The question the plan serves
Can a decomposability measure, validated on systems whose degree is known by construction, read the degree of a freely trained 30-million-parameter transformer that acquired a load-bearing ownership pointer under task pressure? If so, what does it read?
Weekdays and weekends Weekdays carry rulings and spoken go's only, from a queue a session pre-assembles; weekends carry the reviewer sessions John runs by hand and the launches; agents run unattended between weekends.
The weekly re-plan Every Thursday or Friday evening the coming weekend is re-planned from the top of STATUS.md. Kill dates and the wrap-up start never move; what moves is which weekend absorbs a slip, and the re-plan names it.
Weekend by weekend
the outcome, the goals that make it up, and John's part Outcome Amendment A3 (the earlier constructed self-index experiment) closed under Gate A; every registration-blocking decision ruled; the rented slice's seconds per step in the release arithmetic; successor proposal version 2 through its Gate C first-tier review.
Goals
- ✓
W1.1 A3 closure through Gate A, both tiers, and the closure commit landed
DoneJohn and agents
Note Landed 2026-09-25: closure text version 5 appended to the registration file by pull request 49, after both outside reviews, John's rulings, a first-tier check and its re-check
- ✓
W1.2 Every registration-blocking decision ruled and recorded in docs/rulings/2026-09-26-weekend-1-queue.md
DoneJohn
Note Ruled 2026-09-25, all nine pages; the five follow-on decisions from the rehearsal repairs are in docs/rulings/2026-09-25-rehearsal-repairs-rulings.md (pull request 53)
- ✓
W1.3 The 2026-09-22/23 handoff landed at the top of STATUS.md with a 2026-09-24 entry; data/project.toml current; merged worktrees pruned
DoneAgents
Note Landed 2026-09-24 evening: PR 29 and PR 31; no worktree folder was prunable by git's definition (all 69 branches merged), so clearing them is a later git worktree remove pass
- ✓
W1.4 The launcher refuses to rent a machine when the laptop can sleep, and the launch checklist is a gate it checks
DoneAgents
Note Landed 2026-09-24 evening: PR 33 (shared src/launch_gate.sh, 351 and 67 harness checks, charger case and empty-date field covered), checked on PR 36
- →
W1.5 The rented slice run: seconds per step per arm measured on the rented machine and written into the second release's arithmetic; the shutdown path exercised end to end
CarriedJohn and agents
Note Half done 2026-09-25: seconds per step measured on the second attempt (pull request 51, checked on 55) and the registered runs re-priced at $161.90. The half of the shutdown test that runs on the rented machine was not exercised; it moves to weekend 2's development runs (W2.4), the next time a machine is rented
- ✓
W1.6 Rehearsal repairs at toy scale: the named-other condition redesigned (or the risk accepted on record), the nomination procedure shown to be one instrument across arms, control 2 built as stated, the fourth arm attempted if ruled
DoneAgents
Note Landed 2026-09-26 (pull request 52, checked on 58): the named-other redesigns failed, and a later change to the task grammar failed too (pull request 57), so the fallback ruled in advance applies and the risk is on the record; the nomination is one instrument; the fourth, mixed arm passed and is folded into the registration
- ✓
W1.7 Successor proposal version 2 written with the ruled numbers, label and form, the failure-modes checklist exercised against it with commands and output, both checking scripts clean
DoneAgents
Note Landed 2026-09-26 as pull request 54
- ✓
W1.8 Gate C first-tier review of version 2 filed; if clean, the registration text drafted and its Gate A first tier launched overnight
DoneAgents
Note Review filed 2026-09-26 (pull request 56): not clean, one fatal finding and four serious, so the registration text was not drafted. John ruled every finding the same day; version 3 was written from the rulings and is pull request 71, open. Its own review is goal W2.6
- ✓
W1.9 The Saturday ruling packet built and checked (PR 34)
DoneAgents
Note Landed 2026-09-24 evening: PR 34, checked on PR 35, both merged; the packet is docs/rulings/2026-09-26-weekend-1-queue-PROPOSAL.md
John's items
- Run the two A3 closure second-tier reviewer sessions from the committed packets and file both responses verbatim; rule the drafted dispositions the same day
- Rule the pre-assembled queue in one sitting: the eight numbers, the form of the measure, the named-other condition, the middle of the scale, the Wittgenstein criteria, the two record corrections, the spend ruling (successor cap $142; programme envelope $400 to $450)
- Give the spoken go for the rented slice (about $1, hard cap $2)
- Optional, ten minutes: the two Cloudflare secrets and the repository visibility flip, so the site deploys
John's hours about 3
Beside it: Book text lock 2026-09-30 is off the table (done or moved).
Outcome The registration committed through Gate A, both tiers, two weeks ahead of kill date 1; the implementation frozen and tested; the development runs at 10 million parameters landed.
Goals
- →
W2.1 Registration text through Gate A, both tiers; dispositions ruled; text fixes paired-checked; citation and single-source scripts clean on the final text
CarriedJohn and agents
Note Half done: the inside review (RT-237 to RT-246) and John's two outside sessions (RT-247 to RT-256) ran on 2026-10-06 and he ruled all twelve pages. Then the refounding of 2026-10-07 changed the text (the repair, the stop conditions, the renamed outcomes), so version 5 was written and is owed its check. Carried to weekend 3
- →
W2.2 Registration committed (kill date 1 is 2026-10-18)
CarriedJohn and agents
Note Carried to weekend 3 (W3.2). Two things stand in the way on 2026-10-08: the bar for 'the repaired route holds' is unruled, and fifty of the text's citations point at fifteen unmerged branches
- ✓
W2.3 Implementation frozen: generator, arms T and C, the transplant code; unit tests; the even-split rule (RT-58) and the one-scored-token self-test carried over
DoneAgents
Note Done 2026-10-04 (pull request 97): the frozen code reproduces all 502 toy figures and runs end to end at both sizes
- ✓
W2.4 Development runs at 10 million parameters landed and checked by a session that did not write the code; ledger rows opened before any machine existed
DoneJohn and agents
Note Done 2026-10-04 on John's go, about $1.47 for all four; the check (pull request 106) found two of the three built models had switched their built-in route off, which produced the $1.14 repair now written into the registration
- ✓
W2.5 The go packet for the first registered run assembled: launch checklist, ledger row with estimate, sleep gate, scheduled fetch and kill
DoneAgents
Note The development-runs packet served; the first registered run's go waits on the registration and on the reruns' verification
- ✓
W2.6 Proposal version 3 and the rulings on its decisions 20 and 21 merged (pull requests 71 and 72), and version 3 through its Gate C first-tier review by a session that did not write it
DoneJohn and agents
Note Done 2026-10-03: merged, reviewed with nothing fatal (pull request 74), and ruled the same day (docs/rulings/2026-10-03-successor-v3-gate-c-rulings.md). The review session was not a fresh one and says so
- ✓
W2.7 The rehearsal's controls re-run under the registered rules, with the ruled change to how the transplanted piece is chosen; method committed before output; checked by a session that did not run it
DoneAgents
Note Done 2026-10-03 and checked the same evening; three rulings came out of it
John's items
- Added at the 2026-10-03 re-plan, and first: merge pull requests 72 (the rulings on decisions 20 and 21) and 71 (proposal version 3)
- Run the two second-tier reviewer sessions on the registration text and rule the drafted dispositions
- Give the go naming the development runs (about $10, inside the first release)
John's hours about 2.5
Outcome Re-planned 2026-10-08. The registration committed inside kill date 1 with the repair and both stop conditions written in; the three 10-million reruns of the repaired built models landed and verified against the bar John rules; if they hold, the single free-arm run at full size launched on his go. The second release is not this weekend: it is asked for only after both the reruns and the free arm pass, and he gives the go on the figures.
Goals
- ○
W3.1 The twelve open items of version 5 ruled, the bar for 'the repaired route holds' first
PlannedJohn
- ○
W3.2 Version 5 checked by a session that wrote none of it, the fifteen branches merged, and the registration committed (kill date 1 is 2026-10-18)
PlannedJohn and agents
- ○
W3.3 The three 10-million reruns of the repaired built models landed and verified against the ruled bar; a failure stops experiment C here and is written up as instrument research
PlannedJohn and agents
- ○
W3.4 Branch: if the verification holds, the free-arm run at full size launched on John's go and landed; the learn-both gate and the floor read by a paired session
PlannedJohn and agents
- ○
W3.5 The battery draft's third check filed and the measurement rehearsal of its first two entries committed, ahead of Gate A
PlannedAgents
- ○
W3.6 The December result restated under the refounding in one page, so that the kill dates and the new question point at the same thing
PlannedAgents
John's items
- First, before anything else: rule the bar for 'the repaired route holds' (open item 1 of version 5), and the other eleven open items in one sitting
- Merge the fifteen checked branches the registration cites (the sharpness branch, pull request 118, after its check)
- Commit the registration once its check is filed clean
- Give the go naming the three reruns (about $1.14); then, only if the verification holds, the go naming the free-arm run (about $12, about ten hours)
- Ten minutes, optional: the two Cloudflare secrets, so the site deploys
John's hours about 2, spread across the three days
Beside it: Columbus Day makes it three days. The field exercise of 2026-10-13 to 14 follows (blackout, tentative).
Outcome As written 2026-09-24; left as written at the 2026-10-08 re-plan until weekend 3's handoff knows what landed. If the reruns and the free arm both held, this is the earliest weekend the second release (the remaining eight runs, about $150 to $172) can be asked for on the figures; if either stopped experiment C, this weekend starts the instrument-research write-up instead. Kill date 1 passes on Sunday.
Goals
- ○
W4.1 Nomination procedure frozen and committed, method before output
PlannedAgents
- ○
W4.2 Transplants run on fresh episodes on every arm
PlannedAgents
- ○
W4.3 The measure computed on T and C: the validation result, with paired uncertainty
PlannedAgents
- ○
W4.4 Validation findings drafted and Gate B first-tier review launched
PlannedAgents
John's items
- The queue only; one ruling if the nomination freeze needs one
John's hours under 1
Beside it: Kill date 1 is Sunday 2026-10-18.
Outcome John's Gate B ruling on the validation result: R1 (read arm F), or R2 or R3 (closure path). If R1, arm F read on the frozen procedure with confirmation seeds.
Goals
- ○
W5.1 Gate B on the validation result, both tiers, ruled
PlannedJohn and agents
- ○
W5.2 If R1: arm F read on the frozen procedure with confirmation seeds, locally; findings drafted
PlannedAgents
- ○
W5.3 If R2 or R3: closure text drafting starts
PlannedAgents
John's items
- Run the two second-tier reviewer sessions on the validation packet and rule
- If R1: rule on extension E1 (three more free-arm seeds, about $36, an envelope ruling) for weekend 8
John's hours about 2.5
Outcome Findings document for the reading (or the null); Gate B first tier run midweek before; John's two second-tier sessions and ruling. Kill date 2 passes with the runs landed three weeks earlier.
Goals
- ○
W6.1 Findings document for the reading or the null, through Gate B both tiers, ruled
PlannedJohn and agents
- ○
W6.2 Closure text (registered text) drafted the following week
PlannedAgents
John's items
- Run the two second-tier reviewer sessions on the reading's findings and rule
John's hours about 2.5
Beside it: Kill date 2 is Sunday 2026-11-01.
Outcome Gate A on the closure text: first tier midweek, John's two second-tier sessions and ruling on Saturday, closure landed in the successor's registration file on Sunday. Six weeks before wrap-up.
Goals
- ○
W7.1 Closure text through Gate A, both tiers, ruled, and landed in the registration file
PlannedJohn and agents
- ○
W7.2 STATUS.md, data/project.toml and the site carry the result
PlannedAgents
John's items
- Run the two second-tier reviewer sessions on the closure text and rule
John's hours about 2.5
Beside it: The field exercise of 2026-11-10 to 11 follows.
Outcome A slipped gate takes this weekend before any extension does. Otherwise extension E1 (three more free-arm seeds for a tighter degree reading) runs.
Goals
- ○
W8.1 Absorb a slipped gate from weekends 2 to 7, if any
PlannedJohn and agents
Note Re-plan 2026-10-03: expected to be used. The registration slips one weekend, so the full-size runs move to weekend 4 and each of weekends 4 to 7 takes the work of the one before; those entries are left as written until the 2026-10-04 handoff knows what landed
- ○
W8.2 Else: extension E1 launched and landed
PlannedJohn and agents
John's items
- Only if E1 was ruled: the go naming the three runs
John's hours 0 to 1
Beside it: Mini CSS 2026-11-18 to 19 and the mile retest 2026-11-20 follow.
Outcome Slack for a slipped gate; otherwise extension E2 (the existing 10-million-parameter checkpoints read on the frozen procedure, $0) and E1's findings with a short Gate B.
Goals
- ○
W9.1 Absorb a slipped gate, if any
PlannedJohn and agents
- ○
W9.2 Else: E2 read and filed as a side reading; E1 findings through Gate B
PlannedJohn and agents
John's items
- Only if E1 ran: the second-tier sessions on its findings and the ruling
John's hours 0 to 2
Outcome The state-of-the-programme write-up; the write-up to the outside human reader; explainer.md refreshed; site and data/project.toml current.
Goals
- ○
W10.1 State-of-the-programme write-up drafted and reviewed
PlannedJohn and agents
Note Ruled 2026-10-08 (decision 3 of the December-result restatement): not a separate document. The Sentient Horizons essay draft is its public face and the internal write-up folds into the closing STATUS entry at wrap-up
- ○
W10.2 Write-up sent to the outside human reader
PlannedJohn
- ○
W10.3 explainer.md refreshed to cover the registered results, MVM-0, A3 and the successor
PlannedAgents
- ○
W10.4 Site and data/project.toml current with the December result
PlannedAgents
John's items
- Choose and contact the outside human reader from docs/outside-reader-shortlist-2026-09-19.md
- Read and rule on the state-of-the-programme write-up
John's hours about 3
Beside it: Thanksgiving makes it four days.
Outcome No work planned. Mini CSS 2026-12-04 to 08.
Goals
No goals. Nothing is planned this weekend.
John's items
Nothing for John.
John's hours 0
Beside it: Mini CSS 2026-12-04 to 08 (five days, Friday to Tuesday).
Outcome Records closed with actual-after rows; RESUME.md drafted; any extension's Gate B ruling. Planned light: inside the post-field recovery window.
Goals
- ○
W12.1 Compute ledger, STATUS.md, data/project.toml and the site closed with actual-after rows
PlannedAgents
- ○
W12.2 RESUME.md drafted: how to wake the programme in May 2027
PlannedAgents
John's items
- Any extension's Gate B ruling, if one is pending
John's hours 0 to 1
Beside it: Inside the roughly ten-day recovery window after the Mini CSS.
Outcome Final state of the record before wrap-up starts Monday 2026-12-21; nothing launches. The hibernation condition is completed over the exodus break, by 2027-01-04, with John's final ruling that it is met.
Goals
- ○
W13.1 Record in its final pre-hibernation state; nothing in flight
PlannedAgents
- ○
W13.2 Hibernation condition per docs/program-roadmap-2026-09-20.md met and ruled, by 2027-01-04
PlannedJohn and agents
John's items
- Final ruling that the hibernation condition is met (may land during the exodus break)
John's hours about 1
Beside it: Wrap-up starts 2026-12-21. Exodus break roughly 2026-12-21 to 2027-01-03 (dates inferred, not published).