294.18 250.46L456 29.65Z"/>
+<path id="revealed-diamond" fill="#24CDB5" d="M256 237L287 268L256 299L225 268Z"/>
+</g>
+<g id="wordmark" fill="#FFFFFF" aria-hidden="true">
+<path id="letter-i" d="M534 375V228.951H574.872V375ZM554.436 213.965Q545.444 213.965 539.722 208.787Q534 203.61 534 194.619Q534 185.627 539.722 180.313Q545.444 175 554.436 175Q563.7 175 569.422 180.313Q575.144 185.627 575.144 194.619Q575.144 203.61 569.422 208.787Q563.7 213.965 554.436 213.965Z"/>
+<path id="letter-t" d="M657.447 375Q634.831 375 624.204 364.51Q613.578 354.019 613.578 331.131V259.469H590.144V228.951H613.578V194.619H654.45V228.951H692.869V259.469H654.45V326.226Q654.45 336.853 658.946 340.668Q663.441 344.482 672.433 344.482H692.869V375Z"/>
+<path id="letter-h" d="M714.869 375V181.54H755.741V246.935Q761.736 236.035 772.362 230.858Q782.989 225.681 796.068 225.681Q820.591 225.681 832.989 241.213Q845.387 256.744 845.387 280.995V375H804.515V290.804Q804.515 274.455 799.338 266.281Q794.161 258.106 782.444 258.106Q769.91 258.106 762.826 266.689Q755.741 275.272 755.741 291.894V375Z"/>
+<path id="letter-r" d="M861.387 375V228.951H900.079L901.169 258.106Q905.529 242.847 913.975 235.899Q922.422 228.951 935.774 228.951H949.125V262.738H935.774Q918.608 262.738 910.433 269.687Q902.259 276.635 902.259 292.984V375Z"/>
+<path id="letter-l" d="M965.125 375V181.54H1005.997V375Z"/>
+</g>
+</svg>