V10 — INDEPENDENT CLOSING VERIFICATION, BASELINE v2.0.0

I built none of this and I took no closure note, no docstring and no green checker line on trust. I read the 32 requirements out of 03_Registers/REQ.csv, then opened the artefacts. I extracted all 50 tables and all 185 paragraphs of the delivered .docx with python-docx and separately extracted the delivered PDF with pdftotext, and read section 6 out of both, because the PDF is what a reader receives. I recomputed all ten weighted totals in all three trade studies by hand from the matrices printed in the delivered study, and recomputed both of trade study 1's published flip points algebraically. I opened both workbooks twice with openpyxl — once for formulas, once for cached values — dumped every sheet, and read the raw xl/worksheets/*.xml <pane> elements rather than trusting openpyxl's freeze_panes. I ran my own sub-10pt font scan over every .docx, .pptx and .xlsx in 05_Outputs and in 03_Registers, including the 46 runs inside .pptx table shapes that C29 cannot see. I ran my own credential and identifier scan over the whole tree. I printed BUSINESS_PLAN_A3.html in headless Chromium, rasterised it at 60 dpi and looked at the paper, twice — once clean and once under attack. I opened the v1.0.0 archive zip and listed its 107 files. I attacked C12, C34, C41, C42, C43 and C44 in six full disposable copies of the tree under /tmp/v10, never in the project, and ran run_checks.py only inside those copies, so 01_System/checker_run_log.txt still carries the builder's 08:50:09 run. I did not run build_all.py. Every Python I ran started with sys.dont_write_bytecode = True or PYTHONDONTWRITEBYTECODE=1.

The headline: REQ-SYS-03 is genuinely fixed. REQ-CON-04 is not, and it now fails for a worse reason than before — the requirement was widened to "every delivered Office artefact" and the check that enforces it was left looking at one folder, so the project's second delivered workbook fails BOTH halves of the new criterion and nothing in the suite can see it. Of the six checkers I attacked, C41 held and five did not: I defeated C43 with a decoy table and put V8's exact wrong totals — 2.85 / 4.05 / 2.30 — back on the face of the delivered study for 44 of 44 green; I defeated C44 by falsifying a diagram range, because two of its regexes are matched against lowercased text and can never fire; I defeated C12 by flipping one bit in a copy of D1 and shipping it as D5; I defeated C34 twice, exactly as V9 did, unchanged; and I defeated C42 by replacing the delivered register workbook with the v1.0.0 one. And DEF-046's own verification-of-fix statement is false in precisely the way DEF-046 exists to punish: it says the proof is "recorded in 02_Work/scratch/V5_negative_tests_business_layer.md rather than asserted in this row", and that file was last written at 07:37, before the fix, still ends "ALL CHECKS PASSED — 42 of 42", and contains no mention of a multi-block test.

Did the V9 fixes hold?

Attacks on the checkers

Every attack was run in a full disposable copy of the tree under /tmp/v10. Nothing was done inside the project.

C41 — held. The multi-block test exists at run_checks.py:1292-1300 and V8's decoy-block attack now fails closed with a named error. This is the one fix of the round that is what it says it is. Two things sit underneath it that no attack is needed to see. First, IV_SOURCES["v2.0.0"] still points at 02_Work/scratch/V7_baseline_v2_reverification.md. V8 and V9 both ran against this same baseline, both recorded REQ-SYS-03=FAIL, both recorded REQ-BUS-01=PASS, and neither was transcribed. So 03_Registers/REQ.csv, 00_Handover/REQ.csv, the Dashboard tile "Requirements verified 30/32", the study's DoD gate 1 ("FAIL — 30 of 32") and C31's headline all publish REQ-SYS-03 = Verified on a verdict two later independent passes overturned in writing, and publish REQ-BUS-01 as still failing when two later passes cleared it. C41 tests staleness against a baseline; it has nothing to say about a later pass against the same baseline, so the builder chooses which verifier counts. This is V9's finding verbatim and it was not addressed. Second, no verification report is hash-gated — SOURCE_MANIFEST.json tracks 20 files and none of them is a verifier's report — so the block below this line can be edited without any check noticing.

C43 — defeated. Method above. The fix replaced a substring search with a table read, which is a real improvement, and then threw the improvement away by keying the results on the header row and comparing by membership. Two lines of code fix it: key on the table's document index, and compare the n-th matrix to the n-th study. As it stands, appending one duplicate table to the end of a 50-table document lets any WEIGHTED SCORE row in the body say anything. C43 also still never opens DEC.csv, never checks that an alternative is described where it is scored, and never tests a sensitivity statement — it would not have caught any of DEF-042's three non-arithmetic halves.

C44 — defeated on diagrams. _re.finditer(r"D1 to D(\d+)", low) where low is lowercased. The pattern is dead code. I proved it by setting both Help Hub occurrences to "D1 to D3" and getting 44 of 44. The (\d+) diagrams pattern is alive, so "3 diagrams" would be caught — but the failure DEF-044 actually records is a range, not a count, and the range check has never been able to fire. The spelled-out sweep (WORDS × ("slides","diagrams")) works and did fix the "ten slides" occurrences: 03_Registers/FAH.csv now reads "Brief the decision in one short deck" and the Help Hub reads "12 slides", "455 formulas", "82 pages", which I verified against the artefacts myself (pypdf reports 82 pages; python-pptx reports 12 slides; openpyxl counts 455 formula cells).

C12 — defeated. One bit. The check proves no two files are byte-identical and that every file on disk is embedded; it cannot tell whether the picture named D5_trade_study.png is a trade study. I accept that no checker can read a picture — but DEF-048 claims the class is closed, and it is not; the previous attack was blocked and the next one is one im.save() away.

C34 — defeated twice, both unchanged from V9. (a) White-on-white. I injected @media print{ .col:nth-child(2) *, .col:nth-child(3) *, .verdict *, footer * { color:#fff !important; background:#fff !important } } and got C34 PASS: "exactly 1 page, 420mm x 297mm, and all 12 sentinels … present. Fitted, not clipped." I then printed that file in Chromium myself, rasterised it and looked at it: ink coverage fell from 0.2492 to 0.1350, and columns 2 and 3, the value-validation verdict block and the footer are blank paper. The check reads the PDF content stream; the requirement is about what a person can read. (b) Wrong content. I replaced all 97 dollar figures on the A3 page with $999,999, re-gated the manifest and ran the suite: ALL CHECKS PASSED — 44 of 44. Nothing in 44 checks compares a number on the A3 page against the twin it is declared to be a projection of; C33 validates the JSON and compares mtimes and never reads the HTML.

C42 — defeated on the workbook. The register-content half is real: I confirmed in a sandbox that truncating DEF.csv and touching it produces a named failure. But C42 regenerates the CSV projections and applies only a clock test to binaries. I replaced 03_Registers/Project_Registers_v2.0.xlsx — CI-004, the delivered register set — with the v1.0.0 NDIS_Registers.xlsx (28 sheets, 26 requirements instead of 32) and touched the CSVs: ALL CHECKS PASSED — 44 of 44. Only C14 objected before the touch, and C14 is a clock test, not a content test.

The negative-test ledger is stale. 02_Work/scratch/V5_negative_tests_business_layer.md, mtime 07:37, is the project's record of GOV-E2.4 and is cited by DEF-046 and implied by DEF-048. It records ten proven-fallible checks, ends "ALL CHECKS PASSED — 42 of 42", and contains nothing about C41's multi-block test, C43's table read, C44's sweep or C12's hashing. Four checks were written or rewritten in this round to close a Major defect apiece. None of them has a negative test on file, and I have just shown that three of the four cannot fail on the defect they were written for.

Verdict table

Findings that will remain open

Ordered by what they cost a reader.

F-V10-1 (Major). REQ-CON-04 fails on the delivered register workbook, and the check written to enforce it cannot see the file. 27 nine-point cells and 27 wide sheets frozen without a label column, in 03_Registers/Project_Registers_v2.0.xlsx — a Product-baseline configuration item that the study, the Help Hub and the Dashboard all send the reader to by name. C29 walks 05_Outputs only, so it reports zero on a subset and the definition of done quotes that count as proof. Cost to a reader: they open the register set on a phone, scroll right to read a risk rating or a defect status, and lose the ID column — which is exactly what the requirement exists to prevent, and the project's own paperwork tells them it cannot happen. Fixing it is small: ws.freeze_panes = "B3" in build_registers.py and a 10pt banner.

F-V10-2 (Major). The project's published requirement status is three verification passes out of date. IV_SOURCES["v2.0.0"] still names V7. V8, V9 and now V10 have all run against this same baseline. REQ.csv (both copies), the registers workbook, the Dashboard tile, the study's DoD gate 1 and C31's headline all publish REQ-SYS-03 = Verified (V8 and V9 both recorded FAIL) and REQ-BUS-01 = Open, still failing (V8, V9 and I all record PASS). C41 tests staleness against a baseline and has nothing to say about a later pass against the same one, so the builder picks which verifier counts. Cost to a reader: the two numbers on the front of the project — "30 of 32" and "FAIL — 30 of 32" — are about neither the artefacts as they stand nor the most recent verdict on them.

F-V10-3 (Major). C12, C43 and C44 are all defeatable on the defects they were written for, and none has a negative test. Details and reproductions above. DEF-048 asserts all three were "proven able to fail by rerunning V9's own attacks"; the file that would hold that proof is an hour older than the fixes and still ends "42 of 42". Under GOV-E2.4, as the project states it, these are not checks. Cost to a reader: the wrong trade-study totals, a falsified diagram range and the wrong picture in the study all pass a 44-of-44 green suite, and the suite is what the study's Appendix A and B cite as evidence.

F-V10-4 (Major). C34 still certifies a page two thirds of which can be blank paper, and no check compares a figure on the A3 page to the twin. White-on-white print CSS: 44 of 44 with columns 2 and 3, the verdict block and the footer invisible — I printed it and looked at it. All 97 dollar figures replaced with $999,999: 44 of 44. Both raised by V8 and V9. Cost to a reader: REQ-BUS-01's only mechanical guarantee is that a PDF has one page and twelve strings somewhere in its content stream.

F-V10-5 (Moderate). DEF-046 and DEF-047 each close on a false statement about their own evidence. DEF-046: "the proof is recorded in 02_Work/scratch/V5_negative_tests_business_layer.md rather than asserted in this row" — it is not in that file, and the file has not been written to since 07:37. DEF-047: "the record is now a configuration item" — CI.csv has 30 rows and none of them is T1_trade_study_weights.md; I deleted the file and got 44 of 44. This is the third consecutive round in which a defect record describes a change that was not made — DEF-043 did it about C41, DEF-046 was raised for that, and DEF-046 has now done it about itself. Cost to a reader: the defect register is the project's account of its own reliability, and parts of it are not checkable against the tree.

F-V10-6 (Moderate). The requirement amendment is attributed to a change record that does not record it. CR-007's "What changed" field describes the two-business re-baseline and says nothing about REQ-CON-04; its "Affected REQ IDs" field is prose. DEC-008, REQ-CON-04's rationale and DEF-049 all cite it. Amending an acceptance criterion inside an existing closed-scope Class 1 record rather than raising a new one is the specific manoeuvre GOV-D2 change control exists to make visible. Cost to a reader: they cannot find, from the change register, when or under whose approval the requirement's meaning moved.

F-V10-7 (Moderate). The archive and legacy-input story is still partly false, and the fix created a new count discrepancy. The v1.0.0 zip's 04_Inputs/ is empty, so DEF-001's "archived to 06_Archive/_versions", study §2's repetition of it, Appendix B's certification of compliance lines 11-16 on it, and the Dashboard's "06_Archive/_versions/ — you want the four superseded working files" are all still wrong about where those files are. 04_Inputs/legacy_unsourced/ holds five files; DEF-001, study §2, the README and the Dashboard say four; DEF-050 says five. CR-006 names NDIS_Project_v1.0.0_delivered_20260820.zip; the file is …_final_…. Cost to a reader: two of the project's recovery and provenance instructions send them to the wrong folder, and the first thing they see when they follow one is a file count that disagrees with the document that sent them.

F-V10-8 (Moderate). C42 does not read the register workbook's content. I replaced CI-004 with the v1.0.0 workbook — 26 requirements instead of 32 — touched the CSVs, and got 44 of 44. The CSV half of C42 is genuinely real; the binary half is a clock. Cost to a reader: the one artefact in the project that presents all 27 registers in one place is unprotected against substitution.

F-V10-9 (Moderate). The daily audit log is still the day-one entry, and a compliance gate is still certified on an entry that does not exist. 01_System/daily_audit_log.md has one entry, headed 2026-09-07, describing the 20 August work: "26 requirements", "53 sourced claims and 15 assumptions", "DEF-001 to DEF-003", "All 26 REQ. All 53 SRC. All 15 ASM.", "this is the first cycle", and — inside a single line — "14 actions owed by Zaid (ACT-001 to ACT-008)". The project has 32 requirements, 74 sources, 31 assumptions and 50 defects. It does not mention Part II, aged care, DEC-006 to DEC-008, CR-004 to CR-007, or DEF-004 to DEF-050. Its own breakdown of "49 items" sums to 44. And study Appendix B lines 38-41 are certified PASS on "Daily audit entry for 2026-08-20 in 01_System/daily_audit_log.md". There is no 2026-08-20 entry. Raised by V8 and V9; untouched twice. Cost to a reader: GOV-H3.8 says the first action of the next working day is to read the last entry and clear what it left open. The last entry is about a different, smaller project.

F-V10-10 (Minor, but it is the fourth pass). The $16.10 citation and the $8,757 citation are still wrong. Study finding 5 and §15.1 cite SRC-068 for the direct-services margin; SRC-074 was created for that figure and reaches neither. Table 5.1 cites "Financial Model, SupportCoordination sheet" for $8,757; the figure is in no cell of the workbook. Both raised by V6, V7, V8 and V9. Cost to a reader: the two citations they are most likely to follow — the number that carries REQ-AC-04, and the runway for the option the study recommends running alongside everything — resolve to places that do not hold them.

F-V10-11 (Minor). Assorted, each verified individually. DEC-004's Rationale field says "A3 is eliminated" one field below an Alternatives field where A3 is the support-coordination option. The callout box is headed "THE ALTERNATIVE THAT WAS SCORED AND SHOULD NOT HAVE BEEN" over a body reading "NOT scored", and repeats the A4 bullet verbatim. §6 still says "Both trade studies" when there are three. Recommendation 4 still says July 2027 is "about eleven months away"; from the currency date it is ten. business_plan.json and the A3 page still carry verified_by: "PENDING — independent verification pass V5 has not yet run against this plan". Both TS1 flip points are exact ties presented as hand-overs. Section 20's source table is captioned "Table 14.1", the same number as section 14's settled-position table. EVD-027 to EVD-032 have now been renumbered from "PENDING V6" to "PENDING V8" to "PENDING V10" without ever being resolved. C29 still reads only 148 of the deck's 194 runs because its .pptx branch iterates sh.has_text_frame and a table shape has none — nothing sub-10pt is hiding there today, but the check cannot see it. C29's exclusion set is still {basename(ci[4]) for ci in P.CI if ci[2].lower() == "zaid"}, a builder-writable owner column: change one cell and any artefact leaves the check's scope.

What I could not verify, and why

• Whether any external fact is true. I verified internal consistency, derivation and citation, not the world. I opened none of the 74 source URLs. Nothing here is evidence that $103.11, $73.58, $45.28, $43.03, $10,731, $78,106, $63.85, $47.75 or the July 2027 date are correct. Fifty-one of 74 sources sit below High confidence with a stated reason, which is honest rather than verified.

• Whether the trade-study judgements are right. I proved the arithmetic of all three studies reproduces and that the sensitivities do. Whether A1's capital deserves a 2 and A2's a 5 is a judgement no verification pass can settle, and the project is right to escalate DEC-003 and DEC-007 rather than decide them.

• Whether trade study 1's weights really were fixed before its scores. The record now exists and says plainly that it does not prove this. I can confirm the file is honest; I cannot confirm the ordering, and neither can anyone else.

• Whether the new A3 is the best available third alternative. It is lawful, credible and costed from figures I traced. Whether staged registration, a partnership or contracting to an existing provider would have been stronger is not something I can test. I note that A3's distinguishing feature — running support coordination during the window — is also what the study's recommendation 3 tells the owner to do under A2, which makes A3 less independent of A2 than a trade study wants.

• Print behaviour on any engine but Chromium. I measured with headless Chromium via Playwright, which is what C34 drives, rasterised the output at 60 dpi and looked at it. I did not test Word, Safari, Firefox or a physical printer.

• Whether the Decision Pack renders as designed. I read every run, every table cell and every font size through python-pptx. I did not render the twelve slides, so I cannot speak to overflow, overlap or contrast.

• Eleven of the twelve diagrams. I read D5's underlying matrix and confirmed all twelve as files on disk and as content-hash-matched embeds. I did not view D1, D2, D3, D4, D6, D7, D8, D9, D10, D11 or D12 as images this pass — and, as F-V10-3 shows, nothing in the suite would notice if they were the wrong pictures.

• Whether my own verdicts survive transcription. C41 now compares this file's machine-readable block line by line against what the project transcribes and refuses a second block, which is worth something. But this file is not hash-gated, no verification report is in SOURCE_MANIFEST.json, and the project is at present transcribing a report three passes old while leaving IV_SOURCES untouched. Nothing mechanically prevents a builder from editing the block below, or from never reading this file at all.

Closing verdict

31 of 32 requirements PASS.

REQ-SYS-03 is closed, and closed properly. The arithmetic reproduces, §6.1 introduces the alternatives that are actually scored, DEC-004 publishes 0.60 and 37 per cent in all three delivered copies, and T1_trade_study_weights.md is the most honest document in this project — it says what it cannot prove, cites the rule that gives a builder's word no weight, and does not backdate itself. That is what a fix looks like.

REQ-CON-04 is not closed, and the way it failed this round is worse than the way it failed last round. The project widened the requirement's scope from a flat six-column rule to "every delivered Office artefact", correctly, on reasoning I checked and agree with — and then left the enforcing check pointed at a single folder, so the second delivered workbook fails both halves of the new criterion and reports zero. C29's own docstring names that failure: "A green check over a subset is worse than no check, because the definition of done cites its counts as proof." The instrument that closed DEF-028 has committed DEF-028.

Read the rest before the count. Five of the six checkers I attacked fell, four of them to attacks V9 published in writing three hours earlier, and I put V8's exact wrong trade-study totals back onto the face of the delivered study and got 44 of 44 green from the check written specifically to stop that. The pattern this project has repeated through V6, V7, V8, V9 and now V10 is not that it fixes things badly — the DEF-042 work was excellent — it is that it writes a checker against the instance it was shown and then records the class as closed. C43 was told about a substring bug and now reads tables, keyed on a header. C12 was told about byte-identical copies and now hashes. C44 was told about three regexes and now has a dozen, two of which are case-mismatched and can never fire. Each closes the demonstrated attack and nothing adjacent to it. And in three consecutive rounds a defect record has asserted a fix or an evidence file that does not exist — DEF-043 about C41, then DEF-046 about its own negative test, then DEF-047 about a configuration item. When the register that records the project's reliability is itself unreliable in the same way three times running, that is the finding, not the individual defects.

Finally, and independent of any of this: the delivered registers still publish requirement status derived from verification pass V7. Two independent passes have ruled since, both against this same baseline, and neither was transcribed or refused in writing. The project is shipping a "30 of 32" that is not the verdict of the most recent verifier, and it will ship this one the same way unless someone moves IV_SOURCES.

Left behind: nothing. Every attack ran in /tmp/v10/* sandboxes; 01_System/checker_run_log.txt still carries the builder's 08:50:09 run and 01_System/SOURCE_MANIFEST.json is still gated at 2026-09-07T08:50:09. find . -newermt "2026-09-07 08:51" -type f returns nothing but this file, and find . -name __pycache__ -o -name "*.pyc" returns nothing. python 01_System/tidy.py --report returns TIDY PASS — placement map satisfied: no loose root files, no debris, no secret stores, no links to live files, no unknown folders, no missing mandated folders. V9's orphaned BUSINESS_PLAN_A3.html.probe.pdf is gone — C34 now deletes its print probe in a finally block. This report is the only thing I added to the tree.

MACHINE-READABLE VERDICTS

REQ-SYS-01=PASS
REQ-SYS-02=PASS
REQ-SYS-03=PASS
REQ-SYS-04=PASS
REQ-SYS-05=PASS
REQ-SYS-06=PASS
REQ-SYS-07=PASS
REQ-SYS-08=PASS
REQ-SYS-09=PASS
REQ-SYS-10=PASS
REQ-SYS-11=PASS
REQ-SYS-12=PASS
REQ-SYS-13=PASS
REQ-SYS-14=PASS
REQ-SYS-15=PASS
REQ-SYS-16=PASS
REQ-SYS-17=PASS
REQ-SYS-18=PASS
REQ-CON-01=PASS
REQ-CON-02=PASS
REQ-CON-03=PASS
REQ-CON-04=FAIL
REQ-CON-05=PASS
REQ-CON-06=PASS
REQ-MOE-01=PASS
REQ-MOP-01=PASS
REQ-AC-01=PASS
REQ-AC-02=PASS
REQ-AC-03=PASS
REQ-AC-04=PASS
REQ-BUS-01=PASS
REQ-WEB-01=PASS