CloudWorkz Limited v its former directors

How the litigation system works, and how every part fits together

One matter, handled as one strategy with several recipients. Evidence sits in a typed store, instruments are written in a vault, a ladder and a gate test the work, runs and subagents do the heavy lifting, an independent review audits the lot, and only Tom King signs and sends.

State as at 5–6 October 2026 Review project commit c3955ae · 5 Oct 2026 20:55 +0200 Page generated 6 Oct 2026 15:43 CEST (13:43 UTC) Not counsel-reviewed · Inward · Not for dispatch Built from the review project files at that commit, not the live store

1 · The whole system

Five layers and one human gate

The system is easiest to understand as a pipeline that records facts, turns them into documents, tests the documents, and then stops at a person. A second, independent loop (the Review Project) reads everything without changing it and reports what the pipeline cannot see about itself.

Facts enter on the left, become documents in the middle, are tested by the control layer, and stop at Tom King. The dashed loop is the part most systems lack: an independent reader that reports back without touching the matter.

1,114exhibits registered (1,108 held)

690decision rows, last D-693

56live instruments cited across the matter

1,208exhibit citations, 320 distinct exhibits

15validation rungs

67gate checks

461review findings, 126 dispatch-blocking

1 of 53instruments frozen (the one already sent)

The four ideas that make it work

One strategy, many recipients.

The matter is not Ian Wolter alone. It covers Andrew Hobsbawm, Heather Salway, William Nicholl, the shareholders, Crowdcube and the wider legal strategy, and every review treats it as one whole so that letters cannot contradict one another.

Registers are claims, not evidence.

A filename, a register description or a status cell can all be wrong. Where a finding turns on what a document says, the primary is opened and quoted. This rule exists because of real errors (E-620, E-065, E-084).

Decisions are the owner's and are tracked.

Every decision is a numbered D-row. Later work cites it, cannot silently override it, and may only record a "CHALLENGE TO D-xxx" with reasons. Landing markers then test whether a decision actually appears in the documents.

Approval is bound to bytes.

A signature means "I approve this exact text, sha256 …". If the text changes, the hash changes and the approval no longer applies. Only Tom King signs and only Tom King sends.

2 · The matter

What the system is protecting: one claim, ten review lanes

CloudWorkz Limited has pre-action claims against its former directors. Tom King is the only founder and, since February 2024, the sole active director. Each lane below has its own instruments, evidence set and review folder. Pick a lane to see its role and state.

Who the recipients are

Person	Role in the matter	Pack	Posture
Ian Wolter	Executive Chairman, principal target; resigned as director 8 Feb 2024	PACK-1	Core letter before claim, 21 days to respond, 7 to acknowledge, plus schedules and the Deed of No Personal Benefit
Andrew Hobsbawm	Former non-executive director	PACK-2	Lane currently marked DO NOT SEND; only Tom King can release it (D-614)
Heather Salway	Former director; Ian Wolter's sister	PACK-3	Letter must state claim and relief sought (decided 2 Sep), 14 days
William Nicholl	Former director	PACK-4	14 days; invoices reconcile to £26,448.12
Shareholders, Crowdcube	Deed outreach, buy-back facilitation, members' report	PACK-5/6/7	Deed names no one and states no figure; Crowdcube asked to invite every underlying investor to sign or give reasons

Standing rules set by Tom KingNo outward letter states the Company's own weaknesses (D-274). No counsel before settlement (D-214), so the NOT COUNSEL-REVIEWED banner stays and is never reported as a defect. "The Founders" in the plural is wrong everywhere. Director status turns on function, not on filing dates (Tier 0 rule).

3 · Where things live

One home for each kind of thing

Documents live in the vault, the typed record lives in the matter store inside the same vault folder, and the primary originals never leave the LaCie drive and the mailbox exports. The rule since 6 September (D-262) is that everything sits in one location, with only markdown going to Obsidian.

What	Where	Role	Write rule
Matter store (Wolter)	`Litigation/Lexi-Matters/cloudworkz-v-wolter`	evidence.sqlite, exhibits.yaml, DECISION-LOG, FINDINGS-LEDGER, acceptance/*.yaml, tools/, HANDOVER, CLAUDE.md	Working record; writers gated by read receipt (RULE 1B)
Hobsbawm store	`Lexi-Matters/cloudworkz-v-hobsbawm`	Separate register for the Hobsbawm lane	Disagrees with the Wolter store on some facts; 41 of 62 exhibit ids collide
V9 instruments	`Litigation/Wolter/V9-source`	The 69 canonical instruments (letters, schedules, deeds, RFIs, scripts) and their sidecars	Single authoring and approval surface since 31 Aug; generator pipeline retired until sign-off
Primary evidence	`/Volumes/LaCie/Litigation-Cloudworkz-v-Wolter`	Mail exports (about 30 GB for Tom King's mailbox), bank statements, certificates, Google Takeout, 74,595 files, 262 GB	Read-only. Binaries never go into Obsidian
Backups	LaCie `Mac-cleanup-2026-10-04/`	About 2,500 .bak and pre-image files moved off the Mac by Tom King's own cleanup	The LaCie copy is currently the only copy
Review Project	`CloudWorkz-Review-Project/`	The independent audit, its findings and the human-review folders	Writes only here; matter and LaCie are read-only
Advisor master document	`wolter-advisor-review/`	29 July compiled export for an outside reader, seven sections	Snapshot, now superseded by later work

The Lexi engine behind the store

Lexi is Tom King's own tool for assembling structured litigation packs. All privileged matter content stays on his machine and is never written to the code repository or the cloud database, which is schema-restricted to login and billing. The legal-discipline rules live in one Python engine and are called through a defined interface, not re-implemented in the web layer.

The engine exposes named operations: validate a bundle, run discipline checks, build a manifest, assemble a pack, a cold-read pass for human review, dispatch, an audit trail and a final acceptance gate reduced to one yes/no against eight criteria.
Fail-closed: if an AI step fails or is flagged, the system stops, falls back to a deterministic path and flags a human. Raw AI output never lands in a legal document without a human review step.
Check-sync compares the hand-kept exhibit register with the database and refuses to overwrite the register when the database looks unpopulated, so a naive "--fix" cannot wipe hundreds of exhibits.
Of the July advisor snapshot: the test suite collected 603 tests against a stated 574, a small example of a governing document going stale.

4 · Evidence layer

From a file on the LaCie to a quoted exhibit in a letter

Every statement in an outward instrument has to trace back to a held primary. The evidence layer is the chain that makes that checkable.

A document's identity comes from its hash, not its filename. Atoms are mined first and numbered later by a human, so a wrong filename cannot silently become an exhibit number.

What the layer guarantees

All 2,080 held primaries resolve with a full sha256 match at the 4 Oct baseline; 1,108 of 1,114 exhibits are held.
Of 203 exhibits cited by the Wolter letter, all are held; 151 spans were tested and 150 were verbatim, 1 partial, 0 contradicted.
Quotations are tested mechanically; registers are never accepted as proof of what a document says.
Receipt tables record who registered, retagged or corrected what, and when.

What it cannot yet guarantee

37 of 1,114 exhibits point at a different document than the latest registration receipt; 13 are cited and six are genuinely wrong (E-020, E-027, E-039, E-244, E-296, E-547).
Register prose reached outward text three times (E-620, E-065, E-084).
No litigation hold exists on any Company system, and the Wolter Takeout of June 2026 has thin custody records.
130 cited exhibits have no LaCie path; of 477 that do, 102 do not byte-match (46 cited outward, mostly renderings).

Working rule from 2 SeptemberA filename is a claim to be tested, never evidence. A repair pass settled 81 exhibit ids and found misfilings that ran in displacement chains up to five rows deep.

5 · Decision layer

Every choice is a numbered row, and later work has to respect it

The decision log is the memory of what Tom King has decided. It is what stops a new session from reopening settled questions, and what lets an auditor say whether a letter reflects the decisions in force.

DECISION-LOG.md

690 rows (D-001 to D-693, with two undocumented gaps at D-515 and D-519). The canonical log is the one in the live matter store; the earlier vault fork collided on D-numbers and was retired (2 Sep).

FINDINGS-LEDGER.md

About 1,705 F-rows: defects found, their status and owner. Of 485 rows tagged open or for-Tom, 116 are stale and 131 are already answered, so only about 30 are real judgements.

Authority hierarchy

A settled-questions index and a ranking of which document controls when two disagree (Tier 0 key rules, canonical strategies, decision rows, then registers). Held in the Wolter parent folder, which has no git copy.

Landing markers: turning a decision into a test

A marker says "this phrase must be present (or absent) in this file". The decision-landing check reads the markers. Where a marker existed, the review found the decision in place; where none existed, it found four decisions missing or reversed.

101live marker sets

76base decisions covered

D-336coverage starts here by design

0markers for D-542 to D-679

Known weaknessThe gate reads "clean" over roughly 614 decisions it cannot see. Supersession is one-way: a later row records the change but the earlier row's status cell is not updated (for example D-690 still reads LANDED though D-692 withdrew its £250,000 exclusion).

6 · Instruments and sign-off

How one document travels from draft to sent

An instrument is any document that leaves the building or binds someone: a letter before claim, an RFI, a deed, a members' update, a video script. Each has a code in the document index, a sidecar, a chain of dependants and an approval state. Step through the lifecycle below.

The five approval controls

Five separate controls assert that an instrument is ready. They overlap, and only some are tests rather than statements.

Control	What it asserts	Type	Verdict of the review
Sign-off sheet	27 lines: read, conformity, exact-copy approval (line 24)	Prose	Keep Line 24 should carry the freeze hash
D-190 alignment block	Decisions implemented, evidence rested on, gates and conduct limbs	Prose	Simplify Generate a current-state line; prose goes stale
Freeze (signoff.py)	sha256 of approved bytes; a break travels down the document chain	Test	Keep and extend Only 1 of 53 frozen
Read receipt	Four controlling documents declared read before a register write	Test (4 writers)	Keep Add a gate row and version hash
Landing markers	A phrase present or absent in a named file	Test	Keep, extend coverage Caught live defects
Alignment protocol, pre-reliance checklist	Five-step and eleven-dimension passes	Prose	Retire Never run on any send

Where the controls have caught real errorsA senior investor's age stated in an outward letter (removed through a marker), four reversed decisions at D-671, a sidecar overstating a settlement by £766, a citation defect in the Wolter letter.

7 · The run system

How work is actually executed, session after session

The matter is too large for one conversation, so the system is built around handovers, recorded runs and parallel workers. Three kinds of run exist.

The shape of a worker wave

The rules at the centre were learned from failures: three session-limit incidents killed 16, 10 and 5 agents mid-run, which is why waves are capped at five, every worker writes a skeleton first, and every task prompt is saved so a relaunch is one line.

Model tiersTom King's standing rule is to delegate simpler work to cheaper models. The most capable models do judgement, register-sensitive drafting and legal reasoning; Sonnet-class models do searches, batch reads, inventories, extraction and check runs. The review followed this: fable and opus for pack reviews, law, red teams and verification; sonnet for decision sweeps, bundles, open items and the human-review folders.

8 · Rungs and gate

The two machines that test the work

The ladder asks whether each layer of the matter (rules, decisions, evidence, claims, strategy and so on) has been assessed against the current bytes of its inputs. The gate runs 67 mechanical checks across the instruments and registers. Both were audited on 5 October.

What the review found, at commit c3955ae

Computed from 01-FINDINGS.csv (461 rows) on 6 October. The ladder and gate figures further down come from the framework review and were not re-derived in this update.

461findings (9 critical, 131 high, 201 medium, 120 low)

126marked dispatch-blocking (335 not)

251defects (plus 131 gaps, 40 challenges, 34 enhancements, 5 redundancies)

6new Articles findings AR-001 to AR-006 (one high, four medium, one low)

Findings by recipient	Rows	Findings by recipient	Rows
Wolter	97	Members	45
All recipients	84	Salway	39
RFIs	67	Nicholl	38
Hobsbawm	47	Settlement, Crowdcube	22 each

Two findings moved on 5 OctoberSA-001 (the Articles require two directors) fell from critical to medium and is no longer dispatch-blocking. LN-109 (the board minute against a two-director quorum) fell from high to medium and is no longer dispatch-blocking. The nine critical findings that remain are P2-001 to P2-004, LN-301, LN-302 and RW-001 to RW-003. LN-301 is resolved as a decision by Tom King's answer H1a (below) but stays in the file until he records a D-row.

The 15-rung ladder

A rung's verdict is typed by a session into layers.yaml. The code can only downgrade it: to COUNTERFEIT if a check fails, or to STALE if the sha256 of the declared inputs has changed. Hashing is by content since D-428.

12satisfied at HEAD

3stale at HEAD: decisions, claims, strategy

6partly met on the stricter test

9not met on the stricter test

#	Rung	What its checks test	At HEAD	Review verdict

Why "15/15" was wrongA commit titled "rungs 15/15" was made before two later edits to the decision log and defences register. Recomputed with the store's own hash function, three rungs were stale, so the ladder read PLAN BLOCKED. The top rung, gates, has no inputs and no checks, so its hash is the hash of nothing and it can never be anything but satisfied.

The 67-check gate

Two judges exist: gate_runner.py records a run and preservice_gate.py (PG) decides what a result means. They disagree. On a fresh full run PG blocked on 10 checks while the handover said "3 non-zero" and a decision row said the gate was clean except for Tom King's items.

5fully met

49partly met

13not met

34with no usable recorded meaning

Met 5Partly met 49Not met 13

#	Check	Verdict	Why, and the smallest fix

Dispatch readiness: what nothing yet measures

The review's central finding about the machinery is that it reports on itself, not on what will leave the building. It proposes one joining check, check_dispatch_ready.py, with seven tests. The matrix shows each pack against them as of 5 October.

The seven tests

a Frozen with a masked hash
b Review PDF text is clean (no placeholders, no inward banners)
c PDF is complete (every enclosed schedule present)
d Dispatch manifest row matches current source
e Day-of checklist cleared, with who and when
f Sidecars, markers and quotations clean for that file
g Enclosure bundle exists and matches the held primaries

Two sends have already happened

Both went round the release path. The direct-investor email was sent with nine edits made after the session's last version (D-546), and the deed review copy was recorded five days after it was sent (D-681). The as-sent intake proposed by the review would hash whatever Tom King actually sends.

Pack or lane	a Frozen	b/c PDF	e Placeholders	f Sidecar, markers	g Enclosures	Blocking	State
PACK-1 Ian Wolter	Fail	Fail	Fail	Part	Fail	8	Not ready
PACK-2 Andrew Hobsbawm	Fail	Pass	Part	Part	Fail	12	Not ready; DO NOT SEND under D-614
PACK-3 Heather Salway	Fail	Pass	Part	Part	Fail	11	Not ready
PACK-4 William Nicholl	Fail	Pass	Part	Part	Fail	8	Not ready
RFIs (13 + SJP)	Fail	Pass	Part	Part	Fail	21	Not ready
Crowdcube	Fail	Fail	n/a	Part	Fail	13	Not ready; D-317 sign-off not obtained
Members update	Fail	Part	Fail	Part	n/a	6	Not ready; follows Crowdcube
Video 1	Fail	Fail	n/a	Fail	n/a	5	Not ready; cannot record against a deed version that does not exist
Settlement deeds	Fail	Fail	Part	n/a	n/a	13	Not ready; later stage by design
DocuSign signing version	None	n/a	n/a	None	n/a	1+	Does not exist as an instrument yet
Direct-investor email (DIE)	Frozen	n/a	n/a	n/a	n/a	0	Sent 14 September; the only frozen instrument

Failing test (a) is expected while drafting continues: the ordering rule (D-557) says anything that will still move must move before anything is frozen. The defect is that nothing shows the other six tests.

9 · The audit

The Review Project: read everything, change nothing, prove it

On 4 October Tom King commissioned a complete review of the whole matter (brief v6). It ran as a separate project folder with hard rules: the matter folders and LaCie are read-only, nothing is sent or uploaded, only Tom King sends, and questions never stop the run but go to a questions file.

Three goals, in order

1 · Nothing is lost

Every document, decision, evidence item, open item and reason is found, indexed and carried forward. 17,042 files were indexed; 91,637 coverage rows record what was read in full, in part or not at all.

2 · Ready for human review

A cover note, unchanged text with hash, redlines, evidence table, enclosure review, consistency check and sign-off page for every instrument.

3 · Forensic root-and-branch review

Framework, systems, evidence, logic, claims, causation, quantum, law, strategy and conduct risk, handed back as numbered files 00 to 99.

The six phases

Phase	Content	Status
0	Set up the project, prove access (PASS 25, WARN 5, FAIL 0), freeze git baselines and a sha256 manifest of 17,042 files	Done 4 Oct
1	Eight project skills: matter orientation, evidence verification, instrument review, claims/causation/quantum, law check, conduct gate, enclosure test, context keeper (v0, then v1 with changelogs)	Done 5 Oct
2	Nothing lost: master index, context map, decisions map (690 rows), open items (142), orphans and risks (270 rows)	Done 5 Oct
3	Subagent reviews by pack, area and lane; two red teams; law checked online; Companies House; adversarial verification	Done 5 Oct
4	`human-review/`: one folder per instrument across ten lanes, reading order, questions, evidence bundle index, chronology, PDFs	Done 5 Oct
5	Handback: executive summary, findings CSV, scorecard, claims/causation/quantum, strategy, red teams, framework, enhancements, enclosures, coverage, integrity	Done 5 Oct

What the audit found

461 findings by severity

Critical 9High 131Medium 201Low 120

126 are marked dispatch-blocking. Types: defects, gaps, redundancies, enhancements and challenges to recorded decisions.

How findings are written

Every row carries an id, type, severity, confidence (VERIFIED against the primary, PART, or INFERRED), recipient, file and line, a quoted piece of evidence, why it matters, the smallest fix and its owner, the decision affected and whether it blocks dispatch.

Independent checks on the audit itself

Adversarial verification

Every critical finding and every challenge to a decision went through a sceptic pass: 48 survived, none refuted. Severity was corrected where needed (for example the 35-page PDF finding went from critical to high).

Red teams

Two teams worked only from what each recipient would hold: one as Ian Wolter (25 findings, including enclosed warnings that turn into a threat against Tom King), one as the non-executive directors (17 findings, including the unpaid £2,341.97 judgment and Tom King's own 11 February 2024 email to William Nicholl).

Question audit (QAUDIT)

Tom King asked whether every question put to him was genuinely unanswered. Four sonnet agents checked each against the decision log, ledger, settled facts and his chat answers. Result: most were partly or wholly answered already, so the punch list asks only what is still open (69 questions: 49 before dispatch, 20 can follow) and lists 13 as already answered.

Integrity proof

A sha256 manifest before and after shows 0 files added to or modified in the matter by the review. 130 baseline paths absent from the Mac were Tom King's own cleanup, all .bak pre-images, all present on the LaCie (25 of 25 re-hashed matched).

The audit corrected itselfThe review's first critical finding (SA-001, that the Articles required two directors) rested on the Articles of 26 January 2022. Tom King pointed to the SeedLegals folder; the April 2023 Articles allow one director (Art 24) and make one quorate (Art 27.1). SA-001 fell from CRITICAL to MEDIUM and is no longer dispatch-blocking. 122 citations were mapped across both texts (ARTICLES-CROSSWALK-2022-v-2023.md), 13 root files were corrected, and the cover note, sign-off, consistency and evidence table of every human-review instrument now open with the line "Articles correction 5 Oct 2026". The lesson recorded: non-filing at Companies House does not show non-adoption, and when a primary has several dated versions, read all before reporting the operative one. See section 12.

Scorecard of the control framework

Component	Met	Partly	Not met	Comment
Rungs (15)	0	6	9	Strongest: evidence and assertions. Weakest: gates, plan, loss, merits
Gate checks (67)	5	49	13	12 checks cannot exit non-zero; 4 do not measure their names
CLAUDE.md rules	2	9	10	Rules claim "mechanical" where only four writers are gated
Skills and skill rules (SR-01 to SR-05)	All partly met; two retire as duplicates, two need an owner decision			No skill receipt exists for the 3 and 4 Oct sessions

10 · Human review

Where the machinery hands over to Tom King

The review does not decide anything. It prepares each instrument so that Tom King can approve, change or hold it in his own time, and it prepares his answers so that dictation is enough.

Seven files per instrument

COVER-NOTE.md what it is, to whom, the decisions behind it, the evidence described in plain terms, strengths and weaknesses, and numbered decisions for Tom
CURRENT-TEXT.md the V9 text copied unchanged; line 1 is its sha256 so exactness can be checked
PROPOSED-CHANGES.md each change as OLD to NEW with reason; reversals of recorded decisions kept apart as challenges
EVIDENCE-TABLE.md each statement, the primary's quote and a status of VERIFIED, PART, INFERRED or CONTRADICTED
ENCLOSURE-REVIEW.md keep, remove, add or redact per enclosure
CONSISTENCY.md against its chain, sibling letters, decisions and deeds
SIGN-OFF.md the unsigned exact-copy approval block

Control documents around them

READING-ORDER order 05, 01, 03, 04, 02, 06, 07, 08, 09 with time estimates (the Wolter folder alone is about 5.5 hours of deciding and 3.7 hours of exact-copy reading)
PUNCH-LIST numbered, voice-readable questions with lettered options and a default; 49 before dispatch, 20 can follow, 2 for counsel, 5 answered by Tom on 5 October, 13 already answered by earlier decisions (not asked), 8 that are Claude's own work. A phone PDF is at human-review/pdf/A3-PUNCH-LIST.pdf
QUESTIONS-FOR-TOM the full protocol version, each item with a "Checked" line
HANDOVER-FOR-HUMANS the matter on one page, recipient by recipient, and how to resume
EVIDENCE-BUNDLE-INDEX 320 cited exhibits and what each enclosure set would contain
CHRONOLOGY 97 rows by function, not by filing date
pdf/ 72 phone-readable PDFs, re-rendered after the Articles correction

How Tom answersHe reads an id, says the letter or YES/NO, then any reason: "B3, A. Keep it, because the deed is already drafted that way." Saying "all defaults" accepts a whole section. Each answer becomes a D-row he records; the review applies the transcript as redlines. Only the items that cannot be settled from earlier answers or the evidence reach him.

Start here: fifteen genuinely unanswered items

The punch list's own "start here" list (as at commit c3955ae). Everything else in Section 1 is only partly open: the earlier decision stands unless Tom says otherwise.

15start here: C2, C6, D3, D5, D6, E1, H4a, H4b, H5d, H5e, H6c, I1a, I1b, J3c, M1a

49before dispatch in total (the 15 are part of these)

20can follow dispatch

2for counsel; Tom decides meanwhile

Tom's order of work: the start-here items, then Section 1, then Section 2. His answers are recorded as D-rows by him; the review does not write to the matter. Four review defaults run against Tom's own earlier decisions and are flagged as challenges, not applied: D2 (D-681), D4 (D-656/657), E4 (D-205, D-673) and G1 (D-232, D-268, D-287).

11 · Sign-off plan

What now stands between the drafts and dispatch

This replaces the earlier plan on one point. "Appoint a second director and pass a members' resolution" is no longer a pre-dispatch step. The April 2023 Articles allow one director and make one quorate, and Tom has decided not to appoint a second (Claudia Chavez was rejected: not independent, and a second director would raise the quorum to two). Everything below depends on those Articles having been adopted, which Tom confirmed in chat on 5 October. The signed pack has not been seen by the review.

Confirm the Articles on SeedLegals. Tom confirms that the April 2023 Articles were adopted (which text, 6 or 26 April, and the date) and keeps the signed pack with the matter file.

Re-cite the article numbers. In K-A27, CC-MIN, K-WS, SH-UPD, K-NBD-R and the four Q1 schedules, following ARTICLES-RECITATION-INDEX.md (41 sentences re-checked so far, 9 of them now wrong).

Tom signs the No-Benefit Deed first. Any recovery on his own claims goes to the Company. The review copy is unsigned, so a signed copy and date are needed. If he intends to vote on the optional redistribution approval, clause 3.3 must change before he signs.

Sole-director minute and written declaration of interests. Drafts 02 and 03 in 10-CORPORATE-CAPACITY are usable as drafts. The written resolution (draft 01) is a fallback only.

Rephrase the Wolter co-founder passage (AR-006). W-LBA lines 236 to 238 and W-Q1 Q(e). Tom agreed on 5 October. This is the one new finding that blocks the Wolter pack.

Remove E-231 from the enclosure lists (H1a). The four Phase-1 RFIs and W-EV section 8. Tom needs to record a D-row.

Each SIGN-OFF.md condition now reads"Confirm the April 2023 Articles and use the 2023 article numbers."

Counsel questions, decided meanwhile by Tom

Whether Tom may vote his shares to authorise his own interest (Art 28.1(h)). The general-law point: s.239 excludes the interested member from ratification, not from prior authorisation, but there is a s.994 unfair-prejudice risk.
The late Companies House filing (no RES01 in 2023): an offence by the Company and its officers, but it does not make the resolution ineffective (ss.26, 29 to 30).
Wolter's Art 25.1(b) right to reappoint himself and its effect on quorum.
The 55% drag-along threshold (it was 75% in 2022) and the changed transfer and co-sale rules (AR-004, AR-005).

Open: not to be marked done

Item	Why it is open	Status
Signed SeedLegals pack	Tom confirms signing in April 2023; the review has not seen the signed pack	Open
Signed No-Benefit Deed	The review copy is unsigned; no date or witness	Open
Companies House filing	No RES01 or new Articles filed in 2023 (history read in full 4 Oct, re-checked 5 Oct)	Open
Which April text was adopted	6 April or 26 April 2023; the provisions Tom needs are identical, the numbering differs by one article	Open
D-285	Not re-read	Open
Calderbank expiry (13 Nov)	Not re-checked beyond D-206(5)	Open

Also still open on the register: the register of members at adoption and whether the special resolution reached 75% of eligible votes (not checked). Separately, the LaCie archives of the 2,509 backup files Tom moved off the Mac on 4 to 5 October are now the only copy; a second copy is still to be made.

12 · Articles and answers

The Articles correction, the superseded references, and Tom's recorded answers

2022 against 2023

Point	2022 Articles (26 Jan 2022)	April 2023 Articles	Effect
Number of directors	Art 23: not less than 2	Art 24: not less than 1, not more than 9	One director is lawful; Model Article 7(2) becomes available
Quorum	Art 26.1: 2 directors	Art 27.1: 1 director, unless there are 2 or more, then 2 including a Founder Director	A sole director is quorate
Conflicts	Art 27 (27.5, 27.7)	Art 28 (28.1(h) members' authorisation, 28.5, 28.7, 28.10)	The machinery is the same; the numbers moved
Votes	Not compared by this page	Ordinary and Vv shares vote; only Deferred Shares do not (Art 7.3)	Every class Tom holds votes
"Founder"	Tom King alone	"each of Ian Wolter and Tom King"	AR-006: the co-founder passage must be rephrased
Drag-along	75% (Art 18.1, 18.13)	55% (Art 19.1, 19.13)	AR-005: counsel to consider

What on this map relied on the 2022 Articles

I searched this page for Art 23, 26.1, 27.5, 27.7, Model Article 11(3) and s.186. None of those references appears in the page text. Two statements on the page did rest on the superseded reading and are corrected and marked below; the sign-off plan (section 11) drops the second-director step. The instruments themselves still cite the 2022 numbers (finding AR-002), which this page reports but cannot change.

Where	Reliance	Status
Map, section 9 callout (the audit corrected itself)	Reported SA-001 as a two-director requirement	Superseded: rewritten
Map, lane 10 Capacity	Described the capacity pack as a route needed because the board was inquorate	Superseded: fallback only
K-A27 (instrument)	Art 26 and 27 of the 26 Jan 2022 Articles; 27.5 and 27.7 for self-authorisation	Superseded: re-cite to 28.5, 28.7, 28.1(h), 28.10; 6 wrong
CC-MIN (instrument)	Assumes a plural board; cites no article number	Wording: recast as a sole-director record
W-LBA (instrument)	Art 26/27 cite; "Founder" passage	Superseded: 27/28; passage exposed (AR-006)
Reasoning: "Art 23 requires two directors" / "Art 26.1 quorum of two"	Behind SA-001, LN-109 and the original executive summary	Superseded (2023 Arts 24, 27.1)
Reasoning: Model Article 11(3) (a board below quorum may do only two things) and s.186	Behind the second-director route	Superseded: neither is engaged when quorum is one
The review's text still carrying the old reading	SA-017, SA-023, LN-107 article references; 02-SCORECARD; 04-STRATEGY section 13.3; 04A Part A section 3.1; QUESTIONS-FOR-TOM; READING-ORDER; CRPD and K-A27 cover notes	Not rewritten: the correction file governs where they conflict

What survives: s.175(6)(a) still means a sole interested director cannot authorise his own conflict, so K-A27 is a declaration, not an authorisation (SA-017 stands on that ground).

Findings from the correction

ID	Severity	What it says
AR-001	Medium	Wolter's right under Art 25.1(b) to reappoint himself after his 8 Feb 2024 resignation. A neutral watch item, not blocking
AR-002	Medium	Every instrument citing the Articles uses the 2022 numbers; K-A27 names the wrong instrument outright. Blocking
AR-003	Medium	No 2023 special resolution or new Articles on Companies House; the only 2023 resolutions are RES10 and RES11 filed 18 Sep 2023. Counsel
AR-004	Medium	The April 2023 Articles changed the transfer rules governing the 20 Dec 2023 transfer of 2,013,753 shares to Wolter and Tom's intended purchase of Crowdcube investors' shares
AR-005	Low	Drag-along fell from 75% to 55%
AR-006	High	Challenge to D-272: the Articles define Founder as each of Wolter and King, so W-Q1 Q(e) and W-LBA's "four records are against that proposition" hand Wolter an answer. Blocks the Wolter pack; Tom agreed to rephrase

Tom King's recorded answers (phase3/tom-answers.csv)

H1a: NO. Do not enclose the E-231 cap table

It carries about 45 third-party shareholders' data.

Resolves LN-301 (critical) as a decision.
Remove E-231 from the four Phase-1 RFIs (March, Raghunandan, Styles, Agyeman) and from W-EV section 8 (P1-007), and from any enclosure list.
Tom must record a D-row; the review does not write to the matter.

I2: NO. Crowdcube did not launch the July 2026 vote

Now a fact, recorded for the file.

Any sentence in an instrument that implies a vote was launched must be corrected.
Candidates from a keyword search (not yet confirmed as wrong): SH-UPD text lines 16 to 29 ("My interest in Vote 1", "before you vote") and line 129. CC-RRQ line 82 and CC-COV carry no such sentence.
The review has not yet produced its own list of affected sentences. Tom needs to record a D-row or F-row.

13 · How it fits

One loop, from fact to signed instrument and back

A fact enters. A document arrives from the LaCie, a mailbox or SeedLegals. It is hashed and held. Atoms are mined, unnumbered. A person registers an exhibit and a receipt is written.

A decision is made. Tom King decides by dictation. The decision becomes a D-row; the framework requires the author to have read the controlling documents first (read receipt), and landing markers are added so the choice can be tested later.

An instrument is drafted in the vault. The author reads the case theory, strategy and decision log before writing. Quotations are cut from held text, exhibits cited by id, enclosures derived from the letters, and a sidecar records the alignment.

The machinery tests it. Quotation verification, the 67 gate checks, landing markers, the ladder and the freeze test run. Failures are fixed at source or waived on the record.

The review audits it, independently. Subagents re-read primaries, red teams attack outward text from the recipient's side, a sceptic pass tries to refute every critical finding.

It is prepared for a human. A review folder shows the unchanged text, the proposed redlines, evidence and enclosure checks, and the questions that remain.

Tom King decides, signs the hash and sends. Approval is against sha256; the text is frozen; the bytes actually sent are what should be hashed as sent. Only then is the next stage (Day-21 schedules, five-stage additional claims) unlocked.

Everything is logged. The session closes with a ledger entry and handover so the next session continues without loss, and any change goes back into the store as D-rows and corrections.

What is strong and what is weak

Strengths the review would not reopen

Content-hashed rungs did detect the three stale rungs.
The receipt gate stops unreceipted writes on four row tools.
Landing markers caught live defects, including a reversed decision set.
Quotation verification passes 439 spans with no hard failures.
Arithmetic in the Wolter letter (benchmark £5,281,184.04, Seninde £1,999,998, 345 days and others) recomputed consistently.
The drift ledger is complete and the matter was never written to by the review.

The recurring weakness

The machinery reports on itself, not on what will leave the building.
Prose statements of readiness go stale when an instrument moves, and nothing re-cuts them.
Two copies of one fact diverge and no guard compares them (yaml against receipt, status cell against text, the two stores).
Deletions are invisible: rollback files, receipts and plan files have no "should exist" check.
No single plan of record: plan-v5, the 14 September plan, PLAN-TO-SIGNOFF V2 and the handover list all claim to govern.

Proposed next structure

Tom King adopts one plan of record, PLAN-TO-SIGNOFF V3, by a single D-row, built per pack against the seven readiness tests.
Add the end-to-end dispatch check, the as-sent intake, a review-PDF re-render on every instrument edit, and instrument status for the DocuSign signing version.
Make PG the only gate verdict; give the ladder real currency checks; extend landing markers to D-542 to D-693; retire duplicate controls.
Hold the matter drafting-stage: foundation first, then drafting, then sign-off, as Tom King's own sequencing rule requires.

14 · Sources and limits

What this map is built from, and what it is not

Read for this map

Review Project CLAUDE.md, PLAN, HANDOVER-PROMPT and the full LEDGER (4 to 5 October), re-read at commit c3955ae (6 October)
Updated from: 01-FINDINGS.csv (computed), phase3/tom-answers.csv, phase3/CORRECTION-SA-001-2023-ARTICLES.md, phase3/ARTICLES-CROSSWALK-2022-v-2023.md, human-review/PUNCH-LIST.md, human-review/ARTICLES-RECITATION-INDEX.md, 99-INTEGRITY.md sections 8 and 9
05-FRAMEWORK-AND-SYSTEMS in full: rungs, the 67 checks, registers, rules, approval controls, plan and keep/retire list
The SUBAGENT-BRIEFING, the PUNCH-LIST header, the Wolter review pack README and the W-LBA sign-off page
The advisor review's engine and process section (July snapshot)
Tom King's saved litigation notes and working preferences

Not read, or not verified

The live matter store, the Lexi code and the LaCie were not accessed; nothing here was re-verified against them.
The vault copies of the executive summary, integrity and coverage files are early skeletons, so figures for them come from the ledger.
The store's CLAUDE.md, the earlier Codex drafting toolkit (retired with the generator pipeline) and individual instruments were not re-read.
Counts and statuses are as at commit c3955ae and will have moved since. The ladder, gate and scorecard figures were not re-derived in this update.
Not seen by the review or by me: the signed SeedLegals pack, the signed No-Benefit Deed, the Companies House filing, which April text was adopted, D-285, and the Calderbank expiry (13 Nov) beyond D-206(5). The register of members at adoption was not checked.
The I2 sentence list is from a keyword search of three instruments, not a full read. I could not check any E-231 enclosure list for the redline either.

Handle with careThis page summarises privileged, inward working material, including settled facts, case theory, red-team findings and named individuals. It is not for disclosure and not counsel-reviewed. Keep it with the matter folder and do not forward it to any recipient or external service.