DriftWatch phase ledger¶

Status, evidence, commit and the hours the researcher reports, for every phase and gate. The phase-closeout skill updates this file at the end of each phase. Status values: NOT STARTED, IN PROGRESS, DONE, CUT. Gates are decided by the researcher from the gate card in plan §7; the evidence file decides, not mood.

Phases¶

#	Phase	Weeks (dates)	Planned h	Status	Evidence	Commit	Hours (actual)	Notes
1	Foundations and governance	1-2 (12-25 Oct 2026)	8	DONE	CI run 38028624827 (in the private development repository, not public) (Linux + Windows green, 258 tests, 45/45 guardrail scenarios each); guardrail review; ADR-0001; PROTOCOL	d634836	3	Done 10 Oct 2026, before week 1; prompt rev 1.2. See “Phase 1 closeout” below
2	Data ingestion and data cards	1-2 (12-25 Oct 2026)	10	DONE	inventory; notes, Q1-Q7, leakage review; LAMDA card; BODMAS card; ADR-0004 (ACCEPTED); CLAIMS D-01..D-14; `data/manifest.json`	14da248	2	Done 10 Oct 2026, before week 1. Cap 1.5 GB approved; LAMDA peak 0.87 GB (rebuild 0.93), BODMAS 0.56 GB. Guard allowlist widened (ADR-0001). Cache v2 after leakage review. See “Phase 2 closeout” below
3	Time slicer, C1-C3, leakage canary	3 (26 Oct-1 Nov 2026)	8	DONE	slicer review, 2 passes; mutation matrix 15/15; slicing run; window options; ADR-0005 (ACCEPTED); ADR-0004 addendum 1; CLAIMS W-01..W-07	9b8e958	3	Done 11 Oct 2026, before week 3. Canary tested on seeded drift, not the signal-free fixture; real-data canary left to Phase 4 by design. See “Phase 3 closeout” below
4	Models, AUT/A-AUT, E0 baselines	3-4 (26 Oct-8 Nov 2026)	10	DONE	G1 evidence; E0 report; tables; fit times; ADR-0006 (ACCEPTED 15 Oct 2026); CLAIMS E-01..E-12, M-01, M-02, B-01..B-05, T-01	57e2003	2 (reported as under 3)	Done 11 Oct 2026, before week 3. G1 criteria met on LAMDA; BODMAS AUT above 0.90 left as an open alarm; ADR-0006 awaits acceptance; guide sanity band `[verify]`. See “Phase 4 closeout” below
5	Delay and budget simulator, E1	5-8 (9 Nov-6 Dec 2026)	32	DONE	G2 evidence; ADR-0007 (PROPOSED); E1 figure; E1 tables; label influence; chapters 2, 5, 6; CLAIMS S-01..S-06, E-13..E-21	a5c8604	3 to 6 (reported)	Done 13 Oct 2026, before week 5. G2 criteria met except: ADR-0007 not accepted (researcher), second leakage-review pass did not complete (usage limit). Open: which comparator carries RQ1; parallel runs do not fit the 1.5 GiB cap. See “Phase 5 closeout” below
6	Tier 1 grid (E2) and statistics	9-11 (7-27 Dec 2026)	24	DONE	prereg-v1 tag c26aa5f (15 Oct 2026, before any E2 result); ADR-0008 (ACCEPTED); stats review; GRID.md; ch03; grid ledger `results/e2/<dataset>/ledger.jsonl`	e015804 (final evidence); fca8a10 (code), c26aa5f (registered)	10 (reported by the researcher on 20 Oct 2026 for Phases 6 to 8 together; no per-phase split)	E2 complete (54 of 54 cells), tables, figures, CLAIMS E-22 to E-39 and ch07 written, two stats passes. Closed 20 Oct 2026 with Phases 7 and 8 (hours reported together). Open items for the researcher are listed under “Phases 6 to 8: state”
7	Tier 1 write-up and reproduction	13-14 (4-17 Jan 2027)	18	DONE	`driftwatch reproduce table1` (smoke passes; clean-clone full run MATCH, 31 min, CLAIMS E-24); claims audit script; related-work addendum; ch01, ch04 (first draft), ch08 (partial)	e015804 (on phase-08-conformal-e3, which contains phase 7)	included in the 10 h reported for Phase 6	README, ch01 to ch08 drafts, claims audit (PASS), clean-clone reproduction and G3 evidence exist. G3 decided PASS by the researcher on 20 Oct 2026. Closed with Phases 6 and 8
8	Conformal layer and selection (E3)	15-17 (18 Jan-7 Feb 2027)	23	DONE	conformal modules, selectors, `runner_e3`, E3 grid spec and tables built and tested on synthetic streams; addendum v2 draft; ADR-0009 (PROPOSED)	e015804	included in the 10 h reported for Phase 6	`prereg-v2` tagged f38858e (15 Oct 2026, before any E3 result); ADR-0009 ACCEPTED (option A). BODMAS E3 complete (28 of 28 cells, 0 failed, CLAIMS E-25); LAMDA E3 seeds 0 to 4 complete (16 of 16 cells); seeds 5 to 9 (12 cells) not run. Tables, figure, CLAIMS E-40 to E-47, ch07 section 7.5, a stats review and a leakage review exist; open: the trap-arm confound and the hold-out cut (review). Closed with Phases 6 and 7; open: the trap-arm caveat and the hold-out deviation are kept, not fixed
9	Drift monitors and scoring (E4)	18-19 (8-21 Feb 2027)	18	DONE	addendum v3 (tag `prereg-v3` on 1d302ea, 20 Oct 2026 17:52 IST, before any E4 result); ADR-0010 (ACCEPTED); leakage review; stats review; G4 evidence; tables; figure; synthetic checks; E3 cross-check; ch06 section 6.12, ch07 section 7.6; CLAIMS P-17, P-18, E-48 to E-65; tests 910 passed (fast), ruff clean, guardrails 45/45, claims audit PASS	ff0d455 (E4 results at 1d302ea)	10-18 (range reported by the researcher, 21 Oct 2026)	E4: 76 of 76 cells, none failed; seeds 0-9 decided before any result. No monitor beat periodic(3) (M-P); timing was equivalent where testable (M-C). The monitors alarmed in most slots. Two monitor designs (p-uniformity, MMD) were fixed on synthetic no-drift checks before the tag. The Holm floor of 0.0625 was not written into the addendum and is stated after the results (P-18). Branch stacked on PR #4; the worktree E:/DriftWatch-e4 is at prereg-v3 (remove with `git worktree remove`).
10	Tier 3: cross-platform (E5), EMBER2024 (E6)	20-21 (22 Feb-7 Mar 2027); now run 21-28 Oct 2026	0 (10 opt.)	DONE (MODE B, scoped)	Tier 3 decision; coverage (93.9% overall; constraints, queue, conformal and monitors 92% to 100% per module); results audit; `tests/test_configs_parse.py`, `tests/test_validation_branches.py`, `tests/test_fixture_regenerates.py`, `tests/test_audit_results.py`; fast suite 987 passed in 78 s, whole suite 1,128 passed, ruff clean, guardrails 45/45, claims audit PASS	dd3d1ae	4-6 (range reported by the researcher, 22 Oct 2026)	Go checklist failed on the deadline (due by 10 Dec 2026); E5 and E6 not run (future work), so the three Mode A artifacts (addendum v4, `configs/e5_cross_platform.yaml`, `reports/tables/e5_cross_platform.csv`) were not produced. Found and fixed: the slicer mutant `public_order_stratified` survived because an earlier speed-up marked its killer test module slow (the harness skips slow tests); the module is fast again and a test guards the harness suites. No script, config or function was deleted: every script has a test or a cited role, every config parses. The CI job timeout is 25 min (the slow suite takes 8 to 13 min) and the uv cache is on; neither is verified until CI runs. README updated for E4 and the deferred tiers
11	Paper and capstone completion	21-24 (1-28 Mar 2027); now to 10 Dec 2026	24	DONE (compressed: report draft; rewrite pending)	report chapters 0 to 9 (about 20,900 words; length against the guide’s budget in ch00, CLAIMS T-02); bibliography checker, 34 of 34 entries resolved live against Crossref or arXiv; simulated hostile review with responses; licence and ethics audit; claims audit PASS (110 numbered paragraphs, 106 with a resolving id, 4 marked [verify], 0 uncovered); clean-clone check at ed6be75 (CLAIMS E-67: 1,004 fast tests, guardrails 45 of 45, full Table 1 MATCH on 8 rows); fast suite 1,004 passed in 60 s, ruff clean, guardrails 45/45	0bd9992	8 to 12 h (researcher, 22 Oct 2026)	Compressed for the 10 Dec 2026 deadline (report plus code repository). HANDOFF items paper/main.tex and docs/submission_checklist.md are not produced: the 8-page paper and the venue checklist move to after the deadline (decided 22 Oct 2026). Decisions taken by the researcher on 22 Oct 2026 and applied: report retitled (“Malware Detector Lifetime Under Delayed Labels: Small Effects and Null Monitors in a Pre-Registered Time-Aware Study on Android and Windows PE Data”); exploratory lifetime reading added as ch07 section 7.7 (CLAIMS E-68, unregistered); chapter 4.5 cut to one paragraph; stored predictions hashed in docs/predictions_manifest.csv and archived privately outside the repository (CLAIMS E-69); BODMAS licence kept as “used as published”; docs/source removed from the index and git-ignored (history not rewritten); class-matched baseline not run (future work). Still open for the researcher: rewrite and defend the chapters; the faculty template, page limits and AI-disclosure wording (needed by about 5 Nov); the [researcher: …] and [verify] markers (BODMAS labelling, Li et al. against plan Table G.1, LAMDA timestamp field); weeks 1 and 2 hours; pushing and merging the PRs in order (PR #4, phase-09, phase-10, phase-11) and the prereg-v3 tag. Nothing pushed, tagged or published.
12	Release, deployment, maintenance	21-24 (1-28 Mar 2027); run 26-28 Oct 2026	12	IN PROGRESS (built and tested; nothing published)	Phase 12 checks (CLAIMS E-70 to E-72); ADR-0011 (ACCEPTED); RELEASING; MAINTENANCE; licence and ethics audit (updated 27 Oct 2026); architecture section with diagrams and captures; first maintenance run (`reports/maintenance/2026-10-first-run.md` on branch `maint/2026-10-first-run`, its own PR #7; not a site page, so no link); fast suite 1,056 passed, slow suite 142 passed, ruff clean, guardrails 45/45, claims audit PASS, strict docs build and link check pass, `driftwatch reproduce table1` on a fresh BODMAS download MATCH (8 rows, max difference 0.000000)	de2b64b	1 to 3 h (range reported by the researcher, 28 Oct 2026)	Built: release workflow (Trusted Publishing, TestPyPI rehearsal path, wheel smoke on Linux and Windows, draft GitHub Release so the Zenodo DOI is a separate decision), clean-room workflow, `driftwatch smoke` from an installed wheel, Sphinx site with strict build, link check and versioned deploy, maintenance playbook with the arxiv-audit skill and scheduled workflow, issue forms, CHANGELOG, CITATION.cff and .zenodo.json (both validate), filtered-history script (rehearsed), behaviour-preserving cleanup (CLAIMS E-71). Done on 28 Oct 2026 with the researcher’s go (CLAIMS E-76 to E-79): PRs #5, #6 and #7 merged in the private repository; the public repository `Yathindra25KB/driftwatch-lifetime` created from the filtered history (main and the three prereg tags only, branch protection and the `testpypi` and `pypi` environments set, Pages enabled); the release workflow rehearsed on the rc branch (build and the wheel smoke on Linux and Windows passed, the upload was refused only because no publisher is registered); the monthly arXiv workflow and the docs deploy ran on GitHub. Not done, each needs the researcher: registering the Trusted Publishers on TestPyPI and PyPI (account login, two-factor), the TestPyPI upload and `clean-room` run that follow, the signed tag and PyPI upload (the `Private :: Do Not Upload` classifier is still in place on main), publishing the draft release (Zenodo DOI), captures of PyPI and Zenodo. Deviations: the docs-tool ADR is ADR-0011, not ADR-0008 as the prompt names it (ADR-0008 exists; numbers are never reused), so the HANDOFF check lists `ADR-0008-docs-tool.md` as missing; `paper/main.tex` does not exist (paper deferred after the capstone deadline); `paper/REFERENCES_TODO.md` kept (nine open citation items); cleanup of `results/` deletes only the three uncited run folders, because ledger rows E-13, E-14, E-20, E-21 and S-06 cite the others.

Gates¶

Gate	Decide on	Pass criteria (plan §7, summarised)	Status	Evidence	Commit	Notes
G1	Sun 8 Nov 2026 (end wk 4)	Time-aware AUT(F1) below random-split F1 by the PREREG band on at least one dataset; real-data canary moves AUT toward random; constraint, AUT and canary tests green; no fit over 60 s or a shrink plan; reports/G1.md PASS	DECIDED PASS by the researcher on 15 Oct 2026 (ADR-0006 accepted the same day; the BODMAS AUT alarm and the comparator dependence of criterion 1 are recorded in chapter 8, not hidden)	reports/G1.md: criteria met on LAMDA (gap 0.14 to 0.15, canary PASS, tests green, fits under 9 s, reruns identical); BODMAS alarm open	57e2003	Researcher decides. Fail: fix the slicer; no simulator
G2	Sun 6 Dec 2026 (end wk 8)	Simulator end to end on at least one dataset; test_queue_delay and spy test pass; regression fixture holds; E1 figure regenerates; pool ADR accepted; reports/G2.md PASS	DECIDED PASS by the researcher on 15 Oct 2026 (stated in the Phase 6 planning session; ADR-0007 accepted the same day; the second leakage-review pass was then completed, see the Phase 6 stats and leakage notes)	reports/G2.md		Open decision A resolved (per-slot comparator, ADR-0008); open decision B: sequential grid, cap unchanged
Tier 1 results	Sun 27 Dec 2026 (end wk 11)	E2 complete; prereg-v1 predates results; tables and figures regenerate with no fits; stats-reviewer findings resolved	NOT DECIDED			Fail: use buffer week 12, no Tier 2 work in it
G3	Sun 17 Jan 2027 (end wk 14)	Tier 1 results and core chapters exist; reproduce table1 matches on a clean machine; claims audit clean; reports/G3.md PASS	DECIDED PASS by the researcher on 20 Oct 2026, accepting that criterion 2 (core chapters exist) is met by drafts that are under the guide’s page targets and that the chapters are the researcher’s to rewrite and defend	reports/G3.md		Fail: cancel Tier 2 and Tier 3, finish the report
G4	Sun 21 Feb 2027 (end wk 19)	E3 and E4 complete or honestly reported; prereg-v2/v3 predate results; reports/G4.md written	DECIDED PASS by the researcher on 21 Oct 2026	reports/G4.md	ff0d455	Decided early (calendar week 2). E3 56/56, E4 76/76 cells; both tags predate their results
Tier 3 go/no-go	Mon 22 Feb 2027 (wk 20)	G4 passed; cumulative hours at or under plan (151 h at end of wk 19); average at least 9 h/week over the last four weeks; capstone deadline not before end of wk 24	CHECKLIST FAILS on the deadline item (researcher’s numbers, 21 Oct 2026: G4 PASS; hours within plan; at least 9 h per week; deadline earlier than 28 Mar 2027)	reports/G4.md	ff0d455	Fail: Phase 10 in polish mode (MODE B), per the Phase 10 prompt; recorded in reports/G5-tier3-decision.md when Phase 10 starts
Release	Sun 28 Mar 2027 (end wk 24)	Paper at 8 pages or fewer, check_bib.py clean; chapters 1-9 complete; TestPyPI rehearsal passed; docs strict build clean; licence and ethics audit clean; researcher said go for PyPI and DOI	NOT DECIDED			Fail: smaller release; never rush irreversible steps
Hour check	Every week	Average at least 5 h/week over the last four weeks	OK	Weekly hours log		Fail: single dataset; two-dataset comparison becomes future work

Weekly hours log¶

The researcher fills this in weekly (about 20 minutes, plan §5.1 step 9). Compare cumulative hours with plan Table 4.1. Planning assumption: about 8.5 h per working week (confirmed by the researcher on 10 Oct 2026).

Week	Dates	Hours	Cumulative	Notes
0	10-11 Oct 2026 (before week 1)	10	10	Phase 1 (3 h), Phase 2 (2 h), Phase 3 (3 h) and Phase 4 (2 h, reported as under 3), researcher-reported
1	12 Oct - 18 Oct 2026	5 to 9		Researcher-reported as a range on 22 Oct 2026, not an exact figure; cumulative and 4-week average left blank until exact hours are given
2	19 Oct - 25 Oct 2026	5 to 9		Same: range reported on 22 Oct 2026 (week 2 not yet over)
3	26 Oct - 01 Nov 2026
4	02 Nov - 08 Nov 2026			G1
5	09 Nov - 15 Nov 2026
6	16 Nov - 22 Nov 2026
7	23 Nov - 29 Nov 2026
8	30 Nov - 06 Dec 2026			G2
9	07 Dec - 13 Dec 2026
10	14 Dec - 20 Dec 2026
11	21 Dec - 27 Dec 2026			Tier 1 results
12	28 Dec - 03 Jan 2027			Buffer
13	04 Jan - 10 Jan 2027
14	11 Jan - 17 Jan 2027			G3
15	18 Jan - 24 Jan 2027
16	25 Jan - 31 Jan 2027
17	01 Feb - 07 Feb 2027
18	08 Feb - 14 Feb 2027
19	15 Feb - 21 Feb 2027			G4
20	22 Feb - 28 Feb 2027			Buffer; Tier 3 go/no-go
21	01 Mar - 07 Mar 2027
22	08 Mar - 14 Mar 2027
23	15 Mar - 21 Mar 2027
24	22 Mar - 28 Mar 2027			Release

Revision 1.2: memory overrides (8 GB laptop, 7.2 GB usable)¶

The Phase 1 prompt (rev 1.2, 10 Oct 2026) states the real hardware and says it wins over any source document that assumes more memory. Places that assume more:

Source	Location	Assumption	Status under rev 1.2
Implementation plan	p. 1, “Scope you chose” box	“Hardware 16 GB RAM laptop”; verdict condition “LAMDA cached inside 16 GB in Phase 2”	Superseded: 8 GB (7.2 GB usable)
Implementation plan	p. 9, Table 4.5	“LAMDA, dense float32 … 18.4 GB, Impossible on 16 GB”; “LAMDA, dense int8 … 4.60 GB, Possible but wasteful”; “keep peak memory under 6 GB”	Dense int8 (4.6 GB) is no longer safely possible; the 6 GB cap is replaced by a cap the researcher approves in Phase 2
Implementation plan	p. 9, verdict table	“Memory (16 GB) PASS if LAMDA is streamed”	To re-test in Phase 2 against the approved cap
Implementation plan	p. 10, condition 3	“LAMDA is cached inside 16 GB in Phase 2”	Superseded by the Phase 2 cap
Implementation plan	p. 16, Phase 2 description	LAMDA memory hazard sized against a 16 GB machine	Superseded
Implementation plan	p. 17, Phase 2 prompt and acceptance	“Peak memory must stay under 6 GB … because I have a 16 GB laptop”; “Measured peak memory … is below 6 GB”	Prompt text replaced in docs/prompts/phase-02.md (see below); acceptance uses the approved cap
Implementation plan	p. 18, Phase 2 fallback	“LAMDA cannot be cached inside 16 GB”	Read as “inside the approved cap”
Implementation plan	p. 43, answered questions and Q3	“Hardware 16 GB RAM laptop”; Q3 default “8 threads, 16 GB RAM, 50 GB free [estimate]”	Measured 10 Oct 2026: Ryzen 7 6800H, 8 cores / 16 threads, 7.25 GB RAM visible, E: 121 GB free, C: 5.4 GB free
DriftWatch_Phase_Prompts.md	line 97 (Phase 2, deliverable 2)	“Peak memory must stay under 6 GB … 16 GB laptop”	Replaced in docs/prompts/phase-02.md
Build guide	(none found)	The guide states dataset sizes only, no laptop RAM	n/a

Prompt edits¶

Date	File	Edit	Authority
10 Oct 2026	docs/prompts/phase-01.md	Verbatim rev 1.2 prompt; supersedes the Phase 1 text inside docs/source/DriftWatch_Phase_Prompts.md (byte-compared with the pasted prompt)	Phase 1 prompt rev 1.2
10 Oct 2026	docs/prompts/phase-02.md	Replaced “Peak memory must stay under 6 GB (record the measured peak) because I have a 16 GB laptop.” with “I have an 8 GB laptop (7.2 GB usable). Before building, measure free memory and propose a peak-memory cap for me to approve, then record the measured peak against it.”	Phase 1 prompt rev 1.2, deliverable 5
10 Oct 2026	docs/prompts/phase-03.md … phase-12.md	Split unchanged from docs/source/DriftWatch_Phase_Prompts.md (each byte-compared with its source block)	Phase 1 prompt rev 1.2

VISTAS capstone rules¶

Answer given 10 Oct 2026: all unknown. The plan defaults stay in force until verified: deadline not before end of week 24; Markdown sources converted to DOCX/PDF; AI assistance disclosed; no demo.

Rule	Status
Capstone deadline	Early December 2026, by 10 Dec 2026 (researcher, 21 Oct 2026; exact date to confirm). Hand-in: the report and the code repository. Earlier than Sun 28 Mar 2027, so Tier 3 is removed and Phases 11-12 are compressed (see the Phase 10 to 12 rows)
Report template (Word or LaTeX)	unknown, to verify
Page rules	unknown, to verify
AI-assistance disclosure policy	unknown, to verify (see docs/ETHICS.md)
Demo expected	unknown, to verify

Phase 1 decisions (10 Oct 2026)¶

OS Windows-native (Git Bash + PowerShell tools); repository github.com/Yathindra25KB/Driftwatch- (private until G3, public by Phase 12); licence MIT (ADR-0002); PyPI distribution driftwatch-lifetime, import driftwatch (ADR-0003); one long session; branch protection with required CI on main once pushed; about 8.5 h per week.

Deviations from source documents¶

Item	Deviation	Reason
`cli`	Module `src/driftwatch/cli.py`, not a `cli/` package	Phases 7 and 12 name `src/driftwatch/cli.py` (PROTOCOL D11)
Plan Appendix B	MultiEdit dropped; PowerShell covered; hook paths anchored on `$CLAUDE_PROJECT_DIR`; no Bash deny rules	Current Claude Code docs (ADR-0001)
Phase 1 plan: restart after step E	No restart needed	Hooks went live through the settings file watcher (verified live, ADR-0001)
Phase 1 plan: GitHub repo creation	Existing private repo `Driftwatch-` (created 6 Oct 2026) used as `origin`	Researcher supplied it
Guard download allowlist	`json.schemastore.org` added	Validating settings against the docs-linked schema (a schema-rejected settings file disables the guardrails)
phase-closeout skill	Authored with skill-creator guidance; its LLM eval/viewer loop was not run	Deterministic helper scripts are unit-tested; the prompt runs the checklist by hand in Phase 1

Relief levers used¶

None.

Phase 1 closeout (10 Oct 2026)¶

Checklist run by hand, as the prompt requires for the phase that creates the skill.

Check	Result
Clean clone of `phase-01-foundations` at `180368d`: `uv sync --frozen && uv run pytest -q && uv run ruff check . && uv run ruff format --check .`	258 passed; All checks passed!; 60 files already formatted
`uv run python scripts/check_guardrails.py` (local, Windows)	45/45 scenarios passed; latency median: guard 189 ms, reviewer guard 176 ms, formatter 346 ms, re-injection 181 ms, Stop gate 803 ms (runs the tests: justified)
CI run 38028624827 on `0d2b8ca`	lint-test (ubuntu) success: 258 passed, 45/45; windows-smoke success: 258 passed, 45/45. `d634836` only loosens scenario 44 to stop expecting Phase 1 to be IN PROGRESS (local: 258 passed, 45/45); CI reruns on the ledger push
Live checks in Claude Code	Guard blocked real Bash and PowerShell force pushes and an unlisted download; permissions denied a `docs/source/` write and a `.env` read; formatter, scoped rules and skill discovery observed; `/memory` lists CLAUDE.md and the rules, and approval prompts appeared for guardrail edits (researcher check)
`.claude/settings.json` vs https://json.schemastore.org/claude-code-settings.json	ok – validation done
CITATION.cff (`cffconvert --validate`) and `.github/workflows/ci.yml` (`check-jsonschema` vendor.github-workflows)	valid; valid
HANDOFF check (`check_handoff.py 1`)	35 items, all exist
Adversarial review	21 findings: 20 fixed with regression tests (`6cdb589`), 1 accepted fail-closed (review); residual risks in ADR-0001
ADRs	ADR-0001 ACCEPTED (researcher), ADR-0002 ACCEPTED (MIT), ADR-0003 ACCEPTED (`driftwatch-lifetime`)
GitHub	`main` and `phase-01-foundations` pushed to the private repo; `main` protected (required check `lint-test`, strict; PR required with 0 approvals; enforced for admins; no force pushes or deletions)
Low-priority items	`report/ch01_introduction.md` skeleton; `paper/references.bib` with 13 identifier-verified entries; unverified items in `paper/REFERENCES_TODO.md`

Phase 1 config hashes¶

File	SHA-256
uv.lock	`ef115bb63578ddd01a6826c68eb6a52d275cdfe05084e59faa63ac8b8b705d22`

(no configs/*.yaml yet)

Phase 2 prerequisites¶

Disk: E: has 121 GB free (allow about 30 GB for raw files and caches [estimate]). C: has only 5.4 GB free and holds the uv cache and the uv-managed Pythons. Moving UV_CACHE_DIR and UV_PYTHON_INSTALL_DIR to E: before Phase 2 is recommended (researcher’s machine setting).
Start the LAMDA download early in week 1. huggingface.co and the BODMAS host must be added to the guard’s download allowlist by a Phase 2 CHANGE_REQUEST.
RAM: 7.25 GB visible, about 0.6 GB free at the 10 Oct check. Phase 2 measures free memory and proposes a peak-memory cap for approval.

Phase 2 closeout (10 Oct 2026)¶

Check	Result
`uv run pytest -q`; `ruff check .`; `ruff format --check .`	322 passed; All checks passed!; 82 files already formatted
`uv run python scripts/check_guardrails.py`	45/45 scenarios passed
`uv run driftwatch data verify` (data/ and the clean-rebuild root data/rebuild_check/)	OK, 0 problems on both; 1,088 cache files byte-identical between them (cache v2)
Builds under the approved 1.5 GB cap	LAMDA 1,008,381 rows, 120 monthly units, peak 0.87 GB (rebuild 0.93), 93.1 s, 119.7 MB; BODMAS 134,435 rows, 424 weekly units, peak 0.56 GB, 23.4 s, 1,231.2 MB
HANDOFF check (`check_handoff.py 2`)	8 items, all exist
Leakage review (leakage-reviewer subagent)	13 findings (2 high, 4 medium, 7 low). 11 were fixed with tests in `bba2692` (cache v2). 2 are carried to the Phase 3, 4 and 8 ADRs (slot-boundary hold-outs; past-only duplicate sensitivity) and recorded in ADR-0004
ADRs	ADR-0004 ACCEPTED by the researcher (LAMDA month; BODMAS 2 weeks from 2019-08-26; M = 50, strict gaps; placeholder-dated benign rows excluded). ADR-0001 change request (download allowlist) approved
Sources	LAMDA HF commit `ad9614bdd5556767f97ced2fce797c2f06408ebf`; BODMAS Google Drive file IDs in data/manifest.json. The BODMAS download worked without a quota error

Deviations and notes:

LAMDA timestamps come from the root metadata.csv (added, AndroZoo’s collection date), which the README does not describe. year_month matches it for every row.
The BODMAS dataset licence is not stated by its authors (the code is BSD-2-Clause): open question Q5.
Cache v1 (the first build) is archived in the manifest. Its files under data/cache/*/v1/ and the rebuild root data/rebuild_check/ can be deleted by the researcher; the guard blocks recursive deletes outside temp folders.
The Phase 2 prompt’s CHANGE_REQUEST was “none”. The allowlist change was approved at plan time as the CHANGE_REQUEST noted in “Phase 2 prerequisites”.

Phase 2 config hashes¶

File	SHA-256
uv.lock	`1c7380afb87669761538e9cc8fec08f100a1b94d5e5bf719ca9b61b26745afd0`

(no configs/*.yaml yet)

Phase 3 closeout (11 Oct 2026)¶

Check	Result
`uv run pytest -q`; `ruff check .`; `ruff format --check .`	481 passed; All checks passed!; 112 files already formatted
`uv run python scripts/check_guardrails.py`	45/45 scenarios passed
HANDOFF check (`check_handoff.py 03`)	7 items, all exist
Constraints	C1, C2, C3 raise `ConstraintViolation` subclasses (never warnings); tests implement the guide’s definitions literally; property tests run against an independent integer-arithmetic oracle
Mutation checks	15 of 15 deliberate breakages killed by the real suites, baseline 120 tests green (matrix, CLAIMS W-07)
Canary	Planted drift: AUT 0.76-0.79 (true timestamps) to 0.96-0.98 (shuffled), random-split F1 0.96-0.98, 4 seeds. No drift: `NO_GAP`. Leak caught by C2/C1, and by the canary when both are disabled. Limit pinned: leaking 1 of 10 evaluation slots is not seen by the canary alone
Full-dataset slicing	LAMDA 47 windows / 564 evaluation slots, 28.6 s, peak 0.40 GB; BODMAS 2 windows / 24 evaluation slots, 2.9 s, peak 0.69 GB; 0 densified; no C1-C3 violation (CLAIMS W-04)
Leakage review (leakage-reviewer subagent)	Pass 1: 14 findings (0 blocker, 3 major, 11 minor); pass 2: 9 findings (1 major latent, 7 minor, 1 informational). No C1-C3 violation in either pass. All fixed except two carried to Phase 5 (Window object reaches held-back labels; per-row seen-before mask for E0) and one documented as accident-proofing
ADRs	ADR-0005 ACCEPTED by the researcher (LAMDA nT 4, nE 12, stride 3; BODMAS nT 3, nE 6, stride 2; M = 50; cap and flag; sensitivity ratio 5%). ADR-0004 addendum 1 amends item 7 (selectors get only `EvalSlot.public()`)

Deviations and notes:

The committed synthetic fixture has features independent of labels, so the canary movement is tested on seeded in-memory timelines with planted drift (driftwatch.data.synthetic). The fixture caches and a built drifting BODMAS-format cache prove the script end to end.
The real-data canary was deliberately not run: windows must not be chosen after seeing real AUT. Phase 4 runs scripts/leakage_canary.py with the ADR-0005 windows (gate G1).
C3 follows the researcher’s “cap and flag” choice: slots already below the target ratio keep all rows and are flagged. 31 of the 62 LAMDA headline months are below 10% malware; the ADR states the prevalence spread as a threat to validity.
Carried to Phase 5: selectors, policies and monitors take only OpaqueSlot and TrainWindow, never Window or EvalSlot; the spy test lands with the first selector; the E0 duplicate mask is built from the simulator’s real training pool.
Branch phase-03-time-slicer is stacked on the unmerged phase-02-data-ingestion and phase-01-foundations; nothing has been pushed.

Phase 3 config hashes¶

File	SHA-256
uv.lock	`1c7380afb87669761538e9cc8fec08f100a1b94d5e5bf719ca9b61b26745afd0`

(no configs/*.yaml yet; window specs carry their own digest in every slicing manifest)

Phase 4 closeout (11 Oct 2026)¶

Check	Result
`uv run pytest -q`; `ruff check .`; `ruff format --check .`	669 passed; All checks passed!; 133 files already formatted
`uv run python scripts/check_guardrails.py`	45/45 scenarios passed
HANDOFF check (`check_handoff.py 04`)	10 items, all exist
E0 evidence runs (clean commit cd72d83, `--require-clean`, five seeds)	BODMAS 180 window rows, 3,185,890 predictions, 965.7 s, peak 1.40 GiB; LAMDA 420 window rows, 23,210,920 predictions, 2,200.3 s, peak 0.90 GiB (E-01)
Reproducibility	Seed 0 of every variant rerun on the same commit: 16 of 16 cell prediction hashes equal, 120 of 120 window rows equal incl. comparator columns (E-11)
Fit time	Slowest fit 8.6 s (LightGBM default, BODMAS); configured models 1.2 to 3.5 s; the 60 s limit holds, no shrink needed (B-05)
Leakage review (leakage-reviewer subagent)	No temporal leak. 1 major (comparator class prior confounded with time) and 7 minor findings: fixed with tests in `92018af`, except the recovery-time caveat, which is documented in ADR-0006. 4 informational: the bench docstring and LightGBM seed-sensitivity are fixed; the `window_id` join note is in the table builder; sealed labels inside the runner are by convention only, so Phase 5 must give selectors only `OpaqueSlot` and add the spy test
Stats review (stats-reviewer subagent)	5 major, 5 minor, 1 informational: iid bootstrap replaced by moving-block bootstrap, unsupported “gap understated” sentence removed, comparator caveats added, training-prior range and slot denominators corrected, LAMDA reproducibility run and E-rows written before the report was committed. Applied in `2010858`, `2e1702a`, `57e2003`
Gate G1	Evidence in reports/G1.md: criteria met on LAMDA; BODMAS AUT above 0.90 left open. Not decided by Claude
ADRs	ADR-0006 PROPOSED (values frozen 11 Oct 2026 under the researcher’s delegation, before any E0 result); acceptance is the researcher’s

Deviations and notes:

Runner location. driftwatch.sim.{config,results,runner}, not new top-level modules: tests/test_smoke.py pins the package map to guide section 8. driftwatch run <config.yaml> is in cli.py.
First E0 run discarded unread. Started at commit 5893df7, stopped after the leakage review found the comparator confound and missing comparator counts; replaced by the run of the amended code (ADR-0006, process note). Its partial output results/e0/00420e5e86baa4ea/ is not used and can be deleted by the researcher (the guard blocks recursive deletes outside temp folders).
LR max_iter 300 to 1000 after a convergence-only probe on a training window (no scores printed): LAMDA converges at 325 iterations.
Memory. Dense LR keeps float32 and the random-split step frees its pooled copy; the BODMAS full run peaked at 1.40 GiB, 7% under the 1.5 GiB cap. Phase 5 pools will be larger: re-measure before adding labels to the training pool.
Comparator. Headline comparator is prior-matched on training class counts but not per slot (global malware cap), so RQ1’s attribution to leakage in the random split is not tested by E0.
Not done. The guide’s sanity band ([verify]: PDF not decodable here, values not supplied); a per-slot-matched comparator; any investigation of the BODMAS alarm.
Branch. phase-04-models-e0 is cut from origin/main after the Phase 3 PR merged; nothing has been pushed.

Phase 4 config hashes¶

File	SHA-256
configs/e0_bodmas.yaml	`36bc82c2a4c11421cc926dcb5781961ca42cfac63ea407df0f1eefbf565ca0d5`
configs/e0_lamda.yaml	`114dd2337c82dea1fadfdad7c180396dbb4912c833526aeefbe1fdac456954d2`
uv.lock	`2c5a7dfa5000f370ed1b4c1bba4c86c689a18a5bcce570be8b996fddddbca00e`

Run config hashes (code version, config and lock file): BODMAS 608f15e1a6628732eb71e3e2fc1bac3ede1c704f874649b320fd1f9abc9b7957, LAMDA 613c0f98e507a8ad9d2ea70d340da5107aeb3820c773d46083d4ba167395341f.

Phase 5 closeout (13 Oct 2026)¶

Check	Result
`uv run pytest -q`; `ruff check .`; `ruff format --check .`	815 passed, 1 deselected, 84 s; All checks passed!; 160 files already formatted
`uv run python scripts/check_guardrails.py`	45/45 scenarios passed
HANDOFF check (`check_handoff.py 5`)	14 items, all exist
Queue, spy, regression, determinism, resume	S-01..S-06: d in {0,1,2,3} with property tests; spy at component and runner level; fixture ordering; three deliberate breakages caught; a BODMAS arm killed at 3 of 18 units resumed to the identical predictions hash
Evidence runs (clean commit 2bd12d6, `--require-clean`)	Label influence: LAMDA and BODMAS, never / 0.5% / 5%, 3 seeds; E1: both datasets, five seeds, per-slot matched comparator (E-13..E-19). Longest refit 9.9 s
Memory	BODMAS E1 peak 1.47 GiB (98.2% of the 1.5 GiB cap); two BODMAS arms in two threads crossed the cap (E-20)
Leakage review (leakage-reviewer subagent)	Pass 1: no temporal leak; 2 major (latent cache-key hazard, ADR-0007 not yet written), 5 minor, 6 informational. Fixed with tests in `0117287` and later, except three carried (feature-vector hashes through queue and pool; calibration share not yet consumed; `code_version` ignores untracked files). Pass 2 did not complete (API usage limit); rerun on `b3ff456..HEAD`
Stats review (stats-reviewer subagent)	1 blocker (comparator choice moves the RQ1 verdict), 5 major, 4 minor, 2 informational: achieved-ratio, decomposition and spread columns added, class split and per-arm denominators added, “modest” wording removed, run log and CLAIMS rows written
Gate G2	Evidence in reports/G2.md. Not decided by Claude
ADRs	ADR-0007 PROPOSED; new-label weight 1 and no truncation are the researcher’s choices (11 Oct 2026), the other rules await acceptance

Deviations and notes:

Branch. phase-05-simulator-e1 is cut from phase-04-models-e0 (Phase 4 is not merged); nothing has been pushed.
E0 path moved onto the sealed door. The E0 loop now scores through SealedSlot; E0 predictions are unchanged (E-16 reproduces the A-AUT of E-02 and E-03 to four decimals, and E0’s window-level gap is reproduced from the E1 window table).
Config schema. selector, delay, budget_pct, pool and policy: periodic are open; nonconf, monitors and monitor-triggered retraining stay refused. Weight other than 1 and any truncation are refused by ADR-0007. E0 configs hash as before.
Store. Per-unit checkpoints with --resume (needs a clean tree); predictions carry series, policy, selector, delay, budget_pct.
Parallelism. run_many and --workers exist and are tested; at the approved cap they cannot be used on BODMAS (G2, Open decision B).
Failed and superseded runs. Three BODMAS diagnostic runs at 716608f failed on the memory cap; three at fa614c3 were superseded after leakage-review code changes and give an identical table; listed in ADR-0007 and E-21. Directories under results/li/ and results/resume_check/ can be deleted by the researcher.
Chapters. Drafts of chapters 2, 5 and 6; sources retrieved 13 Oct 2026; descriptions from search summaries are marked [verify]; new references are listed in the chapters, not added to paper/references.bib.
Not done. Uncertainty and credibility selectors, conformal layer, monitors (Phases 8 and 9); a stratified within-slot comparator; a noise-floor arm for the label-influence diagnostic; a seen-before mask from the real pool.
Environment. C: was at 100% during the closeout and 291 tests errored once on temporary files; a rerun on the same tree passed. Free space on C: before the next long run.

Phase 5 config hashes¶

File	SHA-256
configs/e0_bodmas.yaml	`36bc82c2a4c11421cc926dcb5781961ca42cfac63ea407df0f1eefbf565ca0d5`
configs/e0_lamda.yaml	`114dd2337c82dea1fadfdad7c180396dbb4912c833526aeefbe1fdac456954d2`
configs/e1_bodmas.yaml	`a1677b3162b1f3070bfedd52621f0837131ea1183855cd86a3c403644d72bce5`
configs/e1_lamda.yaml	`f55e605dd645c625b0e62caacf675f8d142bb030ab214520a2de8ddf6015476d`
configs/li_bodmas_b0.5.yaml	`ca077fd06a0332232da946ff41915183f7d4a99f77308c5a9f33caf1da6def20`
configs/li_bodmas_b5.yaml	`79e2a06f22082633d6324ec09bfee8dcd8c7e05ebee2331ff2eacf2761c0b24f`
configs/li_bodmas_never.yaml	`12592db8311419b049284352ffcceae0731368e24460e732e1d6a09c6c74da36`
configs/li_lamda_b0.5.yaml	`f1e2ce49c04fb1ee8b106e50d5eb92c8befb283f19d0e787a18957fce69749db`
configs/li_lamda_b5.yaml	`cae058f1d6eddcd619b88b8923f8f65f67a14f0bcf30f79f34816674e5b6d2ef`
configs/li_lamda_never.yaml	`3b94e93bb5b95eafdf07b033d6270d1bfb92a3088d37017ffc0cbf7e7945dc44`
uv.lock	`2c5a7dfa5000f370ed1b4c1bba4c86c689a18a5bcce570be8b996fddddbca00e`

Run config hashes (code version, config and lock file; commit 2bd12d6): E1 BODMAS 45ea4054625ca8253d5d27f51a40f36c5057348c20ac5c15b4c5c1070a25b122, E1 LAMDA df138cc96fa9162219b70318b7179da382baf257d3e253a4b536b3b15de44e42; label influence BODMAS never ed348555adec05b60a97e5066b3ea149fd8ae44a42e3915d2c7df91509cf1636, 0.5% 8076e24d690b20ad… (see CLAIMS E-13, E-14 for the 16-digit run directories).

Phases 6 to 8: state at 16 Oct 2026¶

Hours reported by the researcher on 20 Oct 2026: 10 h for Phases 6 to 8 together (no per-phase split; the budgets were 24, 18 and 23 h). Report title chosen by the researcher: “How Long Does a Malware Detector Last? Label Delay, Analyst Budget and Conformal Monitoring in One Time-Aware Protocol”; if the work becomes a paper: “Detector Lifetime Under Delayed Labels: A Pre-Registered Time-Aware Study on Android and Windows PE Malware”. The phases were closed on 20 Oct 2026 after the LAMDA seeds 5 to 9 entered the E3 tables (commit e015804).

Item	State
Branches	`phase-06-grid-e2` (from phase-05), `phase-07-tier1-writeup` (merges phase-06), `phase-08-conformal-e3` (from phase-07; carries all the work, head 6e16222 plus this entry). Pushed to `origin` with the researcher’s go; no pull requests; tags `prereg-v1` and `prereg-v2` are local only (not authorised to push)
Registered	tag `prereg-v1` on c26aa5f (15 Oct 2026, before any E2 result); tag `prereg-v2` on f38858e (before any E3 result); ADR-0006, ADR-0007, ADR-0008, ADR-0009 ACCEPTED; G1 PASS and G2 PASS recorded as the researcher’s decisions
E2 grid	54 of 54 cells complete (27 per dataset), 0 failed, code version c26aa5f; tables `reports/tables/e2_.csv`, figures `reports/figures/e2_.png`, regenerate byte for byte (CLAIMS E-26 to E-39); cell config hashes in `reports/tables/e2_ledger.csv`
E3 grid	BODMAS 28 of 28 complete and LAMDA 28 of 28 complete (Main cells seeds 0 to 9; the LAMDA seeds 5 to 9 cells were added on 20 Oct 2026); none failed; code version f38858e; tables `reports/tables/e3_*.csv`, figure `reports/figures/e3_selection.png` (CLAIMS E-40 to E-47); cell config hashes in `reports/tables/e3_ledger.csv`
Gate evidence	`reports/G3.md` written; the decision is the researcher’s
Claims audit	PASS (`reports/claims_audit.md`)
Tests	839 fast tests pass in about 43 s; the E3 slow modules pass (13 tests); ruff clean; guardrails 45 of 45
Reviews	stats: E2 two passes (review) and E3 one pass (review), major findings applied; leakage: E2 second pass (ADR-0008) and E3 (review): no leak; trap-arm confound and hold-out cut recorded and open
Clean-clone reproduction	`driftwatch reproduce table1` MATCH on a clean clone, 31 min (CLAIMS E-24)

Open items for the researcher (updated 20 Oct 2026)¶

Gate G3 (see reports/G3.md), and whether the chapter drafts count as “core chapters exist”.
Hours for Phases 6, 7 and 8.
The E3 trap arm (confounded with not retraining) and the hold-out cut that deviates from ADR-0004 item 8: rerun under an addendum, or keep the recorded caveats.
LAMDA E3 seeds 5 to 9: done on 20 Oct 2026 (decided by the researcher).
Whether to push the tags and open pull requests.
The [verify] markers in chapters 2, 4 and 5 (including “Li et al., NDSS 2025”).

Housekeeping not yet done¶

phase-06 and phase-07 are contained in phase-08 (no merge needed); the worktree E:\DriftWatch-grid is removed and E:\DriftWatch-e3 (code version f38858e) can be removed once no further E3 run is planned; E:\tmp-clean-clone can be deleted by the researcher. Tags prereg-v1 and prereg-v2 are pushed; one pull request for phase-08-conformal-e3 is the next step.

Phases 6 to 8 config hashes¶

Hashes of the config files and the lock at the closeout (commit e015804). The per-cell run config hashes (config, code version and lock) are in reports/tables/e2_ledger.csv and reports/tables/e3_ledger.csv.

File	SHA-256
configs/e0_bodmas.yaml	`36bc82c2a4c11421cc926dcb5781961ca42cfac63ea407df0f1eefbf565ca0d5`
configs/e0_lamda.yaml	`114dd2337c82dea1fadfdad7c180396dbb4912c833526aeefbe1fdac456954d2`
configs/e1_bodmas.yaml	`a1677b3162b1f3070bfedd52621f0837131ea1183855cd86a3c403644d72bce5`
configs/e1_lamda.yaml	`f55e605dd645c625b0e62caacf675f8d142bb030ab214520a2de8ddf6015476d`
configs/e2_bodmas.yaml	`efabb09e34fce610ec11b51c1369f0b6bb486eb22922286822df67a5aefa74ff`
configs/e2_lamda.yaml	`b57b5b08270ed170c22d4b266cb9627840517ec4d89f34d7ee55247c9fcc8f77`
configs/e3_bodmas.yaml	`06a640b53b40059979dcac714cf70a76b94bfa5e47b9d243bad8935f09d02189`
configs/e3_lamda.yaml	`4f0181081e658b660a6e0dd7048344bab1139c63af4ad8efb0c6f72d43e439d2`
configs/li_bodmas_b0.5.yaml	`ca077fd06a0332232da946ff41915183f7d4a99f77308c5a9f33caf1da6def20`
configs/li_bodmas_b5.yaml	`79e2a06f22082633d6324ec09bfee8dcd8c7e05ebee2331ff2eacf2761c0b24f`
configs/li_bodmas_never.yaml	`12592db8311419b049284352ffcceae0731368e24460e732e1d6a09c6c74da36`
configs/li_lamda_b0.5.yaml	`f1e2ce49c04fb1ee8b106e50d5eb92c8befb283f19d0e787a18957fce69749db`
configs/li_lamda_b5.yaml	`cae058f1d6eddcd619b88b8923f8f65f67a14f0bcf30f79f34816674e5b6d2ef`
configs/li_lamda_never.yaml	`3b94e93bb5b95eafdf07b033d6270d1bfb92a3088d37017ffc0cbf7e7945dc44`
uv.lock	`2c5a7dfa5000f370ed1b4c1bba4c86c689a18a5bcce570be8b996fddddbca00e`

Phase 9 config hashes¶

File	SHA-256
configs/e0_bodmas.yaml	`36bc82c2a4c11421cc926dcb5781961ca42cfac63ea407df0f1eefbf565ca0d5`
configs/e0_lamda.yaml	`114dd2337c82dea1fadfdad7c180396dbb4912c833526aeefbe1fdac456954d2`
configs/e1_bodmas.yaml	`a1677b3162b1f3070bfedd52621f0837131ea1183855cd86a3c403644d72bce5`
configs/e1_lamda.yaml	`f55e605dd645c625b0e62caacf675f8d142bb030ab214520a2de8ddf6015476d`
configs/e2_bodmas.yaml	`efabb09e34fce610ec11b51c1369f0b6bb486eb22922286822df67a5aefa74ff`
configs/e2_lamda.yaml	`b57b5b08270ed170c22d4b266cb9627840517ec4d89f34d7ee55247c9fcc8f77`
configs/e3_bodmas.yaml	`06a640b53b40059979dcac714cf70a76b94bfa5e47b9d243bad8935f09d02189`
configs/e3_lamda.yaml	`4f0181081e658b660a6e0dd7048344bab1139c63af4ad8efb0c6f72d43e439d2`
configs/e4_bodmas.yaml	`ce502a98352ff9630bfa7e6a3f461de51433409567f3fe5ef5c7cf561366e733`
configs/e4_lamda.yaml	`e2e856bb3c160a969389d13d63d3f6c497ef5b3e46da610bffd31ac12ff35b0e`
configs/li_bodmas_b0.5.yaml	`ca077fd06a0332232da946ff41915183f7d4a99f77308c5a9f33caf1da6def20`
configs/li_bodmas_b5.yaml	`79e2a06f22082633d6324ec09bfee8dcd8c7e05ebee2331ff2eacf2761c0b24f`
configs/li_bodmas_never.yaml	`12592db8311419b049284352ffcceae0731368e24460e732e1d6a09c6c74da36`
configs/li_lamda_b0.5.yaml	`f1e2ce49c04fb1ee8b106e50d5eb92c8befb283f19d0e787a18957fce69749db`
configs/li_lamda_b5.yaml	`cae058f1d6eddcd619b88b8923f8f65f67a14f0bcf30f79f34816674e5b6d2ef`
configs/li_lamda_never.yaml	`3b94e93bb5b95eafdf07b033d6270d1bfb92a3088d37017ffc0cbf7e7945dc44`
uv.lock	`2c5a7dfa5000f370ed1b4c1bba4c86c689a18a5bcce570be8b996fddddbca00e`

Phase 10 config hashes¶

No config file changed in Phase 10; the hashes are those of the Phase 9 table above (E4 configs as registered at prereg-v3).

Phase 11 config hashes¶

No config file changed in Phase 11 (git log phase-10-polish..HEAD for configs and uv.lock is empty); the hashes are those of the Phase 9 table above, and uv.lock is 2c5a7dfa5000f370ed1b4c1bba4c86c689a18a5bcce570be8b996fddddbca00e.

Phase 12 config hashes¶

No config file changed in Phase 12 (config_hashes.py gives the Phase 9 hash for every file under configs/). uv.lock changed: it now carries version 0.1.0 of the project and the docs dependency group (Sphinx, MyST, Furo and their dependencies; sphinx-copybutton was added and then removed); no runtime or development pin moved. Committed lock on phase-12-release: 5a94849d17b7ae4626dc3bafa97a3d6329525ec2cae65e304a8f58e944e20919 (84 locked packages besides the project). The stored results of E0 to E4 carry the earlier lock hash in their run records and are unaffected; a new run would carry the new one. The maintenance branch maint/2026-10-first-run upgrades the lock further (see reports/maintenance/2026-10-first-run.md) and is a separate pull request.