hermes-blind 0.1.3__tar.gz

hermes_blind-0.1.3/CHANGELOG.md

@@ -0,0 +1,76 @@
+# Changelog
+
+All notable changes to `hermes-blind` will be documented here.
+
+Format: [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
+Versioning: [SemVer](https://semver.org/). The 0.x line remains experimental;
+minor versions may change the public surface before 1.0.
+
+## [0.1.3] — 2026-07-19
+
+0.1.3 is the first public 0.1.x release. The public comparison is
+**0.0.6 → 0.1.3**. Versions 0.1.0 through 0.1.2 were unpublished development
+candidates.
+
+### Added since public 0.0.6
+
+- A dependency-free CLI: `hermes-blind apply`.
+- Deterministic recovery-scaffold generation from Claude Code and Codex JSONL.
+- `goals`, `first-sentence`, and `full` anchor modes.
+- Intent-debias and target-scope preambles used by hermes-rubric.
+- `placebo` and `gate-only` mechanism-isolation variants.
+
+### Changed from the unpublished 0.1.2 candidate
+
+- Goal-set extraction replaces first-sentence truncation as the default.
+- Recovery output records the source filename instead of its absolute path.
+- Recovery output refuses to overwrite its input and preserves existing output
+  files unless `--force` is explicit.
+- Public docs and metadata now separate tested mechanics from unproven
+  behavioral efficacy.
+- Added a privacy-safe anchor-fidelity report with sanitized per-session
+  counts, statistical context, receipt commitments, and public synthetic
+  mechanics fixtures; the report and sanitized data ship in the source
+  distribution.
+- The public CLI excludes unfinished experiment and analysis harnesses tied to
+  sibling repositories.
+- Removed the staged placeholder seal; it is not a release requirement.
+
+### Evidence boundary
+
+- A frozen 66-goal audit represented 40 pre-listed goals with goal-set
+  extraction, versus 7 with the previous first-sentence heuristic: a 50.0
+  percentage-point increase, with improvement in 7 of 9 sessions and ties in
+  2. This supports substantially better mission representation in the
+  recovery artifact.
+- Bias reduction, downstream model adherence or task outcomes, automatic
+  drift detection, and optimal timing remain unproven.
+
+## [0.0.6] — 2026-04-24
+
+### Added
+- Initial release. Five scaffold variants (`null`, `micro`, `short`, `v1`, `full`)
+  exposed via `wrap(prompt, variant)`.
+- `extract_disclosure(response)` for parsing the model's disclosure line from
+  a completed response.
+- `token_estimate(variant)` rough token count per variant.
+- 19 unit tests covering: variant correctness, caller-prompt preservation,
+  length-ordering invariant, disclosure extraction (seven shapes incl. edge
+  cases), scaffold structural invariants.
+- `INTENT.md`, `PLAN-v1.md`, `PLAN-v2.md` — design trail.
+- Three hermes-rubric audit runs on the plan under `rubric-runs/`:
+  phase1-claude.json (5.4/10 on v1), phase1-v2.json (5.7/10 after iteration),
+  phase3-adversarial.json (5.5/10 — low aggregate = useful adversarial
+  signal; three real gaps surfaced: self-referential evaluation loop,
+  length-vs-content confound, backend-generalization runtime check).
+
+### Known not-yet-validated
+- Empirical variance-reduction effect of the scaffold. Phase 4 ablation
+  study not yet executed. Do not treat this release as a proven debiaser.
+- Cross-model convergence across Opus / Sonnet / Haiku / Ollama qwen3.5.
+- Behavior on long targets (>10k tokens) or multi-turn scoring.
+
+### Not in this release
+- CLI (library-only for v0.0.x)
+- Integration into `hermes-rubric` backends (gated on Phase 4 pass)
+- Non-English scaffold variants

hermes_blind-0.1.3/EVALUATION.md

@@ -0,0 +1,125 @@
+# Recovery-anchor fidelity evaluation
+
+This is the public, privacy-safe report for the anchor-fidelity evidence cited
+by hermes-blind 0.1.3. It covers one question only:
+
+> When a long first user turn contains several goals, does the default
+> `goals` recovery anchor represent more of those goals than the previous
+> first-sentence anchor?
+
+It does **not** evaluate the prompt-wrapping primitive, automatic drift
+detection, downstream model adherence, or final task outcomes.
+
+## What “recovery” means
+
+Recovery is the package's second main surface, not a third mode. Given a
+Claude Code or Codex JSONL session, `hermes-blind apply --session ...` reads
+the first user turn and creates a markdown anchor that can be reviewed and
+reinserted into a long session.
+
+That recovery surface has three anchor modes:
+
+- `goals` (default): extracts up to 12 goal-carrying sentences;
+- `first-sentence`: retains the legacy compact anchor; and
+- `full`: includes up to 4,000 characters of the first turn.
+
+The package does not decide when to inject the anchor, inject it
+automatically, or make a model call.
+
+## Method
+
+The design was frozen before corpus mining. Fresh-context model reviewers
+were separated by role:
+
+1. one reviewer saw only each turn-one mission seed and pre-listed its goals;
+2. hermes-blind generated a legacy `first-sentence` anchor and a `goals`
+   anchor for the same seed; and
+3. a separate Haiku reviewer marked each pre-listed goal `REPRESENTED` or
+   `ABSENT` in each anchor.
+
+The evaluated extraction implementation was commit
+`ca433597e8b3683d38dcd656fa7a35ad1c985114`.
+
+The frozen plan called for 10 qualifying sessions. The completed receipt
+contains 9 sessions and 66 goals, and does not record why the tenth session
+was absent. The figures below use the observed denominator; no missing result
+was imputed.
+
+## Results
+
+| Anchor | Goals represented | Coverage | Wilson 95% CI |
+|---|---:|---:|---:|
+| Legacy `first-sentence` | 7 / 66 | 10.6% | 5.2–20.3% |
+| Default `goals` | 40 / 66 | 60.6% | 48.5–71.5% |
+
+The default anchor represented 33 additional goals: a **50.0 percentage-point
+increase** and **5.7×** the legacy coverage. At the paired goal level, all 7
+goals represented by the legacy anchor remained represented; 33 previously
+absent goals became represented; and none regressed.
+
+Across sessions, `goals` improved coverage in 7 of 9 sessions and tied in 2.
+An exploratory two-sided exact sign test over the 7 non-tied sessions gives
+`p = 0.015625`. This test was calculated after the frozen evaluation rather
+than pre-registered, and the sessions come from one private operating
+environment, so it should be read as corroborating evidence rather than a
+population-wide effect estimate.
+
+The sanitized per-session counts and exact calculations are in
+[`evaluation/anchor-fidelity-20260718/results.json`](evaluation/anchor-fidelity-20260718/results.json).
+
+## What this establishes
+
+For these nine long mission seeds, the 0.1.3 `goals` extractor preserved
+substantially more of the original mission in the generated recovery artifact
+than the legacy first-sentence heuristic. Mission representation is a
+necessary first step for recovery: a goal absent from the artifact cannot be
+restored by that artifact.
+
+## What this does not establish
+
+This evaluation does not show that reinserting the artifact makes a model
+follow the mission, improves the final answer, prevents drift, or identifies
+the right intervention time. Those are separate links in the causal chain and
+need separate evaluations. The goal listing and representation judgments were
+model-produced, not human-adjudicated, and the private corpus is too small and
+too concentrated to support a general claim across users, languages, or
+session types.
+
+## Privacy and auditability
+
+The source material consists of real private session openings that may contain
+personal or unpublished project context. Raw text, local paths, complete
+session identifiers, and individual judgments are therefore not published.
+The public JSON contains only relabeled per-session counts, aggregate results,
+method metadata, and SHA-256 commitments to the frozen internal receipts.
+
+This protects the operator's private context, but it also means an outside
+reader cannot independently audit the goal labels from this packet alone. A
+future public or consented corpus is required for full external replication.
+
+## Reproduce the public mechanics
+
+The repository's synthetic fixtures are small, fabricated Claude Code and
+Codex JSONL records—not sanitized private conversations. They verify the
+reproducible software mechanics: parsing both log shapes, extracting multiple
+goals, preserving a later goal after context-only prose, using safe markdown
+fences, refusing unsafe overwrites, omitting parent paths from errors, and
+deduplicating Codex's paired event shapes.
+
+Run them with:
+
+```bash
+pytest -q tests/test_apply.py
+```
+
+These fixtures reproduce the mechanism and regression boundary. They do not
+reproduce the private 40/66 behavioral audit.
+
+## Receipt commitments
+
+| Internal receipt | SHA-256 |
+|---|---|
+| Frozen specification | `05b4afc471eb47a15212fd3dbf225321ca61eeb698252b247594890ef3956555` |
+| Source results | `376a5fbd69accb644171bac0f0363cca09e93d54f298f1adcc20ee4b9ca6fbcf` |
+| Results report | `4774264dd9b8ea16b13fb63a2a81a73c240b68a9783b264c586e104173e92d2a` |
+| Private corpus commitment file | `a241c04d94c572ae109e2e99f2e22e1c9cf3ae1df42c649010882685b8820c71` |

hermes_blind-0.1.3/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Hermes Labs
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

hermes_blind-0.1.3/MANIFEST.in

@@ -0,0 +1,3 @@
+include CHANGELOG.md
+include EVALUATION.md
+recursive-include evaluation *.json

hermes_blind-0.1.3/PKG-INFO

@@ -0,0 +1,178 @@
+Metadata-Version: 2.4
+Name: hermes-blind
+Version: 0.1.3
+Summary: Deterministic prompt and session-recovery scaffolds for LLM workflows
+Author-email: Hermes Labs <roli@hermes-labs.ai>
+License-Expression: MIT
+Project-URL: Source, https://github.com/hermes-labs-ai/hermes-blind
+Project-URL: Issues, https://github.com/hermes-labs-ai/hermes-blind/issues
+Project-URL: Changelog, https://github.com/hermes-labs-ai/hermes-blind/blob/main/CHANGELOG.md
+Project-URL: Evaluation, https://github.com/hermes-labs-ai/hermes-blind/blob/main/EVALUATION.md
+Keywords: llm,prompt-engineering,session-recovery,scaffold,evaluation
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Developers
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Provides-Extra: dev
+Requires-Dist: build>=1.2; extra == "dev"
+Requires-Dist: pytest>=7.0; extra == "dev"
+Requires-Dist: ruff>=0.4; extra == "dev"
+Requires-Dist: twine>=5.0; extra == "dev"
+Dynamic: license-file
+
+# hermes-blind
+
+**Deterministic prompt and session-recovery scaffolds for LLM workflows.**
+
+[![PyPI](https://img.shields.io/pypi/v/hermes-blind.svg)](https://pypi.org/project/hermes-blind/)
+[![Python](https://img.shields.io/pypi/pyversions/hermes-blind.svg)](https://pypi.org/project/hermes-blind/)
+[![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](LICENSE)
+[![Status: experimental](https://img.shields.io/badge/status-experimental-orange.svg)](#evidence-and-limits)
+
+hermes-blind provides two small, standard-library-only primitives:
+
+1. prepend an evidence-gating scaffold to an evaluation prompt; and
+2. recover a compact turn-one goal anchor from a Claude Code or Codex session log.
+
+It makes no model calls, sends no network requests, and does not claim to
+detect drift automatically.
+
+## Install
+
+```bash
+pip install hermes-blind
+```
+
+Python 3.10+.
+
+## Recover a long session
+
+```bash
+hermes-blind apply \
+  --session /path/to/rollout.jsonl \
+  --format auto \
+  --anchor-mode goals \
+  --turn 9 \
+  --out recovery.md
+```
+
+The default `goals` mode scans the first user turn for goal-carrying
+sentences and preserves up to 12 of them. Use `first-sentence` for the
+legacy compact behavior or `full` to include up to 4,000 characters.
+`--format auto` recognizes Claude Code and Codex JSONL shapes.
+
+The `--turn` value is output metadata. It is not an automatic trigger and
+does not imply that turn 9 is an empirically optimal intervention point.
+Existing output files are preserved unless `--force` is passed; the input
+session file can never be used as the output path.
+
+Recovery output includes user-authored text. It records only the session
+filename, not its absolute path, but you should still inspect the markdown
+before sharing it.
+
+## Wrap an evaluation prompt
+
+```python
+from hermes_blind import wrap
+
+prompt = wrap(
+    "Rate this paper on novelty from 0 to 10 and cite the target text.",
+    variant="v1",
+)
+```
+
+Or from the CLI:
+
+```bash
+hermes-blind apply --variant v1 --prompt "Score this artifact from quoted evidence."
+```
+
+Available variants are `null`, `micro`, `short`, `v1`, `full`,
+`placebo`, and `gate-only`. The `null` variant is an exact no-op for
+controlled comparisons.
+
+## Rubric framing helpers
+
+The package also exposes the dependency-free intent and scope preambles used
+by hermes-rubric:
+
+```python
+from hermes_blind import compose_intent
+
+framed = compose_intent(
+    "Evaluate whether this release is ready.",
+    scope_class="results-bundle",
+    intent_debias=True,
+)
+```
+
+## Evidence and limits
+
+Validated mechanics for 0.1.3:
+
+- deterministic prompt wrapping and disclosure parsing;
+- Claude Code and Codex JSONL parsing;
+- three recovery anchor modes;
+- prompt preservation and scaffold-family invariants;
+- package build, clean installation, CLI invocation, and unit tests.
+
+A frozen 66-goal extraction audit found that goal-set extraction represented
+40 of 66 pre-listed goals, compared with 7 of 66 for the previous
+first-sentence heuristic: a 50.0 percentage-point increase, with improvement
+in 7 of 9 sessions and ties in 2. This demonstrates substantially better
+**mission representation in the generated recovery artifact**, a necessary
+first step for recovery. It does not establish that reinserting the artifact
+causes downstream model adherence or better task outcomes. See the
+[privacy-safe evaluation report](https://github.com/hermes-labs-ai/hermes-blind/blob/main/EVALUATION.md)
+for the method, sanitized
+per-session results, statistical context, limitations, and receipt hashes.
+
+Public synthetic fixtures in `tests/test_apply.py` reproduce the parsing,
+extraction, safety, and output mechanics with fabricated Claude Code and Codex
+JSONL. They do not reproduce the private 40/66 audit.
+
+Not established:
+
+- reliable bias reduction from the evaluation prefix;
+- successful behavioral recovery after inserting a generated anchor;
+- automatic drift detection or an optimal intervention turn;
+- adversarial prompt-injection resistance;
+- non-English behavior.
+
+Earlier single-shot experiments did not establish the original bias-reduction
+hypothesis. A later multi-turn research harness remains internal and is not
+part of the public runtime because its efficacy study is incomplete.
+
+Treat the output as a transparent scaffold for a human or agent to inspect,
+not as a security boundary or independent evaluator.
+
+## Development
+
+```bash
+python -m venv .venv
+source .venv/bin/activate
+pip install -e ".[dev]"
+ruff check src tests
+pytest -q
+python -m build
+twine check dist/*
+```
+
+## Version note
+
+0.1.3 is the first public 0.1.x release. The public comparison is therefore
+**0.0.6 → 0.1.3**. Versions 0.1.0 through 0.1.2 were internal development
+candidates; 0.1.3 is a patch over the unpublished 0.1.2 candidate that adds
+multi-goal extraction and aligns the public package surface.
+
+## License
+
+MIT. See [LICENSE](LICENSE).
+
+Part of [Hermes Labs](https://hermes-labs.ai).

hermes_blind-0.1.3/README.md

@@ -0,0 +1,150 @@
+# hermes-blind
+
+**Deterministic prompt and session-recovery scaffolds for LLM workflows.**
+
+[![PyPI](https://img.shields.io/pypi/v/hermes-blind.svg)](https://pypi.org/project/hermes-blind/)
+[![Python](https://img.shields.io/pypi/pyversions/hermes-blind.svg)](https://pypi.org/project/hermes-blind/)
+[![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](LICENSE)
+[![Status: experimental](https://img.shields.io/badge/status-experimental-orange.svg)](#evidence-and-limits)
+
+hermes-blind provides two small, standard-library-only primitives:
+
+1. prepend an evidence-gating scaffold to an evaluation prompt; and
+2. recover a compact turn-one goal anchor from a Claude Code or Codex session log.
+
+It makes no model calls, sends no network requests, and does not claim to
+detect drift automatically.
+
+## Install
+
+```bash
+pip install hermes-blind
+```
+
+Python 3.10+.
+
+## Recover a long session
+
+```bash
+hermes-blind apply \
+  --session /path/to/rollout.jsonl \
+  --format auto \
+  --anchor-mode goals \
+  --turn 9 \
+  --out recovery.md
+```
+
+The default `goals` mode scans the first user turn for goal-carrying
+sentences and preserves up to 12 of them. Use `first-sentence` for the
+legacy compact behavior or `full` to include up to 4,000 characters.
+`--format auto` recognizes Claude Code and Codex JSONL shapes.
+
+The `--turn` value is output metadata. It is not an automatic trigger and
+does not imply that turn 9 is an empirically optimal intervention point.
+Existing output files are preserved unless `--force` is passed; the input
+session file can never be used as the output path.
+
+Recovery output includes user-authored text. It records only the session
+filename, not its absolute path, but you should still inspect the markdown
+before sharing it.
+
+## Wrap an evaluation prompt
+
+```python
+from hermes_blind import wrap
+
+prompt = wrap(
+    "Rate this paper on novelty from 0 to 10 and cite the target text.",
+    variant="v1",
+)
+```
+
+Or from the CLI:
+
+```bash
+hermes-blind apply --variant v1 --prompt "Score this artifact from quoted evidence."
+```
+
+Available variants are `null`, `micro`, `short`, `v1`, `full`,
+`placebo`, and `gate-only`. The `null` variant is an exact no-op for
+controlled comparisons.
+
+## Rubric framing helpers
+
+The package also exposes the dependency-free intent and scope preambles used
+by hermes-rubric:
+
+```python
+from hermes_blind import compose_intent
+
+framed = compose_intent(
+    "Evaluate whether this release is ready.",
+    scope_class="results-bundle",
+    intent_debias=True,
+)
+```
+
+## Evidence and limits
+
+Validated mechanics for 0.1.3:
+
+- deterministic prompt wrapping and disclosure parsing;
+- Claude Code and Codex JSONL parsing;
+- three recovery anchor modes;
+- prompt preservation and scaffold-family invariants;
+- package build, clean installation, CLI invocation, and unit tests.
+
+A frozen 66-goal extraction audit found that goal-set extraction represented
+40 of 66 pre-listed goals, compared with 7 of 66 for the previous
+first-sentence heuristic: a 50.0 percentage-point increase, with improvement
+in 7 of 9 sessions and ties in 2. This demonstrates substantially better
+**mission representation in the generated recovery artifact**, a necessary
+first step for recovery. It does not establish that reinserting the artifact
+causes downstream model adherence or better task outcomes. See the
+[privacy-safe evaluation report](https://github.com/hermes-labs-ai/hermes-blind/blob/main/EVALUATION.md)
+for the method, sanitized
+per-session results, statistical context, limitations, and receipt hashes.
+
+Public synthetic fixtures in `tests/test_apply.py` reproduce the parsing,
+extraction, safety, and output mechanics with fabricated Claude Code and Codex
+JSONL. They do not reproduce the private 40/66 audit.
+
+Not established:
+
+- reliable bias reduction from the evaluation prefix;
+- successful behavioral recovery after inserting a generated anchor;
+- automatic drift detection or an optimal intervention turn;
+- adversarial prompt-injection resistance;
+- non-English behavior.
+
+Earlier single-shot experiments did not establish the original bias-reduction
+hypothesis. A later multi-turn research harness remains internal and is not
+part of the public runtime because its efficacy study is incomplete.
+
+Treat the output as a transparent scaffold for a human or agent to inspect,
+not as a security boundary or independent evaluator.
+
+## Development
+
+```bash
+python -m venv .venv
+source .venv/bin/activate
+pip install -e ".[dev]"
+ruff check src tests
+pytest -q
+python -m build
+twine check dist/*
+```
+
+## Version note
+
+0.1.3 is the first public 0.1.x release. The public comparison is therefore
+**0.0.6 → 0.1.3**. Versions 0.1.0 through 0.1.2 were internal development
+candidates; 0.1.3 is a patch over the unpublished 0.1.2 candidate that adds
+multi-goal extraction and aligns the public package surface.
+
+## License
+
+MIT. See [LICENSE](LICENSE).
+
+Part of [Hermes Labs](https://hermes-labs.ai).

hermes_blind-0.1.3/evaluation/anchor-fidelity-20260718/results.json

@@ -0,0 +1,83 @@
+{
+  "evaluation_id": "anchor-fidelity-20260718",
+  "release": "0.1.3",
+  "question": "Does goals-mode represent more pre-listed turn-one goals than the legacy first-sentence anchor?",
+  "scope": "recovery-artifact mission representation only",
+  "evaluated_commit": "ca433597e8b3683d38dcd656fa7a35ad1c985114",
+  "method": {
+    "design_frozen_before_corpus_mining": true,
+    "planned_sessions": 10,
+    "observed_sessions": 9,
+    "protocol_deviation": "The completed receipt contains nine sessions and does not record why the tenth planned session was absent.",
+    "goals": 66,
+    "goal_listing": "A fresh-context model reviewer saw only each turn-one seed and pre-listed its distinct goals.",
+    "representation_check": "A separate Haiku reviewer marked every pre-listed goal REPRESENTED or ABSENT in both anchors.",
+    "human_adjudication": false
+  },
+  "privacy": {
+    "source": "private real-world session openings from one operating environment",
+    "raw_text_published": false,
+    "local_paths_published": false,
+    "complete_session_ids_published": false,
+    "individual_goal_judgments_published": false,
+    "external_label_audit_possible_from_public_packet": false
+  },
+  "per_session": [
+    {"session": "S01", "goals": 12, "legacy_represented": 0, "goals_mode_represented": 7},
+    {"session": "S02", "goals": 10, "legacy_represented": 0, "goals_mode_represented": 4},
+    {"session": "S03", "goals": 3, "legacy_represented": 1, "goals_mode_represented": 1},
+    {"session": "S04", "goals": 6, "legacy_represented": 1, "goals_mode_represented": 5},
+    {"session": "S05", "goals": 4, "legacy_represented": 3, "goals_mode_represented": 3},
+    {"session": "S06", "goals": 12, "legacy_represented": 0, "goals_mode_represented": 5},
+    {"session": "S07", "goals": 5, "legacy_represented": 2, "goals_mode_represented": 4},
+    {"session": "S08", "goals": 12, "legacy_represented": 0, "goals_mode_represented": 9},
+    {"session": "S09", "goals": 2, "legacy_represented": 0, "goals_mode_represented": 2}
+  ],
+  "pooled": {
+    "legacy": {
+      "represented": 7,
+      "total": 66,
+      "coverage": 0.106,
+      "wilson_95_ci": [0.052, 0.203]
+    },
+    "goals_mode": {
+      "represented": 40,
+      "total": 66,
+      "coverage": 0.606,
+      "wilson_95_ci": [0.485, 0.715]
+    },
+    "additional_goals_represented": 33,
+    "absolute_coverage_increase": 0.5,
+    "coverage_ratio": 5.714285714285714,
+    "paired_goal_transitions": {
+      "absent_to_absent": 26,
+      "absent_to_represented": 33,
+      "represented_to_absent": 0,
+      "represented_to_represented": 7
+    }
+  },
+  "session_level": {
+    "improved": 7,
+    "tied": 2,
+    "regressed": 0,
+    "exploratory_two_sided_exact_sign_test": {
+      "non_tied_sessions": 7,
+      "positive_differences": 7,
+      "p_value": 0.015625,
+      "pre_registered": false
+    }
+  },
+  "limits": [
+    "Does not evaluate downstream model adherence or task outcomes.",
+    "Does not evaluate automatic drift detection or intervention timing.",
+    "Model-produced goal labels and representation checks were not human-adjudicated.",
+    "The private corpus is not included, so external readers cannot audit labels from this packet alone.",
+    "Nine sessions from one operating environment do not support population-wide generalization."
+  ],
+  "source_commitments": {
+    "frozen_spec_sha256": "05b4afc471eb47a15212fd3dbf225321ca61eeb698252b247594890ef3956555",
+    "source_e3_results_sha256": "376a5fbd69accb644171bac0f0363cca09e93d54f298f1adcc20ee4b9ca6fbcf",
+    "source_results_report_sha256": "4774264dd9b8ea16b13fb63a2a81a73c240b68a9783b264c586e104173e92d2a",
+    "private_corpus_commitment_file_sha256": "a241c04d94c572ae109e2e99f2e22e1c9cf3ae1df42c649010882685b8820c71"
+  }
+}