correct: Say why the entry was invalid when suggesting a repair

AGENTS.md

@@ -86,11 +86,16 @@ Resolve scope questions from the changed code and the invariants first. Ask only
 if a material security decision remains unclear; identify the decision rather
 than stopping on general uncertainty. Ask before running a recommended security review.
 
-Follow `docs/developer/AI_POLICY.md` for contributions and `SECURITY.md` for
-private vulnerability reporting. Local edits, checks, and atomic commits are
-allowed; do not automatically push, open pull requests, write replies, contact
-maintainers, or claim authorship. A human publishes contributions. Never commit
-to master or push to an open pull request branch without explicit user direction.
+Before acting on a pull request, read `CONTRIBUTING.md` and
+`docs/developer/AI_POLICY.md` from its current target branch. Follow that merged
+contribution policy when the topic branch contains an older version. AI may
+write PR descriptions and review replies under the tool-neutral policy.
+
+Follow `SECURITY.md` for private vulnerability reporting. Local edits, checks,
+and atomic commits are allowed. Publish only within the user's authorized scope;
+permission to use AI does not itself authorize contacting others. Never claim
+independent human authorship, commit to master, or push to an open pull request
+branch without explicit user direction.
 
 When asked to commit, keep commits focused and independently passing unless the
 user explicitly requests a combined commit. Use an imperative subject of at

CONTRIBUTING.md

@@ -11,7 +11,7 @@ Getting Started
 New contributors are very welcome and needed.
 
 If you use AI tools while contributing, please read and follow the [AI
-policy](/doc/AI_POLICY.md).
+policy](docs/developer/AI_POLICY.md).
 
 In-depth reviewing and testing are the bottleneck of the project, and are the
 most effective way anyone can start to contribute. It will teach you much more

MANIFEST.in

@@ -1,6 +1,6 @@
 include LICENSE README.md SECURITY.md
 recursive-include LICENSES *
-recursive-include docs *.csv *.html *.json *.md
+recursive-include docs *.html *.md
 prune docs/planning
 recursive-include requirements *.txt
 recursive-include tests *.py *.json *.md

docs/developer/AI_POLICY.md

@@ -1,38 +1,34 @@
 # AI Policy
 
-Using AI (i.e. LLMs) as tools for coding is welcome, provided that such use adheres to the following policy, adds value, and does not waste the community's time.
-
-A high bar is held for all contributions to this project.
-
-**AI should not be used to generate comments when communicating with maintainers and other contributors**.
-Comments are expected to be written by humans.
-Comments that are believed to be written by AI may be moderated.
-
-If you are opening an issue, you should be able to describe the problem in your own words.
-
-You should only open a pull request if you:
-
-- know the language the code is written in
-- could have written the code yourself
-- understand the existing surrounding code and the effect of the suggested changes
-
-You must also be able to explain the pull request changes in your own words.
-This includes the pull request body and responses to questions.
-**Do not copy responses from the AI when replying to questions from reviewers.**
-
-This project requires a human author in the loop who understands the work produced by AI.
-**Pull requests should not be opened or driven by autonomous agents**.
-A human author must choose the work, understand the change, and be responsible for the contribution.
-Do not include agents as authors or co-authors of your commits for these reasons.
-Pull requests that appear in violation of this can be closed without notice.
-
-If you wish to include context from an interaction with AI in your comments, it must be disclosed as such.
-It must be accompanied by human commentary explaining the relevance and implications of the context.
-
-Questions or proposed changes to this policy can be discussed in [bitcoin-core/meta].
-
-This policy was adapted from [ripgrep's AI policy], which was adapted from [uv's AI policy].
-
-[bitcoin-core/meta]: https://github.com/bitcoin-core/meta
-[ripgrep's ai policy]: https://github.com/BurntSushi/ripgrep/blob/f0cec341ab95c25c691ad3d5754d4bd9eedde21f/AI_POLICY.md
-[uv's ai policy]: https://github.com/astral-sh/.github/blob/c5187e200db51bfe11d56e13053d29bd3793fdd8/AI_POLICY.md
+Use whatever tools help you, AI included. This project has no rules about
+which tools you may use for what.
+
+The expectations are the same for every contribution, however it was made:
+
+- Do your best.
+- If you open an issue, you should be able to describe the problem in your
+  own words.
+- If you open a pull request, you must be able to explain the changes in your
+  own words when asked.
+- A human author must choose the work, understand the change, and be
+  responsible for the contribution. Understanding means knowing what the
+  change does, why, and what it affects well enough to defend it in review,
+  not memorizing every implementation detail. If you can't defend it, don't
+  submit it.
+- Do not include agents as authors or co-authors of your commits.
+- Respect reviewers' time. Review is the bottleneck. Unverified or low-effort
+  issues, pull requests and comments may be closed without detailed review.
+
+Saying which tools you used, and how, is welcome when it helps review. It is
+not required.
+
+Changes to share generation, recovery or checksums get a security review
+however they were written.
+
+Questions or proposed changes to this policy can be raised in an issue.
+
+This policy follows Linus Torvalds' [reply to Laurent Pinchart] on 13 October
+2026 and the Linux kernel's [tool-generated content guidelines].
+
+[reply to Laurent Pinchart]: https://www.theregister.com/ai-and-ml/2026/10/14/linux-kernel-devs-drop-obligatory-confession-of-ai-tool-usage/5303140
+[tool-generated content guidelines]: https://docs.kernel.org/process/generated-content.html

docs/developer/alignment-benchmarks.md

@@ -1,42 +1,38 @@
 # Alignment correction benchmark evidence
 
-Measured 2026-09-10 with one shared **inclusive capture-mass ceiling `<=1`**.
-
-Host: **AMD Ryzen 7 7735U with Radeon Graphics**, Python **3.13.12**, `Linux-7.0.10+deb14-amd64-x86_64-with-glibc2.42`.
+Informational only: these timings help set correction limits on slower devices. They are not timing guarantees.
 
-## Execution conditions and limits
-
-The host rebooted repeatedly during this task. Accessible user journals did not establish the cause; kernel journals and persistent crash records required administrator access. Remaining cases resumed under first a 50%, then a 25% single-core CPU quota, 1 GiB memory limit, low priority and pauses between cases. The final run also waits for a CPU sensor reading at or below 75°C before each case, stopping if cooling takes over one minute. Capped rows are marked and excluded from promotion evidence. These precautions are not a diagnosed hardware fix. **No public timing guarantee is promoted from this interrupted, mixed-condition run.**
+Measured 2026-09-10 with one shared **inclusive capture-mass ceiling `<=1`**.
 
-Only complete JSON records were preserved across restarts. Cached syndrome tables persist within each process; restarted processes rebuild them. Timed calls retain the ten-second deadline. Cooldown pauses occur outside the timed call.
+Host: **AMD Ryzen 7 7735U with Radeon Graphics**, Python **3.13.12**, `Linux-7.0.10+deb14-amd64-x86_64-with-glibc2.42`. Cached syndrome tables persist within each process. Timed calls retain the ten-second deadline.
 
 ## Public required-depth measurements
 
-All 1156 cases requested `A<=2` and `G<=2`: 1152 completed, 1156 returned no candidates, 284 were capped, and 872 uncapped cases completed without candidates below eight seconds. Maximum elapsed time was **10.001 seconds**; maximum admitted mass was **0.992777219241**.
+All 1156 cases requested `A<=2` and `G<=2`: 1152 completed and none returned candidates. Maximum elapsed time was **10.001 seconds**; maximum admitted mass was **0.992777219241**.
 
-The exhaustive sweep stopped after 1120 of 1980 planned cases because the temperature gate made it impractically slow. The additional 36 representative cases cover every existing public profile length, known/unknown targets, and frozen/unfrozen five-character headers at offset zero with no explicit erasures. The partial matrix additionally exercises offsets -8/-4/-3/-2/-1/0/1/2/3/4/8 and erasures 0/1/2/4/8. The tool retains the full matrix for a stable host. Inputs use valid header syntax and deterministic random remaining symbols. Some offsets have no reachable required layer. No-candidate status is recorded rather than assumed.
+The exhaustive sweep stopped after 1120 of 1980 planned cases because it was impractically slow on this host. The additional 36 representative cases cover every existing public profile length, known/unknown targets, and frozen/unfrozen five-character headers at offset zero with no explicit erasures. The partial matrix additionally exercises offsets -8/-4/-3/-2/-1/0/1/2/3/4/8 and erasures 0/1/2/4/8. Inputs use valid header syntax and deterministic random remaining symbols. Some offsets have no reachable required layer. No-candidate status is recorded rather than assumed.
 
-| Profile | Displayed D | Full body B | Mutable M (frozen / unfrozen) | Completed | Capped | Maximum seconds |
-|---|---:|---:|---:|---:|---:|---:|
-| ms | 48 | 45 | 40 / 45 | 224/224 | 4 | 2.288 |
-| ms | 54 | 51 | 46 / 51 | 224/224 | 4 | 2.768 |
-| ms | 61 | 58 | 53 / 58 | 224/224 | 4 | 3.860 |
-| ms | 67 | 64 | 59 / 64 | 224/224 | 12 | 4.616 |
-| ms | 74 | 71 | 66 / 71 | 224/224 | 224 | 6.005 |
-| ms | 127 | 124 | 119 / 124 | 20/24 | 24 | 10.001 |
-| cl | 74 | 71 | 66 / 71 | 4/4 | 4 | 5.589 |
-| bip39_12w | 56 | 46 | 41 / 46 | 4/4 | 4 | 2.158 |
-| bip39_24w | 82 | 72 | 67 / 72 | 4/4 | 4 | 5.821 |
+| Profile | Displayed D | Full body B | Mutable M (frozen / unfrozen) | Completed | Maximum seconds |
+|---|---:|---:|---:|---:|---:|
+| ms | 48 | 45 | 40 / 45 | 224/224 | 2.288 |
+| ms | 54 | 51 | 46 / 51 | 224/224 | 2.768 |
+| ms | 61 | 58 | 53 / 58 | 224/224 | 3.860 |
+| ms | 67 | 64 | 59 / 64 | 224/224 | 4.616 |
+| ms | 74 | 71 | 66 / 71 | 224/224 | 6.005 |
+| ms | 127 | 124 | 119 / 124 | 20/24 | 10.001 |
+| cl | 74 | 71 | 66 / 71 | 4/4 | 5.589 |
+| bip39_12w | 56 | 46 | 41 / 46 | 4/4 | 2.158 |
+| bip39_24w | 82 | 72 | 67 / 72 | 4/4 | 5.821 |
 
-The implementation requires `A<=2` / `G<=2` before surfacing suggestions. Required-phase interruption suppresses suggestions. Character depths 3/4 remain best effort. Runtime under a CPU quota must not be extrapolated into an uncapped guarantee.
+The implementation requires `A<=2` / `G<=2` before surfacing suggestions. Required-phase interruption suppresses suggestions. Character depths 3/4 remain best effort.
 
 ## Deeper public search
 
-0/16 depth-3/4 probes completed within the deadline at displayed ms lengths 48 and 127, with both target-knowledge and header-freezing states. No deeper cutoff is promoted. Under the CPU quota, long-profile probes may expire in the required phase before deeper work begins.
+0/16 depth-3/4 probes completed within the deadline at displayed ms lengths 48 and 127, with both target-knowledge and header-freezing states. No deeper cutoff is promoted. Long-profile probes may expire in the required phase before deeper work begins.
 
 ## Coordinates and synthetic kernels
 
-`B = D - len(hrp + "1")`; checksum selection uses `X = len(bech32_hrp_expand(hrp)) + B`. Mutable length excludes the frozen header, while reverse indices and the syndrome retain full-body coordinates. The CSV records all eligible target body lengths and the selected checksum size. Public profile validation is unchanged.
+`B = D - len(hrp + "1")`; checksum selection uses `X = len(bech32_hrp_expand(hrp)) + B`. Mutable length excludes the frozen header, while reverse indices and the syndrome retain full-body coordinates. Public profile validation is unchanged.
 
 Synthetic kernels use **body lengths** 40/66/93/127/1015/1023 with explicit short/long period selection. They need not form valid public application strings and never promote public cutoffs. Character depths 2/3/4 sample 32 hypotheses per structural class. Group depth 2 completes the structural-generation and incremental-syndrome sphere; this is distinct from complete BCH decoding of that sphere.
 
@@ -55,16 +51,15 @@ Synthetic kernels use **body lengths** 40/66/93/127/1015/1023 with explicit shor
 | 1023 | yes | 356,746 | 5.020 |
 | 1023 | no | 648,466 | 8.897 |
 
-The JSON records generation/syndrome throughput, BCH throughput for E=0/1/2/4/8 and S=0..4 wherever E+2S<=8, root scans of degrees 1..4, sampled full-result dedup cost, retained allocation counts and traced peaks. BCH microcases exercise successful locators; public timings primarily exercise failed hypotheses. Allocations cover sampled batches, not a retained complete sphere.
+The `--kernel` output records generation/syndrome throughput, BCH throughput for E=0/1/2/4/8 and S=0..4 wherever E+2S<=8, root scans of degrees 1..4, sampled full-result dedup cost, retained allocation counts and traced peaks. BCH microcases exercise successful locators; public timings primarily exercise failed hypotheses. Allocations cover sampled batches, not a retained complete sphere.
 
 Separate public tracemalloc runs had a maximum peak of 116,160 bytes. Tracing results are neither timing evidence nor process RSS.
 
 ## Implementation and artifacts
 
 Incremental syndromes compose small piece-table segments from shifted prefixes. Full candidate bodies materialize only after BCH survival. Repeated two-erasure product-table construction was removed after profiling identified it as a dominant cost; changing erasure locations use direct field multiplication. Constant locators avoid unnecessary root scans. Search follows admitted capture-volume layers. No MITM is used.
 
-- [Public and deeper per-case results](alignment-public-benchmarks.csv).
-- [Kernel, memory and measured runtime source hashes](alignment-kernel-benchmarks.json).
+The raw kernel JSON from these runs is in git history at commit `753cf61`. Reproduce with:
 
 ```sh
 .venv/bin/python -m tools.alignment_benchmark --public --depths 2
@@ -73,12 +68,4 @@ Incremental syndromes compose small piece-table segments from shifted prefixes.
 .venv/bin/python -m tools.alignment_benchmark --public --lengths 48 --depths 2 --deltas 0 --erasures 0 --memory
 ```
 
-Commands emit JSON Lines with a host header. Capped portions of these recorded runs used systemd `CPUQuota=50%` or `CPUQuota=25%` as recorded, `MemoryMax=1G`, `CPUWeight=10`, and `nice -n 10`; each case was followed by an untimed pause equal to its elapsed time, capped at ten seconds. These settings are execution precautions, not correction policy. Full 13/15 consecutive-erasure completion at mass exactly 1 is additionally covered by API tests.
-
-## Validation
-
-After failure cleanup, normal and optimized full pytest runs each pass **714 tests**. Both runs used a 25% single-core CPU quota, 1 GiB memory limit, cooling pauses, and the 75°C temperature gate. The CLI tests now observe prefills at the editable-input boundary and expect the current diagnostic text. The approved package size budget is **4,500** nonblank/noncomment lines, retaining recursive counting; the formatted package measures **4,088**.
-
-Ruff lint and formatting, mypy (19 source files), all 57 frozen PR #70 correction cases, all 768 wallet differential records, and staged whitespace checks pass. No public timing cutoff is promoted by this validation.
-
-The kernel artifact preserves the measured source hashes in `source_sha256` and records the subsequently formatted files in `post_cleanup_source_sha256`. The failure cleanup changed test expectations, formatting, and the approved package size budget; it did not change runtime semantics.
+Commands emit JSON Lines with a host header. Full 13/15 consecutive-erasure completion at mass exactly 1 is additionally covered by API tests.