This is the sole canonical backlog for future work. It contains unchecked work that remains after the deepest red-team review and hardening pass (2026-08-01).
v1 was cut on 2025-09-03 (tag v1 on commit
7c5452b, still an ancestor of main); pyproject.toml version is 0.2.0Verified against the tracked tree at this head. Each needs an owner decision before it is changed; none was modified by either review pass.
examples/vfe/node_modules/ is committed (631 tracked files) even though
the root .gitignore lists examples/vfe/node_modules/. Git ignore rules do
not affect already-tracked files, so the vendored tree stays in the index.
Owner decision: untrack the directory (keep it ignored) and rely on
npm install per examples/vfe/README.md, or retain it intentionally.examples/package-lock.json records no dependencies
({"name": "examples", "lockfileVersion": 3, "requires": true, "packages": {}})
with no sibling examples/package.json. Owner decision: remove it, or add the
missing examples/package.json that it would describe.examples/vfe.jsx is duplicated verbatim at the examples/ root and in
examples/vfe/vfe.jsx (both 56,986 bytes). examples/README.md and
examples/AGENTS.md document only the examples/vfe/ copy. Owner decision:
remove the unused root duplicate or document why it exists.coverage.json is tracked at the repository root (387,508 bytes) and its
meta block records "timestamp": "2026-07-18T13:33:29", while the measured
total at this head is 89.1% (docs/_generated/gate-timings.json). No gate
arm, workflow, script, or document reads it, and .gitignore ignores
.coverage/coverage.xml but not coverage.json, so a --cov-report=json
run silently rewrites a stale committed file. Owner decision: delete it and
add the pattern to .gitignore, or make the gate consume it.pyproject.toml sets [tool.ruff] target-version = "py310" while
requires-python = ">=3.11" and the documented matrix is 3.11/3.12
(docs/environment.md); [tool.black] target-version already lists
py311/py312. The linter therefore grades against a floor two minor
versions below the package’s declared floor. Owner decision: set
target-version = "py311" and re-run start-gate lint.v5.0.3, not the estate’s v5.1.0. .github/gnn-pin.json
records ActiveInferenceInstitute/Generalized_Notation_Notation at
8f293d9ba810e6eda5c09842d9c7ed9ffa5f92c1 (v5.0.3), and
src/repos/cloning.py hard-codes the same revision for the gnn clone target
with the comment “GNN v5.0.3; the shared canonical revision for the education
workstream”. Start does not import the upstream GNN package:
src/gnn/main.py locates a GNN checkout and runs its pipeline with that
checkout’s own interpreter, so the estate’s v5.1.0 importer migration does
not apply here and the older pin is a deliberate choice rather than an
unfinished migration. If that decision changes, the
repository maintainer updates .github/gnn-pin.json and
src/repos/cloning.py together, then re-runs start-gate gnn and
start-gate gnn-exemplar (the exemplar arm needs network).pymdp carries no version pin, and is not on 1.0.5. pymdp is only a
clone target (src/repos/cloning.py: docxology/pymdp, branch textbook);
it is absent from pyproject.toml dependencies and from uv.lock, and no
module under src/, scripts/, or learning/ imports it (only generated
curriculum text under data/ mentions it). Owner: repository maintainer, if
a versioned dependency is ever wanted.data/written_curriculums/, data/translated_curriculums/, and
data/visualizations/ is generated output carrying a run timestamp. The
authored tree (data/prompts/, data/config/, docs/, learning/, src/)
contains no learning-objectives artefact, no glossary term list, and no
learner plan: git grep -i glossary matches only the prompt template
data/prompts/translation.md and generated output under data/. Until that
authored content exists, no Education-contract bundle can be emitted from
this repository and the migration cannot start here.
Owner: the Institute’s curriculum/education lead supplies the objectives,
glossary terms, and learner plan; the repository maintainer then wires them
into the generator. Blocked outside this repository; no code change here can
unblock it.Scoped and implemented during the mega-deep documentation pass. Severity definitions: Minor = typo, broken link, formatting; Medium = stale section rewrite, doc restructure, added missing guide; Major = large doc system overhaul or cross-cutting refactor.
.aii/SKILL.md (tracked file in
a public repo); replaced with a generic clone instruction.
(commit docs: sanitize local path in .aii/SKILL.md (public repo))learning/ guides documenting --input/--output flags that
start-curriculum rejects (exit 2); the canonical CLI uses
--visualizations-dir, --translation-dir, --curriculum-dir.
(commit docs: fix CLI flags in learning guides to canonical start-curriculum)src/pipeline/ module (canonical orchestrator) in
src/AGENTS.md and src/README.md; refresh the stale src/config/
description (now catalog.py + schemas.py + languages.py).
(commit docs: document src/pipeline in source references)docs/environment.md: mypy command now full-scope
(mypy src scripts learning --ignore-missing-imports), matching CI;
fix docs/TESTING.md relative path; align model env defaults with
src/perplexity/clients.py.docs/curriculum_gui.md: remove the “only the Python standard
library” overclaim (GUI imports project pipeline modules).README.md: unify Zenodo DOI (badge/citation use
10.5281/zenodo.17047617; inline text said ...19).docs/clones.md: add missing gnn and cerebrum repositories;
correct clone destinations to the code keys (rxinfer,
activeinference) and the after-cloning structure tree.docs/data_outputs.md: research filenames use display names
(e.g. Barry Bonds_research_20250903.json), not stable IDs; stable IDs
appear in diagrams/manifests.docs/docs_and_deployment.md: point at the existing
.github/workflows/deploy_docs.yml instead of describing an Actions
workflow as a to-be-added alternative.docs/FAQ.md: merge the duplicated “Where do outputs go?” entry.learning/curriculum_creation/README.md: correct Input lines
(config-driven), dependency list (drop pydantic, pathlib), and
--input/--output CLI examples.learning/curriculum_creation/USAGE_GUIDE.md: fix --input/
--output examples and the broken an# H1.docs/AGENTS.md (docs hub list): add methods.md, manuscript.md,
operations.md, curriculum_gui.md to the file index.docs/README.md: remove duplicate H1 (# START Documentation —
README vs # START documentation).docs/index.md: fix ** Professional Domains** bold-marker typo;
align the inferant-stream link label with its relative path.docs/getting_started.md, docs/pipeline.md,
docs/environment.md: docs/-prefixed intra-doc paths are wrong from
inside docs/; use relative names.docs/TESTING.md: annotate the --cov-fail-under=80 example as
the matrix baseline (release floor is 90).docs/conventions.md: use ./run_docs.sh --serve consistently.tests/AGENTS.md: remove duplicate test_repos_cloning.py /
test_run_script_integration.py entries.examples/AGENTS.md: fix off-by-one line counts for vfe.jsx
(1180) and vfe-compiled.html (9403).AGENTS.md root: complete the stable-commands list
(start-clone, start-curate-artifacts, start-run-history)..env.example: replace default_model_name placeholders with the
real defaults used by src/perplexity/clients.py.here.md: use $$/$ math delimiters so equations render on
GitHub (backslash delimiters are not rendered).None open. The artifact boundary and the provider matrix moved to the 2026-11-12 completed section below. What still blocks a release is one publication gate each, tracked where it is worked: the provenance floor in P1 and the coverage floor in P2.
data/artifact-manifests/current.json). Retain a small canonical example
set, archive or remove unexplained timestamped duplicates, and make the
curated manifest reproducible without deleting user-owned evidence. Owner:
repository maintainer — this is a human retention decision, and
uv run python -m scripts.audit_artifacts --check --fail-on-unprovenanced
(the release floor in .github/workflows/release-gate.yml) is the gate that
must pass once the debt is cleared.--dry-run and --estimate-cost
(learning/curriculum_creation/generate_custom_curriculum.py:1551-1556); the
only current dry-run coverage is at orchestrator/runner level
(tests/test_production_hardening.py:1111,
tests/test_gnn_runner.py:383). The rest of the earlier list is covered:
JSON output and exit codes by
test_cli_json_failure_is_bounded_and_machine_readable and
test_gui_entrypoint_has_stable_argument_failure_code, resume by
test_pipeline_dependencies_resume_and_optional_failures, budget refusal by
test_pipeline_budget_stops_before_downstream_stage_and_does_not_publish_failures,
authenticated remote binding by
test_gui_remote_binding_requires_explicit_authentication, and redaction by
test_structured_logging_redacts_sensitive_values. Owner: whoever adds the
tests; nothing external blocks this.docs/_generated/gate-timings.json, measured 2026-11-12T13:22:55Z) to the
90% release floor, and add per-module floors for every destructive
repository/path operation. The two floors differ today: the CI full arm
enforces --cov-fail-under=80 (scripts/gate.py) while
.github/workflows/release-gate.yml enforces --cov-fail-under=90, so the
release floor is currently unmet. Owner: repository maintainer. Neither floor
has hosted evidence, because the estate’s Actions runs are billing-locked.macos-smoke job. The configuration exists — .github/workflows/ci.yml runs
fast on 3.11, full on 3.12, and macos-smoke (lock plus
smoke-macos); docs/environment.md documents 3.11/3.12; all ten console
entry points are listed in AGENTS.md. What is missing is evidence: the
estate’s Actions runs are billing-locked, so neither job has a recorded green
hosted run. Owner: whoever holds the organization billing/runner decision;
local arm output is the interim evidence.audit arm (pip-audit --strict) and
.github/workflows/codeql.yml. Missing: no SECURITY.md, no secret-scan
configuration or scan record, and no triage or threat-model artefact in the
repository. Owner: repository maintainer with a reviewer; CodeQL results are
also unobservable while Actions runs are billing-locked.None of these is blocked; each needs an owner decision from the repository maintainer before it becomes work.
None currently. The one quality Major (mypy debt) was resolved in the 2026-08-01 comprehensive pass and is now enforced in CI.
These were validated and intentionally left unchanged after a review decision; they are documented here so the reasoning is not lost.
src/perplexity/curriculum.py raises without writing any output when any
section fails, so an incomplete curriculum is never mistaken for a complete
one. This is pinned by test_core_hardening.py::test_curriculum_generation_
_is_transactional. The earlier “preserve partial sections” suggestion was
evaluated and REJECTED because it contradicts this intentional, tested
guarantee.src/common/io.py
read_text/read_json reject a symlinked target file but intentionally
allow symlinked parents, because the repository is documented to run under
symlinked install/clone paths (monorepo workspaces); rejecting parent
symlinks would break those legitimate installs. Writes still reject both.src/perplexity/entity.py extract_entity_description falls back to the
full input when no Description: line exists; existing tests pin this.Verified against tracked files at this head. Nothing here rests on a hosted run, because the estate’s Actions runs are billing-locked.
data/artifact-manifests/current.json
(data/audience_research, data/domain_research, data/entity_research,
data/translated_curriculums, data/visualizations,
data/written_curriculums) and the artifacts arm runs
scripts.audit_artifacts --check --baseline data/artifact-manifests/current.json,
scripts.validate_outputs --check, and the offline fixture replay. The
retired “before the first release commit” framing was wrong: release v1 was
cut on 2025-09-03.tests/test_production_hardening.py
covers the RequestLimiter concurrency cap, cancellation interrupting
backoff, provider-reported cost and usage accounting, structured-log and
credential redaction, retry exhaustion, Retry-After parsing (including the
HTTP-date form), and payload validation.
tests/test_release_edge_matrix.py::test_provider_failure_matrix_redacts_secrets_and_honors_cancellation
covers 429 retry plus redaction plus cancellation, and
tests/test_release_edge_matrix.py::test_pipeline_contracts_usage_and_run_artifacts_are_serializable
covers usage normalization and aggregation at the pipeline-contract layer.
Budget refusal is covered by
test_pipeline_budget_stops_before_downstream_stage_and_does_not_publish_failures.
Both adapters run through ProviderAdapter/complete_chat_result;
tests/test_clients.py asserts the Perplexity and OpenRouter base URLs and
the credential requirement.Additional improvements landed after the first comprehensive pass (all four gates green: pytest exit 0, ruff clean, black clean, mypy 0):
ci.yml type-check step was expanded from a
small subset (src/pipeline src/config/schemas.py --follow-imports=skip) to
the full clean scope (mypy src scripts learning). This prevents regression
of the now-zero typing debt.PerplexityConfig/OpenRouterConfig gained fallback_models and
input/output_cost_per_million fields (validated) plus a to_chat_policy()
helper.CurriculumConfig gained perplexity_fallback_models,
openrouter_fallback_models, and the four cost-rate overrides (validated).domain, entity, curriculum, and translation provider paths thread
fallback_models and cost rates into ChatPolicy, so a model-level failure
now falls back to the next configured model and cost overrides reach usage
accounting.to_chat_policy mirroring and CurriculumConfig
validation; the existing negative-cost error contract is preserved.The earlier comprehensive + red-team passes (StageResult partial-failure,
skip_existing freshness, post-clone git hardening, prompt-injection data
framing, legacy 3_Intro cleanup, schema/io/repos/GUI/system hardening)
remain complete as recorded below.
Implemented and verified (pytest, ruff, black, mypy all green) during the comprehensive pass. This adds to the earlier red-team hardening:
src, scripts, and learning
(was ~150). Fixed the untyped orchestrator self.results, Optional-style
issues, API/typing inference across repos/system/learning/scripts.src/perplexity/clients.py): ChatPolicy.fallback_models
— on a model-level failure the client tries the next model before giving up;
attempts are reported across the pool. Covered by a test.src/perplexity/clients.py): documented
default-cost lookup so the default OpenRouter path reports a non-zero
estimated cost; explicit rates still win; unknown models stay honestly 0.src/common/prompts.py + domain/entity/
curriculum): untrusted research/entity/foundation content is wrapped in an
explicit “treat strictly as data” boundary before reaching the model.3_Introduction_Visualizations.py reduced from ~725 lines of dead
chart logic to a thin delegator to the canonical src/visualization runner.Implemented and verified (pytest, ruff, black all green; no new mypy errors) during the deep red-team pass. These are removed from the active backlog:
MAJOR (fixed): silent partial failure. StageResult.__post_init__
(src/pipeline/contracts.py) now re-assesses status after deriving item
failures, so a stage containing a failed item reports ok=False; the
pipeline no longer reports a run green on partial failure. Regression tests
added; the existing test that encoded the buggy behavior was corrected.
MAJOR (fixed): stale skip_existing cache. process_research_directory_detailed
(src/pipeline/stages.py) now verifies the existing output’s declared
research/fep_actinf input hashes (sibling JSON) before skipping; a stale
cache hit is regenerated instead of silently published as authoritative.
MAJOR (fixed): post-clone git hardening. update_repository and
get_repository_status (src/repos/cloning.py) now run git with committed
hooks disabled (-c core.hooksPath=/dev/null, --no-optional-locks,
GIT_CONFIG_NOSYSTEM), refuse repos whose .git/config enables
hooks/filters/fsmonitor/url-rewrites, and use fetch + fast-forward merge
instead of pull. Real-git behavior preserved and covered by tests.
perplexity/translation.py: skip_existing now verifies the recorded
source_sha256 (regenerate when provably stale, legacy behavior preserved);
failure messages no longer assume a colon prefix.
StageSpec/runner early-stop records skipped required stages; resumed runs
report real duration; usage/parsers/config schema hardening (see prior pass).
src/terminal/AGENTS.md (typewriter_effect,
glitch_effect, input_dialog signatures), src/system/AGENTS.md (dependency
attributes), docs/curriculum_gui.md (no longer overclaims “only stdlib”),
learning/curriculum_creation/AGENTS.md (deduplicated thin-delegator entry).config/schemas.py: reject CR/LF in source URLs.config/catalog.py: output_exists no longer raises on odd display names.pipeline/schemas.py: accept numeric schema_version=1.0; reject newline
section names.common/prompts.py: log unresolved non-strict placeholders; return JSON-safe
sorted variables list.common/io.py: symlink-target guard on reads; list_files glob confinement;
load_key_from_file first-wins + empty-vs-absent.common/paths.py: ensure_dir delegates to hardened ensure_directory.common/config.py: CRLF-tolerant frontmatter.common/env.py: prefer repo-root .env with cwd fallback.pipeline/history.py: defensive sort + malformed timestamps handled.pipeline/usage.py: total_tokens recomputed when inconsistent.pipeline/parsers.py: fence-aware heading parse + duplicate-heading warning.perplexity/clients.py: empty-content retries; http:// only for loopback
hosts; JWT redaction; jitter-before-cap; dead-code removal; redacted
client-construction errors.repos/manager.py: repo-name traversal refused on update/status/delete.repos/cloning.py: branch safety check; credential redaction; safe
cleanup_failed_clones (staging/empty dirs only).visualization/runner.py: unreadable curriculum skipped, not run-aborting.generate_curriculum_gui.py: cross-site POST rejection (CSRF), broader path
redaction, run-slot freed on any config exception.system/reporting.py: macOS memory fallback + explicit encodings.system/dependencies.py: generated synthetic artifact is now a warn-only
content check, not a required dependency.terminal/menu.py: all-disabled menu leaves no phantom selection.common/logging_utils.py: JWT redaction pattern added.tests/test_redteam_sweep.py (schema/io/history/usage/parsers/
runner/run_history coverage); 4_Translate_Introductions.py cleaned to a
thin delegator (dead block + validate_languages removed).An item is complete only when implementation, tests, documentation, generated artifacts, and release validation are aligned. Remove completed items instead of adding a history section; add newly discovered future work here.