outerloop-science 0.3.0rc2__tar.gz → 0.4.0__tar.gz

{outerloop_science-0.3.0rc2 → outerloop_science-0.4.0}/.gitignore

@@ -5,6 +5,10 @@ __pycache__/
 dist/
 build/
 
+# Docs site (scripts/docs/build.sh; dist/ is the Python build)
+dist-docs/
+site/
+
 # Environments
 .venv/
 
@@ -12,6 +16,9 @@ build/
 .pytest_cache/
 .mypy_cache/
 .ruff_cache/
+.coverage
+.coverage.*
+htmlcov/
 
 # Agent run artifacts (transcripts contain target-repo code — never commit)
 runs/

{outerloop_science-0.3.0rc2 → outerloop_science-0.4.0}/.pre-commit-config.yaml

@@ -5,6 +5,9 @@ repos:
       - id: check-added-large-files
         args: [--maxkb=500]
       - id: end-of-file-fixer
+        # byte-exact archived-writer fixture: save_record must round-trip it,
+        # and that writer emitted no trailing newline
+        exclude: ^tests/fixtures/author_route_legacy\.json$
       - id: trailing-whitespace
       - id: check-merge-conflict
       - id: check-yaml

{outerloop_science-0.3.0rc2 → outerloop_science-0.4.0}/CHANGELOG.md

@@ -6,6 +6,381 @@ Versions follow [SemVer](https://semver.org).
 
 ## [Unreleased]
 
+## [0.4.0] - 2026-10-20
+
+### Upgrading
+
+Operator actions before upgrading (everything else needs no action; details
+in each entry):
+
+- Rename `autoresearch:*` labels on target repos to `outerloop:*` and update
+  workflow `if:` conditions that match them; old labels and comment markers
+  are no longer recognized.
+- Change an open research-log issue's `<!-- autoresearch:research-log -->`
+  marker to `<!-- outerloop:research-log -->`, and finish, close, or re-mark
+  issues and PRs that carry other `autoresearch:` markers; otherwise the
+  kernel may re-post a review, re-pick a claimed issue, or open a duplicate.
+- Rename the contract key `climb_job_minutes` to `attempt_job_minutes`, and
+  finish or rebind runs parked on a commit whose contract still uses it.
+- Finish or rebind parked runs whose benchmark was renamed or removed.
+- Run `outerloop harness upgrade codex` (deploys that run harness upgrades do
+  this automatically).
+- The one-idea rule is no longer enforced; a repository that wants it writes
+  it into its guidance files.
+
+What changes for runs in flight: a parked run may get one tool-update note at
+its first wake; a measurement between infrastructure retries may get one extra
+retry; codex costs captured before the upgrade still overstate resumed
+sessions.
+
+Before rolling back: remove `regression` blocks from contracts, drain parked
+runs, and do not roll back while self-reported claims exist.
+
+### Documentation
+
+- The docs site is now a manual: getting started, task guides, concept pages
+  with figures, and reference pages for contract fields, the operator CLI,
+  agent syscalls, inbox messages and environment variables.
+- Tests, comments and design notes use generic cluster, partition and
+  hardware names (`gpu-large`, `cpu-short`, "cluster A") instead of site-specific ones.
+  The docs site drops its footer line.
+
+### Changed
+
+- Harness upgrades kill an install only after `OUTERLOOP_HARNESS_TIMEOUT_SECONDS`
+  without installer output, with a one-hour overall backstop, so a slow venv
+  build completes instead of dying every tick. The installers now print while
+  they download and unpack (uv runs verbosely, curl shows progress bars, git
+  clone reports progress), so a healthy install never looks stalled. A busy
+  upgrade lock now says another upgrade is running rather than `[Errno 11]`.
+  Upgrading: the variable keeps its name and default but now means time since
+  the last output, so an install that keeps printing may run for up to an hour.
+
+- The author verb `end` is renamed `stop` (`stop --report FILE`,
+  `stop --withdraw REASON`), with no alias: agents read `end` as "end this
+  turn" and called it when they meant to wait for results, so the refusals
+  it hit now each name the next step (`sleep`, `cancel`, `stop`). Upgrading:
+  a staged request an older kernel's CLI wrote with `"type": "end"` is
+  still honoured; the verb is now `stop`.
+
+- The git safety checks now run once at each hand-off between a session and
+  kernel code (session start and return, wake start, panel read) instead
+  of on every kernel git call. Kernel git calls after a hand-off reuse its
+  filter-driver listing; launching untrusted code on the workspace drops
+  it. A workspace no hand-off covers keeps the per-call check (#512).
+
+- The advisory-review examples and install guide track the `v0` release tag
+  (`@v0` with `reviewer_ref: v0`) instead of `main` or an exact release. `v0`
+  moves only when a final release is cut (a new `RELEASING.md` step), so
+  callers follow releases without hand bumps and never run the moving branch
+  with their reviewer key (#494).
+
+- Research taste moved out of the kernel and into a configurable GUIDANCE
+  layer. Kernel agent-facing text (the author brief, the panel claim, the
+  verifier rubric) now carries only mechanics; a new top-level contract key
+  `guidance` selects the pack: `default` (the moved texts minus the
+  one-idea rule), `none`, or a repo-relative directory holding `author.md`
+  and `panel.md` (recommended location `.github/outerloop/guidance/`).
+  Guidance is read via git objects from the same ref as the contract
+  (never the working tree; a symlink blob is refused), capped at 8,000
+  characters per file, carried last in the author brief and in the panel
+  judges' prompt — outside the fence that holds the claim as untrusted
+  data — and labelled advisory; it may never point into `.outerloop/`.
+  The panel now blocks only on validity — harness
+  exploitation, ruler/seed fishing, data leakage, a confirmed claim-vs-evidence
+  contradiction, and scope and measurement facts; aggregation and landscape
+  live only in the pack (so `none` leaves validity criteria only), and the
+  taste halves of the overfitting/unsupported-claim/measurement-gap rubric
+  are advisory-only. Upgrading: the one-idea rule is no longer enforced
+  anywhere — a repository that wants it writes it into its guidance files
+  (the panel will not block on it); the default hypothesis, done criteria,
+  PR body and brief wording changed accordingly, and stored briefs from
+  earlier versions still load unchanged.
+- A second pass moved the remaining research advice out of the kernel brief
+  into the default guidance pack (when to check the sibling view, budgeting
+  experiments against the final eval, running the repo's checks before a
+  submit) and deleted the divergence-debt aside from the research-line
+  section; under `guidance: none` the brief now states mechanics only.
+### Added
+
+- Operator `outerloop message <run-id> (<text> | --file <path>)` to append one
+  note to a run's inbox, delivered at the run's next wake — no PR needed.
+  Reuses the existing `note`/`kernel` message shape, so older kernels read it
+  without a compatibility change. Upgrading: no action needed; it appends an
+  ordinary note from the kernel; a parked run that is not waiting on jobs wakes
+  at the next tick to read it, and one waiting on jobs reads it when they end.
+- `OUTERLOOP_UNCONTAINED=1` runs a cluster deployment without a container
+  runtime: sessions and evaluations run as Slurm jobs the way local mode runs
+  them without an image, the image is ignored, and the panel stays opt-in
+  (`OUTERLOOP_PANEL_UNCONTAINED=1`, now also carried by the chain's deploy
+  step). Without it, a cluster deployment with no image still services nothing,
+  and its log now names the setting. `outerloop start` now checks the fleet
+  default author with the tick's own author check, as it already did for
+  overrides, reading settings with the precedence the launched tick applies,
+  so a default the tick would refuse (a Codex or Hermes author without an
+  image, a Codex author without a Codex model) fails at start instead of
+  idling every tick.
+- Multi-node benchmarks: a benchmark's `nodes` (default 1, up to 64) spreads
+  every dispatched job over that many nodes, with `gpus` per node. The eval
+  job `srun`s one task per node; each runs the command in its own checkout
+  with `NNODES`, `NODE_RANK`, `MASTER_ADDR`, `MASTER_PORT` and torchrun's
+  `PET_*` defaults set, node 0's output is the measurement, and a failing node
+  stops the rest. A requeued job clears the earlier attempt's per-node results
+  before it starts, and those rendezvous names (and any `PET_*`) are refused as
+  a `seed_env`. GPU-hour charges, budget checks, sweep pacing and the
+  operator's GPU ceiling count `nodes x gpus`. Uncontained Slurm deployments
+  only: a contained or local deployment refuses to launch attempts on such a
+  contract. Upgrading: one-node benchmarks keep their measurement signatures,
+  eval slots and cached baselines unchanged; cached baselines now record
+  their node count, and entries without one read as one node. A contract
+  that sets `nodes` fails to load on older kernels (unknown keys are
+  refused), so upgrade every fleet that services a target before adding it.
+- `OUTERLOOP_OPERATOR_HARNESSES` lists harness CLIs the operator supplies, such
+  as a site-managed `claude` that authenticates by itself: `outerloop harness
+  upgrade` (and the deployment's automatic `--used` run) never replaces them or
+  rewrites their recorded paths, and `outerloop harness status` reports
+  `OPERATOR` instead of drift. Unknown names are refused. The chain's deploy
+  step re-reads it from `.env` every tick.
+
+### Documentation
+
+- Document `OUTERLOOP_CODEX_WEB_SEARCH` and the steward and panel judge key
+  files (`OUTERLOOP_STEWARD_KEY_FILE`, `OUTERLOOP_PANEL_KEY_FILE`,
+  `OUTERLOOP_PANEL_CODEX_KEY_FILE`, `OUTERLOOP_PANEL_HERMES_KEY_FILE`) in the
+  install guide, and mark the verification-loop and dispatcher design notes
+  as built (#337).
+- `docs/design/compute-cloud.md` (#348): Slurm on a cloud as the 0.4
+  deployment, with operator guidance and the kernel gaps it exposed; clouds
+  without Slurm (git in, artifact out; a meta-provider; the run-artifact
+  store) as the long-term direction.
+- The pages in `docs/` build into a MkDocs Material site (`mkdocs.yml`,
+  `scripts/docs/build.sh`), with the README and project pages as site pages
+  and `noindex` staging until the public launch (#523).
+
+### Changed
+
+- Refreshed locked dependencies to current releases admitted by the existing
+  bounds: cryptography 50.0.1→50.0.2, pydantic 2.13.5→2.14.0 (pydantic-core
+  2.46.5→2.50.0), mypy 2.3.1→2.4.0, ruff 0.16.6→0.17.0, and transitives;
+  workflows pin `astral-sh/setup-uv@v10.2.0` (#337).
+- The review and verify posting CLIs now share one gate for the untrusted
+  emitted envelope, so a security fix to the decode/identity/kind checks lands
+  in both (#337). Upgrading: no action needed; envelopes already emitted keep
+  their shape and are read the same way, so a round started before the upgrade
+  posts normally.
+- The `.git` guard's object read no longer derives filter-driver overrides
+  (`cat-file -e` never converts working-tree files), dropping one
+  `git config --get-regexp` subprocess from every guarded kernel git call
+  with no change to what is refused or neutralized (#512).
+- A review round refused because the pinned harness is too old for the model,
+  or because the runner's harness install is missing or stale, now says which
+  and how to fix it instead of calling it an API outage (#475). Upgrading: no
+  action needed; the next skipped round names the cause, and skip comments
+  already posted keep their text.
+
+### Fixed
+
+- A wake that upgrades the author's syscall tool now queues the "tool was
+  updated" note once per tool version, keyed by the installed tool's content
+  digest. The kernel records that digest in the run record when it installs or
+  refreshes the tool (a fresh climb records it without a note), so a wake under
+  the same kernel has nothing to announce. A wake whose recorded digest differs
+  appends the note before saving the new digest: an interruption between the
+  two re-sends at most once, and the inbox's key dedup prevents a duplicate.
+  Upgrading: an old `tool:` inbox key written by an earlier kernel no longer
+  suppresses the new note, so a parked run may see one update note at its first
+  wake after upgrading.
+- Reserve infrastructure retries before submission so a crash before recording
+  the job id cannot exceed the two-retry cap; undo the reservation only on a
+  capacity rejection before submission. Upgrading: a measure that was between
+  infrastructure retries when the kernel was upgraded may get one extra retry,
+  once, because older journals counted failures rather than reserved retries.
+- A dispatched baseline or cleanly completed eval that ends with no result is
+  now an infrastructure failure: two uncharged retries, then a park naming the
+  `infra-retries.json` an operator deletes to resume (#474). Upgrading: a
+  missing journal or `stage.wait_note` reads as zero retries / no note.
+- A codex run that slept and resumed was priced once per leg: `exec resume`
+  reports the thread's cumulative token usage as each leg's usage, so
+  `outerloop status` overstated a resumed session's cost by roughly the
+  number of legs (#479). Session totals now price a codex thread once, at
+  its newest leg. Claude Code sessions keep summing each leg's reported
+  cost: cumulative resume totals arrived only in Claude Code 2.1.277, and
+  the CLI version that ran a session is not recorded — an author override
+  can pin an older CLI whose legs are per invocation — so summing, which
+  can overstate a resumed session on a current CLI, never undercounts one
+  on an older CLI.
+  Upgrading: index entries written before this change carry no session id
+  and keep contributing their recorded cost, so resumed codex sessions
+  captured before the upgrade still overstate; totals are accurate for
+  sessions captured after upgrading.
+- A label-triggered review round no longer restarts the round counter at 1
+  after a summarizer (panel) round: both paths post as the same underlying
+  reviewer, so the count continues (#373). Upgrading: no action needed; round
+  stamps already posted are still counted, and the next round continues the
+  reviewer's count instead of restarting at 1.
+- A manual climb now refuses a relative codex/hermes panel key path, like the
+  service preflight always did (#498).
+- A session that ends in error while an out-of-memory kill happens in the job
+  ended its run with only the session's symptom; the ending now notes when an
+  out-of-memory kill was observed in the job during the session (#483).
+  Upgrading: no action needed; endings already recorded keep their text, and
+  only a session that ends after the upgrade carries the note.
+- A model-API rate limit (HTTP 429) is now recognized as an outage in every
+  backend's wording, and always takes the latch's short 5-minute cooldown,
+  so throttled runs park briefly instead of ending aborted (#481). Upgrading:
+  no action needed; a latch stamped before the upgrade keeps the cooldown it
+  recorded (up to 45 minutes), and a throttle stamped after it takes 5 minutes.
+- An author session that stops without staging anything (an empty reply after
+  long reasoning, or a line of narration with no command) no longer ends its
+  run as "ended without a submit". The kernel resumes the session with a note
+  that nothing ran and the run is still open, up to three times per leg. A stop
+  after a kernel note, a gate verdict, or with a PR open is still taken as the
+  agent's answer, and `stop` (formerly `end`) still ends the run. Upgrading: no action needed; a
+  run already ended as "ended without a submit" stays ended, and the resume
+  applies to a leg that stops after the upgrade.
+- A Codex session that recovers from a transient error no longer ends its run
+  as aborted. Codex reports a dropped or timed-out model stream as an `error`
+  event ("Reconnecting... 1/5") and carries on; the kernel now counts an error
+  as fatal only when no completed turn follows it, the turn fails, or Codex
+  exits non-zero (#491).
+- Codex installed from npm (the review workflow's install) runs through a
+  script launcher whose vendored binary carries its own code-mode host; the
+  session no longer refuses it for lacking a host beside the launcher. A
+  contained session refuses a launcher up front (the image has no Node.js).
+- `examples/review.yml` referenced a nonexistent `@v1` tag; it now pins the
+  current release (`v0.3.0`, with a matching `reviewer_ref`) (#475).
+- Native Codex sessions failed on the first turn because Codex 0.160's code-mode
+  host was not installed. The installer now installs and verifies it, and a
+  session refuses to start without it.
+  Upgrading: run `outerloop harness upgrade codex`; deploys that run harness
+  upgrades do this automatically. Existing single-binary installs are detected
+  as drift and replaced with a verified pair.
+- Uncontained sweep tasks now receive `SWEEP_INDEX`: the bare evaluation
+  branch runs the command under `env -i` and dropped it, so every task of a
+  sweep without an image ran the default configuration.
+
+### Changed
+
+- A run whose benchmark is no longer in its contract (renamed or removed)
+  stops with one clear message naming the benchmarks the contract has,
+  instead of continuing on default settings (no GPU budget, default eval
+  walltime, research lines off). A contract that fails to load still leaves
+  parked runs parked until it is fixed. Upgrading: finish or rebind a parked
+  run before upgrading if its benchmark was renamed or removed; on its next
+  wake it ends aborted with a message naming the benchmarks the contract has.
+
+### Removed
+
+- Upgrading: labels and comment markers are recognized only under the
+  `outerloop:` prefix; the retired markers are `followup`, `body-edit`,
+  `claimed`, `claim-released`, `withdraw`, `unmeasured-merge`,
+  `advisory-review`, `verification-review`, `maintenance-digest`,
+  `round-skipped`, `contract-alarm`, `outage-release` and `research-log`, and
+  the retired labels are `review`, `steward`,
+  `no-review` and `task`. Rename any `autoresearch:*` labels on target repos
+  (`gh label edit autoresearch:review --name outerloop:review`), update
+  workflow `if:` conditions that match the old label, and change an open
+  research-log issue's `<!-- autoresearch:research-log -->` marker to
+  `<!-- outerloop:research-log -->`, or the kernel opens a new one; on old
+  comments and issue or PR bodies it no longer sees the retired markers, so it
+  may re-post a review or addendum, re-pick an already claimed issue, or open
+  a duplicate digest issue — finish or close them before upgrading, or edit
+  the marker to `outerloop:` where the kernel must still recognize it.
+- Upgrading: the contract budget key `climb_job_minutes` is no longer read,
+  and a contract that still uses it fails validation, so rename it to
+  `attempt_job_minutes` before upgrading. A wake reads both the current base
+  branch's contract and the contract at the commit the run was parked on, so
+  finish or rebind a run whose parked commit still uses the old key: it fails
+  on wake and stays stuck even after the branch is renamed.
+
+### Added
+
+- Add optional per-benchmark `regression` allowances with a free tolerance,
+  exclusive hard cap, and optional measured climbed-gain requirement, in
+  relative or absolute units. Suite rows and reports identify the applied rule.
+  Upgrading: no action for existing contracts; their gate behavior and measurement
+  signatures are unchanged. Legacy suite rows without `rule` default to
+  `legacy-floor`; the new row field is additive. Contracts using the new block
+  require this kernel version and must remove it before rolling back.
+
+### Changed
+
+- Codex sessions never get Codex's built-in sub-agent tools; multi-agent work
+  goes through the kernel's channels. `OUTERLOOP_CODEX_WEB_SEARCH` controls
+  Codex's built-in web search: `auto` (default) leaves it to Codex on native
+  provider sessions and turns it off on endpoint sessions, where it cannot run;
+  `off` turns it off everywhere. This fixes Codex sessions on chat-only
+  endpoints, which failed with "unsupported bridge request" after the 0.160 pin.
+  Upgrading: no action.
+- `OUTERLOOP_TOKEN_PRICES` in the deployment `.env` now reaches ticks and the
+  jobs they submit, so session cost is priced from the operator table instead of
+  staying unknown. Upgrading: no action.
+- GPU job sizing is an operator setting: `OUTERLOOP_EVAL_CPUS_PER_GPU` and
+  `OUTERLOOP_EVAL_MEM_GB_PER_GPU` (defaults 8 and 64) set the per-GPU core and
+  memory floor for GPU evals and author launches, for nodes with fewer cores or
+  less memory per GPU. Upgrading: no action; unset keeps the defaults.
+- Apply the current contract's channel settings before wake message delivery or
+  workspace fetches; a run keeps the verification mode it started with, so a
+  contract change applies to runs that start after it. Allow measured results to
+  replace an earlier self-reported leader without inheriting its claimed baseline.
+
+- Bump the pinned Codex CLI from 0.130.0 to 0.160.0 and update its Linux
+  archive digest. All contained Codex roles bind a kernel-owned, read-only
+  managed-hooks-only policy; uncontained sessions disable lifecycle hooks and
+  plugins (including built-in cleanup hooks) because this CLI has no per-session
+  requirements-file override. Rebuild user config before every launch/resume to
+  discard persisted hook trust, while
+  retaining session history. Reject the hook-trust bypass flag.
+  Upgrading: no manual action or run-record migration is needed; existing
+  session homes are sanitized automatically on their next launch. The policy
+  ships with the kernel and needs no image rebuild or system configuration.
+- Record the Codex 0.160.0 project-config finding in `SECURITY.md`: writable
+  sessions auto-trust project config, so repository model settings can apply;
+  project provider settings are filtered. The external containment boundary
+  remains necessary and is unchanged.
+- Retain private per-invocation session evidence: delivered prompt, bounded native
+  log, timestamps, identifiers, usage, and SHA-256 hashes beside existing captures.
+  Record missing logs and truncation without failing the author. Capture Codex and
+  Hermes tokens and price them with `OUTERLOOP_TOKEN_PRICES`; unpriced cost is
+  unknown, while explicit zero prices remain zero. Status exposes known spend
+  and unpriced session counts. Bound native discovery by depth, entries, and
+  elapsed time; withhold unresolved secret prefixes at truncation boundaries.
+  Incomplete Codex turn usage remains unpriced, and final-message reads refuse
+  symlinks at every path component and apply a byte cap. Redact provisioned
+  credential file contents before retention, leave failed Codex invocations
+  unpriced, and hash Hermes's delivered resume prompt. Cost totals persist in a
+  kernel-written, atomically replaced index outside the workspace and session
+  HOME. Separate readers and restarted kernels retain totals; any uncontained
+  invocation marks them unverified, also shown in status.
+  Upgrading: legacy sidecars are not imported into cost totals; pre-index history
+  has unknown coverage.
+  Upgrading: new session sidecars and stage usage fields are additive; older
+  records without them remain readable, with unknown evidence coverage and no
+  fabricated token backfill. Numeric legacy costs retain their meaning. New
+  unknown costs use JSON null; older kernels that coerce parked costs with
+  `float()` cannot resume those records without an update. Drain parked runs
+  before rolling back to such a kernel.
+- Add benchmark `verification: self_report` for explicitly labelled author claims,
+  with required candidate and baseline values at submit, scope enforcement and
+  improvement floors, without evaluation or panel work. Self-reported PRs never
+  arm automatic merging. Add contract channel switches for sibling views,
+  inter-agent messages, shared reports and other agents' line branches. Defaults
+  retain gated verification and all sharing. Document the existing explicit
+  backend/model panel setting for fixed judges.
+- Upgrading: run records, briefs, board rows and submission/leader ledger rows
+  gain optional policy or provenance fields. Missing fields retain gate mode,
+  enabled channels and measured provenance. Default records and rendered text
+  retain their previous bytes; measured ledger rows omit provenance and remain
+  readable by older strict readers. Default startup adds no fetch. No backfill
+  is needed; legacy and HEAD-output fixtures cover compatibility.
+  Start fresh runs for channel isolation because earlier context and Git objects
+  cannot be removed reliably. Keep the updated kernel while self-reported claims
+  exist: older strict ledger readers cannot read explicit self-reported
+  provenance, and older kernels cannot enforce the new contract settings.
+
+
+## [0.3.0] - 2026-10-05
+
 ### Upgrading
 
 Operator actions (everything else needs no action; details in each entry):
@@ -226,7 +601,7 @@ Operator actions (everything else needs no action; details in each entry):
   selections without changing legacy routes. New endpoint authors save
   `model[endpoint=profile]` selectors; keep those profiles until runs finish, and finish
   endpoint runs before rolling back. Hermes source/runtime must be upgraded
-  with the harness. See [compatibility and rollback](docs/endpoints.md#upgrade-and-rollback).
+  with the harness. See [compatibility and rollback](docs/endpoints.md#changing-a-profile).
 
 ### Fixed
 
@@ -385,7 +760,7 @@ tasks to finish.
 - A review panel skipped at preflight or in the follow-up now sends the author a kernel note and records the reason in the submit’s PR addendum. These publishes never arm auto-merge.
 
 - A submitted review measurement could go unreported when its push failed. The number is now posted once, as soon as it is known. Review submits push a bot commit with the report’s first line and measured number as its message.
-- On local compute an author-sleep wake could lose an improvement: the gate answers inline there, and the wake only knew how to end a run, so the result became an aborted ending with no PR (cluster0, three runs). Fresh and resumed climbs now share one terminal: the report, the line notebook, a PR or an ending record, and the issue note. A wake that opened its PR but died before recording it reconciles to that PR instead of opening a second, and a park's snapshot is released only once the run has left waiting.
+- On local compute an author-sleep wake could lose an improvement: the gate answers inline there, and the wake only knew how to end a run, so the result became an aborted ending with no PR (local mode, three runs). Fresh and resumed climbs now share one terminal: the report, the line notebook, a PR or an ending record, and the issue note. A wake that opened its PR but died before recording it reconciles to that PR instead of opening a second, and a park's snapshot is released only once the run has left waiting.
 - A research line lost its snapshot when the author's session reset its branch to main: the seal parented on main, the push was refused, and the run left no notebook entry (#368). The kernel now records the line head itself and the line only moves forward from that record. Memory files a reset dropped from the tree come back at the seal; files the session deleted on the line stay deleted.
 - Intake dropped an issue from a private org member without a word: the App's token sees such an author as CONTRIBUTOR, and only OWNER/MEMBER/COLLABORATOR qualified. Every skipped issue is now logged with its reason, a maintainer's `outerloop:task` label vouches for an issue regardless of association, and new Apps request members read so private members read as MEMBER.
 
@@ -676,7 +1051,7 @@ removes them. The full history since the first pre-release follows.
 - GPU BENCHMARKS: a per-benchmark contract field `gpus` (default 0) sizes
   every dispatched job of that benchmark — gate measures and author
   launches — with `--gpus-per-node=N` on the job (the per-job `--gpus` form
-  is rejected by Torch's submit plugin as a "CPU job" on a GPU partition —
+  is rejected by the cluster's submit plugin as a "CPU job" on a GPU partition —
   the first fleet speedrun attempt aborted on it) and `--nv` on the jail. GPU jobs go
   to a dedicated lane (`AUTORESEARCH_GPU_PARTITION`, optional
   `AUTORESEARCH_GPU_ACCOUNT`); a `gpus > 0` benchmark on a deployment with no
@@ -1040,7 +1415,7 @@ removes them. The full history since the first pre-release follows.
 - Self-deadline: climbs arm a SIGALRM at walltime-minus-margin (default
   120s, floor 60s, `--deadline-margin-s`) raising into the ordinary
   containment — the only pre-kill warning on clusters that deliver no
-  signals to job processes (measured on Torch 2026-08-08). The tick passes
+  signals to job processes (measured on the cluster, 2026-08-08). The tick passes
   each climb its own walltime via `--job-minutes`.
 
 - The advisory reviewer posts one comment PER ROUND (numbered, stamped
@@ -1055,7 +1430,7 @@ removes them. The full history since the first pre-release follows.
   job is terminal — Slurm truth plus grace, outage never reads as dead.
   Previously only crashes (Python exceptions) were contained; kills
   stranded the record forever. A SIGTERM handler also raises into the
-  ordinary containment, but measurement (Torch, 2026-08-08) showed Slurm
+  ordinary containment, but measurement (on the cluster, 2026-08-08) showed Slurm
   delivers no signal to job processes there — on such clusters the sweep
   and the self-deadline (below) are the real teardown paths; the handler
   covers direct kills and other sites.
@@ -2034,7 +2409,7 @@ removes them. The full history since the first pre-release follows.
   agent-01 right after its own win).
 - A job the site relocated off its submitted partition counts as lost only
   when it is also starving: eligible (dependencies done, slot passed) and not
-  started past the grace window. Relocation alone is routine on Torch and
+  started past the grace window. Relocation alone is routine on the cluster and
   cancelling a relocated job only reset its queue age.
 - `AUTORESEARCH_BOT_ALIASES` names the kernel's former logins, and every
   "is this ours" check (own issues, claims, alarms, own PRs, own comments)
@@ -2139,7 +2514,7 @@ removes them. The full history since the first pre-release follows.
   Also fixes the sha pin itself (here and in `scripts/install_hermes.sh`):
   hermes v-tags are ANNOTATED, so `ls-remote` had yielded the tag OBJECT's
   sha — every verify comparing `rev-parse HEAD` (a commit) against it would
-  have always failed, stubbing out all hermes lenses and making the Torch
+  have always failed, stubbing out all hermes lenses and making the cluster
   installer refuse. Pins are now the dereferenced `^{}` commit sha, and the
   installer's three paths (fresh / idempotent / tamper re-pin) were executed
   live against the real repo (a 222 MB clone — the cache earns its keep).
@@ -2151,7 +2526,7 @@ removes them. The full history since the first pre-release follows.
   that finished in minutes. The park-time afterany wake job (zero-latency
   primary) remains planned; the sweep stays the backup.
 - A genuine API refusal the Claude CLI reports as a `subtype: "success"` error
-  is no longer misclassified as a `session-error`. Surfaced on Torch: a session
+  is no longer misclassified as a `session-error`. Surfaced on the cluster: a session
   hit the author key's workspace usage cap and came back `is_error` with a
   content-free `subtype: "success"` and the real cause only in
   `result="API Error: 400 ... usage limits"`. The parse stamped `"success"` into

{outerloop_science-0.3.0rc2 → outerloop_science-0.4.0}/CITATION.cff

@@ -13,5 +13,5 @@ authors:
 repository-code: "https://github.com/outerloop-science/outerloop"
 url: "https://outerloop.science"
 license: Apache-2.0
-version: 0.3.0rc2
-date-released: "2026-10-04"
+version: 0.4.0
+date-released: "2026-10-20"

{outerloop_science-0.3.0rc2 → outerloop_science-0.4.0}/CLAUDE.md

@@ -18,7 +18,7 @@ uv run pre-commit run --all-files
 
 - **The bot never merges and is never a code owner** — do not weaken this in any
   code or config change.
-- **autoresearch is never a target of itself**; the contract file, roadmap, and
+- **outerloop is never a target of itself**; the contract file, roadmap, and
   `.github/` are forbidden write paths everywhere, regardless of contract YAML.
 - Budget caps are load-bearing safety features, not tunables to raise casually.
 - Never commit credentials, transcripts, or run artifacts (SECURITY.md).
@@ -29,7 +29,7 @@ uv run pre-commit run --all-files
   need.
 - Merge commits only; never rebase, squash, or force-push.
 - **Review until quiet**: development PRs iterate advisory-review rounds
-  (after a fix commit, remove then re-add the `autoresearch:review` label
+  (after a fix commit, remove then re-add the `outerloop:review` label
   once the push settles; the authorizing round must be the most recent,
   run against the head commit). Termination is judged, not literal: code
   PRs stop when a round yields no new medium+/behavior-affecting findings;
@@ -37,7 +37,7 @@ uv run pre-commit run --all-files
   then escalate to the PI instead of cycling. Merge only on an explicit
   `ci` success AND that quiet round. Read every review before merging —
   green is not read.
-- Imports are absolute (`from autoresearch...`); deps go in with their code +
+- Imports are absolute (`from outerloop...`); deps go in with their code +
   `uv lock`; CHANGELOG under `[Unreleased]`.
 - A PR that changes state read across kernel versions (run records, PR
   branches, ledger files, inbox messages, caches) carries a compatibility