Roadmap
Current direction: agent-native, OMP resumption authorized (19 Sep 2026)
Robert approved the direction in 01-premise and the supersession recorded in decision-log. Mo is another option, not a BEAM replacement. Agents are the code authors and readers; humans judge behavior and evidence. The complete, fast, trustworthy agent feedback loop for existing features takes priority over expansion. Robert’s evening instruction resumes the two paused units with Astra leading in the existing OMP pane and fresh OMP/Sol/high workers. The lead reviews and independently runs acceptance checks; workers write implementation code. The Claude Code/Opus workflow is superseded.
Current checkpoint, 20 Sep 2026, 8:30 AM ET: Step44 accepted on Darwin, main db515f9b; last accepted full suite 272/272. Step42 and server remain unaccepted. Linux, live provider, model-driven task and Pi comparison remain owed.
| board | current work |
|---|---|
| Now | Server generation42/42,15 records, protected F1 unchanged; grant09 runs remaining20 correctness commands exclusively. No server acceptance/benchmarks yet. Step42 ownership plus bounded native/test quality corrections source-only before AFTER freeze; private patch context rebased. Step44 accepted272/272 plus30 controls; measured line cost disclosed |
| Done today | Accepted on main: the rebuilt mo-application-workspace-v1; mo-harness-in-mo steps 1, 8 and 2; the raw-memory runtime fix (toolchain-raw-memory-report); interpreter-step-40 (a scope that holds against links, Fs.replace); interpreter-step-43 (the auditor’s PR 15: every number in source held to its range, accepted 1:59 PM ET). mo-harness-end-to-end-v1 (the Mo agent end to end on the machine with the scripted Logstat repair, accepted 2:20 PM ET). mo-agent-report-cap (the report cap and defect D1, accepted 3:06 PM ET); the report-claim-check calibration. interpreter-step-41 (Exec, accepted on macOS 3:33 PM ET, full suite 268 of 268). the first x86_64 Linux full suite 249 of 249 on step 40’s tree |
| Next | Review remaining server provenance and Step42 ownership repair; serialize later validation and measurements, with no automatic benchmark grant. Finish Step42 and ci-gate; independently accept mo-workspace-server-4a without changed clients or weakened half-close/operator predicates, then part B (command through Exec, machine, cutover); plan steps5–7; provider login, model-driven task, Pi comparison |
| Waiting on Robert’s presence | Interactive subscription login when ready; offline work continues |
| Recently done | interpreter-step-44, accepted20 Sep8:30AM ET: bounded byte source, full272/272,30 controls,20 interleaved line samples; Linux owed. Earlier: mo-workspace-http-v1, mo-workspace-recovery-v1, mo-application-build-v1, mo-provider-bridge-v1, mo-coding-fixture-v1, mo-workspace-foundation, mo-provider-auth-v1 |
Program 7 is on hold; its superiority thesis and runtime-claim retirement framing are superseded. Its sealed specs and evidence remain historical; none of its old gates is declared passed. Capabilities requirements and the separate generation-ten contract-catch rule are not retired. Step 39 remains unaccepted. New upstream TLS/script edits are not acceptance evidence.
Next work in order
- Application chosen: mo-first-coding-harness, written in Mo and initially maintaining Mo applications. Robert approved the lead/oracle recommendation. Pi is the practical comparator; do not build a second Elixir harness first. Mac arrival and Herdr launch are verified. Establish the dedicated execution boundary through mo-executor-foundation, then the brief’s provider/executor, protected verification, worker write scope and budget decisions before declaring implementation readiness.
- Under the overnight authority: truthful instruments, one complete headless coding path, 401 policy calibration, then two or three bounded application tasks and a matched Pi comparison. Add an affordable model after calibration. Version the Agent-specific policy and settle its recipe binding first. Requirements, implementation and verification remain distinct obligations; the pilot is not proof of language value. No self-editing or compiler changes.
- Agree and explicitly version program 7’s replacement scope, metrics and relationship to retained audit obligations before any new sealing or build.
- Keep moving under Robert’s overnight authority. Record and verify each bounded slice; handle approval decisions as lead and continue independent work around prerequisites requiring Robert’s physical presence.
The final DeepSeek-labelled attachment duplicated the supplied Kimi body exactly; it is not a fifth distinct assessment. Four distinct reviews remain. Preserve agent-native-independent-review-prompt v1.0 unchanged and compare evidence, not votes. The application choice supersedes the earlier open workload question.
Four reports received and reviewed with the oracle
| brief | decision it informs | status |
|---|---|---|
| Unfamiliar-language learning | Minimal guide and teaching feedback; compilation versus behavior | hypothesis to test |
| Text versus semantic editing | Safe targeted operations alongside ordinary source | defer new API pending measured need |
| Useful language restrictions | Compiler, runtime, OS and application responsibilities | no law changes approved |
| Trustworthy verification | Evidence identity, actual coverage and verifier authority | useful principles; proposed gates not ratified |
agent-native-research-synthesis separates findings, selective source checks, limitations and proposals; raw reports are preserved with hashes. The evidence attachment’s auditor designation was encountered during the batch, so this is not a cold independent audit reading. No hidden suite or old gate changes. Later model comparisons include affordable/open-weight models and count failed attempts and stronger-model assistance; smaller-model reliability is a hypothesis.
Proposed implementation sequence, not a work authorization
Trustworthy evidence and bounded execution; security/durability containment; one complete discover/learn/edit/check/run/inspect/repair/verify slice on an existing application; measured diagnostic and latency improvements; then a versioned program 7 and broader model evaluation. Exact scope and ordering remain open. A smaller program-7 core, TLS make-versus-wrap, specific law changes, incremental compilation and acceptance thresholds are proposals, not ratified implementation decisions.
Historical board before the reframe
The board and phase tables below preserve the earlier queue and evidence. Their “Next”, “Later” and “Waiting” sections are not current execution instructions; in particular MCP/agent tooling is no longer relegated to later work. Outstanding defects remain outstanding; the new direction does not accept prior work.
The board first: what is in flight, what comes next in order, what waits, what just landed. One line each, with the page that holds the detail. Fable rewrites it at every acceptance and every pause; the dates say when. Below it, the phases in detail and the history of every step.
Lead transition and audit review, 18 Sep 2026 (Amp; reproductions at 5:28 PM
ET). Robert authorized Amp as lead. PR 12 is integrated after the lead’s
reading was filed first. No worker has restarted: A/B are pushed, C/D are saved
WIP, E/F unstarted, step 39 unaccepted. Repair acceptance instruments before
using their counts again. The orb’s private transfer/worktree state is
unverified; HANDOFF.md and the new decision-log rows hold the continuation
boundary.
Now, in flight
| what | who | since | page |
|---|---|---|---|
Step 39: the TLS brick’s corrections (two auditor readings showed the client accepting four certificate chains OpenSSL rejects; also ALPN past 64 names, the fuzz count, a corpus test that tested nothing, plaintext records after the handshake keys) and F_FULLFSYNC on macOS. The TLS brick is not complete until it lands. Its gate is the auditor’s own script | paused; A/B pushed, C/D patch saved, E/F unstarted; worker ended | paused 18 Sep, 4:33 PM ET | interpreter-step-39 |
| Program 7: revision 2 sealed; fresh-server baseline/skip list unfinished. PR 12 requires R2/RC1/R4 reconciliation before the seal. Lead readiness findings also require binary-safe input, snapshot/log replay and durability resolution. No builds yet | lead; ratified-rule decisions and auditor’s sealing session with Robert | 18 Sep, after PR 12 | 07b-redis-subset-revision-2, decision-log |
With the auditor: generation-six-ready-001, step-38-ready-001, the comparisons of five readings filed this morning | the auditor | 18 Sep, 11:35 AM ET | audit/ |
Next, in order
| # | what | why now | page |
|---|---|---|---|
| 1 | Program 7 readiness: repair acceptance instruments with negative controls; finish baseline/skips; resolve R2/RC1/R4 and the binary-input/replay/durability gaps; explicitly version any spec changes; auditor’s seal, then matched Linux builds | the primary claim needs a valid experiment; step 38’s acceptance does not prove abuse stimulus coverage | 07-redis-subset, 07b-redis-subset-revision-2, decision-log |
| 2 | A toolchain step from generation six: the simulator bounded on a process with a perpetual timer (the 13.4 GB process), for _ in 0..n as a counter, an Fs row that syncs a folder; then the allocator’s resident set behind generation four’s two states | a maintainer’s plausible design took a machine down; a Mo program cannot make a rename durable by the book | erosion-round, gen4-memory-control-probe |
| 3 | Change 7 and generation seven, with the suite smoke-run on generation six’s programs before its seal and the /queues sentence settled | the language rule is read at generation ten; the ledger stands at zero catches of two after six | erosion-round |
| 4 | Placement for what main starts, and a unit test for the step-aside: the step 34 follow-up | echo-1k and the binary’s kv-10k-get still slower at 14 cores | interpreter-step-34 |
| 5 | Chapter 10’s other sections as steps: §2 the restart budget diagnostic, §3 the counted laws to mo.toml, §5 MO0317 naming the changed module; the contract-cost diagnostic generation four asked for | decided on the language page | 10-language-after-the-rounds |
| 6 | The brick cache shared by source hash; the compile benchmark at 5,000 modules | a brick change recompiles once per corpus program today | — |
Later
The compaction copy per reference (half again on jobq’s log) and the one-second restart at 100,000 jobs; mo prove (tier 3); the runtime surface as MCP tools and measurement 4, the incident round, which tests whether the surface shortens an agent’s diagnosis; round 9’s Opus-in-Pi baseline when Pi has an Anthropic key; the package registry once an outsider runs a real service; program 8, the toolchain in Mo.
Waiting on Robert
Blocking program 7: ratified measurement reconciliation (R2’s fault rig,
RC1’s clock, R4’s hundred edits) and the auditor’s sealing session, which only
Robert opens. program-7-spec-ready-002 remains published, but PR 12 shows that
publication is not proof the measurement setup is compliant. Lead
recommendations are rows in decision-log; no rule was amended by this
review.
Rows marked “for Robert” in the decision-log, newest first: the morning of 18 Sep on the Mac (step 38 accepted; generation six read, the laws silent a sixth time; on macOS Mo’s fsync does not reach the disk, the Mac’s speed comparisons with Go withdrawn; three corrections of the lead’s own: the sealed seventh suite was invalid and had never been run, generation five’s third-suite column was never measured, P7’s threshold re-read with evidence in view; five auditor readings compared, every finding conceded, the TLS brick not complete; program 7’s spec to a revision 2; the 13.4 GB process named); the night of 17 to 18 Sep on the VM (step 37 and its fix, the VM’s wedge, the memory control probe); the bricks page’s two rows (the shelf boundary and program 7’s two bricks; the audit items, caps, and cost); the auditor role taken up, M-3 accepted, the bricks prerequisite, and three disagreements with the ratified rules (program 7’s shape against the capabilities rule, the R-A trigger at zero catches, Fable’s reading of the runtime rule’s open clauses) (17 Sep, afternoon); research PR 2 read and merged, chapter 10’s attribution corrected (17 Sep); generation five’s reading, the first Mo-only defect, the laws still silent; the change 4 Mo queue nine times slower on the lease path, speed recorded per generation; generation four; chapter 10 §2’s budget as a value; generation three, the BEAM’s row answered; P6 on Mo’s change 2; round 9’s Haiku row and its reading; generation two; chapter 10 §1 as built (step 31); chapter 10 itself; P6 on Elixir; measurement 1’s completeness row; the BEAM row after round 10; the outage, probed and read; round 8 read; the 16 lint issues from his history bundle. Apart from the audit session, none blocks the work: Fable decides and records, Robert overturns.
Recently done
| when | what | page |
|---|---|---|
| 18 Sep, after 5:28 PM ET | PR 12 integrated; lead reading filed first; current harness false-success paths reproduced with unchanged mock probes; TLS findings qualified by their pre-A/B anchor; program 7 measurement discrepancies conceded; no implementation accepted | decision-log |
| 18 Sep, 11:59 AM ET | Step 38: a Conn reads while a write on it waits, TCP_NODELAY, 64 handshake abuse cells; the lead’s suite 237 of 237, the mutant over ten runs, a byte-exact pipelining probe, the table against cells committed before it was seen | interpreter-step-38 |
| 18 Sep, 11:33 AM ET | Generation six: nothing eroded, no defect under the corrected seventh suite, Mo’s generation-five defect fixed, the lease path back to change 3’s rate; P3, P4, P6 failed, P7 void; the sealed suite invalid and generation five’s third suite never run, both corrected | erosion-round |
| 18 Sep, 10:00 AM ET | Five auditor readings and one research note received, Fable’s four readings filed first, every finding conceded: the TLS brick not complete, the seventh suite’s defects, program 7’s spec to a revision 2, step 38’s cells fixed in advance | audit/ |
| 18 Sep, 4:45 AM ET | The generation-four memory control probe: forty runs; the 46 MiB was one draw from a bimodal spread independent of the contract; the rate loss is the walk; both readings retire | gen4-memory-control-probe |
| 18 Sep, 4:16 AM ET | Change 6 sealed and generation six pre-registered: the archive pruned, a speed budget, the rename rule corrected | 01g-job-queue-change-6 |
| 18 Sep, 3:50 AM ET | Step 37, the TLS brick, part two: the client, the chain, ALPN, KeyUpdate; 1,000 sessions against OpenSSL at 0 mismatches, a fuzz hour at 0 crashes; two runtime defects found by the lead’s probes and fixed the same night | interpreter-step-37 |
| 18 Sep, 3:50 AM ET | Program 7’s spec sealed: mored, a Redis subset against Redis’s own tests | 07-redis-subset |
| 17 Sep, 10:30 PM ET | Step 36’s two readings compared: four points conceded to the auditor, all closed in step 37 | interpreter-step-36 |
| 17 Sep, afternoon | The bricks page: the shelf boundary as three tests, the cost measured against Zig’s std (about 70,000 lines to read, 12,000 to 15,000 to write, estimated), five audit items and a surface cap per brick, bricks written once in Zig for both runtimes, the vendored-C fallback; program 7 needs crypto and TLS | bricks-and-the-cost-of-zero-dependencies |
| 17 Sep, afternoon | The auditor role taken up: three stopping rules ratified by Robert, linked from chapter 1, chapter 10, and the state page; M-3 accepted; Fable’s disagreements filed as rows | audit/README.md |
| 17 Sep, morning | Research PR 2 (the Hermes lane, wiki only) read and merged: restart budgets, error-path coverage, crash consistency, capability confinement; chapter 10’s 3-in-5 budget reattributed to Elixir Supervisor; erosion2-* and erosion5-* pushed at last | hermes-daily-2026-09-17 |
| 16 Sep, 22:55 | Generation five: change 5 by four maintainers in 11 to 25 minutes; the sixth suite Python 94, Elixir 94, Go 92, Mo 93 of 94; the first Mo-only defect, the laws silent; the speed row per generation | erosion-round |
| 17 Sep, 10:15 PM ET | Step 36, the TLS brick, part one: a TLS 1.3 server engine written once for both runtimes, a Conn that stays a Conn; 1,026 and 1,196 handshakes a second, bulk at 2.4× and 1.8× plain over AES-GCM; two defects behind the worker’s reported green found and fixed; the suite 225 of 225 | interpreter-step-36 |
| 17 Sep, 19:25 UTC | The cause of generation four’s nine-times loss named: change 4’s sweep postcondition walks every finished job on every request; contracts off restores generation three’s rate; a chapter 10 candidate | erosion-round |
| 17 Sep, 18:16 UTC | Step 35, the crypto brick: one Zig file over std.crypto for both runtimes, the rows on seven names and the capability Random, the standards’ vectors in the tests and the corpus, the differential run 0 of 18,000 per runtime, the fuzz 0 crashes; List(UInt8) at 6.3× and 2.9× the raw call carried | interpreter-step-35 |
| 16 Sep, 21:20 | Step 34: placement with the starter, the crossing made cheap (100k asks across schedulers 4 s to 0.14), measured at 1, 4, and 14 cores; the change 4 Mo queue found nine times slower on the lease path | interpreter-step-34 |
| 16 Sep, 17:05 | Change 5 sealed, generation five pre-registered: a lease handed off, a queue renamed with jobs in flight | 01f-job-queue-change-5 |
| 16 Sep, 15:05 | Generation four: change 4 by four maintainers; the fifth suite Mo 77, Python 77, Go 76, Elixir 76 of 77; nothing new eroded in four generations | erosion-round |
| 16 Sep, 13:03 | Change 4 sealed: idempotent creates, the archive | 01e-job-queue-change-4 |
| 16 Sep, 13:05 | Step 33: the crash report freed in both runtimes (46 MiB a restart to under 0.3), the interpreter’s abort on a large log fixed | interpreter-step-33 |
| 16 Sep, 11:20 | Generation three: change 3 by four maintainers in 10 to 23 minutes; the fourth suite Mo 55, Go 55, Python 55, Elixir 53 of 55; the Mo queue killed under load back in 106 ms, the BEAM’s row answered | erosion-round |
| 16 Sep, 10:20 | Change 3 sealed: the store restarts itself, a budget, a chaos switch | 01d-job-queue-change-3 |
| 16 Sep, 10:20 | Step 32: crash reports kept apart from the ring, both runtimes; the reopening restart in the corpus; the P6 probe lists the crash with the default ring | interpreter-step-32 |
| 16 Sep, 09:05 | P6 on Mo’s change 2: the queue crashed under load through the surface, 503 within 2 ms, nothing lost, no restart; the outage closed, the restart the program’s | erosion-round |
| 16 Sep, 08:45 | Round 9 closed with Haiku 4.5: wrong in every language in under ten minutes (Mo 28, Go 22, Python 18 of 189) | control-run-9 |
| 16 Sep, morning | The wiki as a site, the state page, seven maps | state-of-the-project |
| 16 Sep, 02:00 | Erosion round, generation two: nothing eroded in Mo, Go, Python; Elixir refuses a torn line after a full disk | erosion-round |
| 16 Sep, 01:15 | Step 31, the deferred reply, in all three runtimes | interpreter-step-31 |
| 16 Sep, 01:00 | Measurement 1 complete: twelve of twelve regenerations at 1.0 | bodies-as-cache |
| 15 Sep, night | Chapter 10, the language after the rounds; P6 on Elixir; the Mac scaling run; round 10 | 10-language-after-the-rounds, control-run-10, mac-scaling-run |
The phases in detail
Rewritten 15 Sep 2026 after program 6, when Robert asked for the forest and agreed to Fable’s order. The rows above the line are what the first three days built; the rows below are the roadmap from here. A step is one brief to one fresh worker session; this table is the count.
| phase | status | briefs left, roughly |
|---|---|---|
| Step44, bounded byte input (interpreter-step-44) | accepted20 Sep on Darwin: full272/272 and independent socket/CLI controls; line measurements disclose observed cost, not zero overhead | server adoption separately; Linux owed |
| The design, the corpus, the interpreter milestone (sessions 1–5) | done | — |
The toolchain under real programs (steps 5–28): formatter, stdlib, Net, Http, the C backend, fibers, the runtime surface, the derived deadline, the delayed send, the one-line if, the keywords as names, a never at rest, a map written in place | done | — |
| Programs 1–6: jobq, logstat, kv, notes, agent, ledger | done; each found two to four toolchain bugs and several gaps | — |
| Control rounds 1–7: rounds 1–6 on agent time (Mo slower, round 6 failed on all four), round 7 on Robert’s measure (held on all four; reliability level at 0 defects each) | done | — |
| Outside reviews of 13 and 14 Sep, the thesis restated (chapter 1), the measure restated (chapter 8), the closure audit, the reading pack; Fable’s earlier readings (empirical-validation-plan, ecosystem-strategy, research-agenda-2026-09-response) | done | — |
Step 29, the runtime honest (interpreter-step-29): restart: :never honoured, platform.exit with a pending delayed send, replay streamed instead of held in one Open update, the simulated-time jump gap, the invariant-untrippable rule as a diagnostic | done 15 Sep; part C’s memory bound unmet on a real log | — |
| Step 29b, replay memory on a real log (interpreter-step-29b): Fable’s acceptance probe replayed a 1M log an HTTP session wrote and passed 8 GB fifteen seconds after the fold, where the worker’s generated log peaked at 2.5 GB; the rule made to hold in every loop shape | done 15 Sep; the evidence log replays at 2.3 GB where it was killed past 9 | — |
| Step 34, placement (interpreter-step-34): a process an update starts placed with its starter, the cross-scheduler crossing made cheap, measured on the Mac at 1, 4, and 14 cores | done 16 Sep, 21:20, accepted: the queue’s pairs and kv-10k-get under mo run level across cores, the crunchers 7.3×, echo-1k and the binary’s kv-10k-get still slower at 14 by the main-starts rule; carried: placement for what main starts, a test for the step-aside, the compaction copy measured on the ledger | 1 (the follow-up) |
Step 35, the crypto brick (interpreter-step-35): SHA-2, HMAC, HKDF, AES-GCM, ChaCha20-Poly1305, X25519, Ed25519, Argon2id, constant-time equality, OS entropy, written once in Zig and linked into both runtimes; Random as a capability | done 17 Sep, 18:16 UTC, accepted: the differential run 0 mismatches in 18,000 per runtime, the fuzz 0 crashes in an hour and 3.9 million inputs, Fable’s probe equal to python’s cryptography under both runtimes; SHA-256 of 1 MiB 245 and 532 MB/s against 1,548 raw; carried: a byte-string value for List(UInt8)’s 6.3× and 2.9× | the TLS brick (step 36) |
Step 36, the TLS brick, part one (interpreter-step-36): a TLS 1.3 server engine over bytes, both suites, X25519, Ed25519 and P-256 certificates, KeyUpdate; platform.tls and a Conn that stays a Conn | done 17 Sep, 10:15 PM ET, accepted: 1,026 and 1,196 handshakes a second, bulk 2.4× and 1.8× plain, fourteen abuse rows, 22 lead probes, 225 of 225 after two fixes the worker had reported green | step 37 |
| Step 30, processes on every core (interpreter-step-30): a scheduler per core, messages across threads, the store’s fsync off the scheduler, measured against Go on the queue and the ledger | done 15 Sep, accepted; the queue is fsync-bound on this disk (round 8 reruns the baselines on it); the ledger doubled from the fsync pool; CPU-bound work 3.7× at 4 cores; two 4-core bugs found by measuring. The Mac run (15 Sep, night): every row fastest at 1 core, the cross-scheduler ask the cost; placement is a step after 31 | 1 (placement) |
| Round 8, the maintenance round (control-run-8): round 7’s three finished job queues handed to fresh agents with a changed spec; a second hidden suite for the change; defects and regressions counted; the experiment the laws were written for | done 15 Sep, night: held on all five (regressions 0/0/0, defects 0/1/0, 1,420 pairs a second against 428 and 325, the loop 0.81 s against 14.4 and 7.5, 0 dependencies); the conjunction survives, the laws’ value for the second agent still unshown (no check caught a bug, Mo 2.2× Go’s time); the fourth oracle found an outage on an impossible log record, for Robert | — |
| Round 9, the small-model round (d41-small-model-round, Robert, 15 Sep): the same pre-registered task and hidden suite with a small model in all three panes, several models of different sizes, open-weights ones among them (Robert, locked); reliability and loops are the columns that move with the model | done 16 Sep, 08:45 (control-run-9), five of five models: reliability moved with the model on the Mo side for the cloud models (kimi 1, deepseek 2, gpt-5.5 0 defect causes; every Go change 1, every Python 0) and in every language for Haiku 4.5 (Mo 28 checks over 4 causes, Go 22 over 6, Python 18 over 2, in 7 to 9 minutes, none green by every check); the diagnostics carried the cloud models to green (first fix right in 21 of 23 loops); the local 27B made no edit in any language; P1 held, P4 for four of five, P2 (open-weights and Haiku), P3, P5 failed; the Opus-in-Pi baseline unmet | — |
| The erosion round, generations two to five (erosion-round): changes 2 to 5 to all four programs by fresh maintainers, a hidden suite per generation, the fourth oracle, P6, the speed row | generation two done 16 Sep, 02:00: nothing eroded in Mo, Go, Python on the old suites; Elixir refuses a torn line and cannot restart after a full disk; third suite Mo 65, Go 66, Python 65, Elixir 62 of 66. Generation three done 16 Sep, 11:20: change 3, the store that restarts itself; the fourth suite Mo 55 of 55 under both runtimes, Go 55, Python 55, Elixir 53; the Mo queue killed under load back in 106 ms (Elixir 285 to 694), nothing lost; P1, P2, P5 held, P3 and P6 half, P4 failed; nothing new eroded in three generations. Generation four done 16 Sep, 15:05: change 4, the key and the archive; the fifth suite Mo 77 of 77 both runtimes, Python 77, Go 76, Elixir 76; P1, P2, P5, P6 held, P3 and P4 failed; nothing new eroded in four. Generation five done 16 Sep, 22:55: change 5, the handoff and the rename, the seam a law; the sixth suite Python 94, Elixir 94, Go 92, Mo 93 both runtimes (the first Mo-only defect: a folder its own verify refuses after a compaction and a rename); P1 and P5 held, P2, P3, P4, P6 failed; the speed row per generation added, and generation four’s Mo queue found nine times slower on the lease path | changes 6 to 10, one generation each |
| Round 10, the Elixir round (control-run-10, d42-elixir-round): the BEAM null hypothesis run; round 7’s job queue and round 8’s change in Elixir with dialyzer, credo, and ExUnit, the same suites | done 16 Sep, night (brought forward): 2 defect causes against Mo’s 0, twice Mo’s speed at 32 workers, the loop 7 s against 0.8, zero run-time dependencies, half Mo’s time to write; the null hypothesis alive; P6 probed on the Mac 15 Sep: the queue killed from outside three times under load, service back in under 600 ms each time, 0 acknowledged writes lost, the BEAM’s row, bounded by the default restart intensity (four kills in 2 s and the node exits). P6 on Mo’s change 2 (16 Sep, morning, erosion-round): the queue crashed by an input through the runtime surface under 10,000 requests a second, every request 503 within 2 ms from then, nothing acknowledged lost, no restart by the program’s :never; the outage is closed, the restart is the program’s to take (change 3), since a restarted process re-runs its state initializers. Change 3 took it (16 Sep, 11:20): the Mo queue killed under load back in 106 ms, nothing lost, against Elixir’s 285 to 694 ms; the BEAM’s P6 row is answered | — |
| Measurement 1, bodies as cache (bodies-as-cache, direction 43): every program regenerated from its stripped spec, twice | done 16 Sep, 01:00: twelve of twelve regenerations at completeness 1.0 (logstat, kv, notes, jobq with the hidden suite, ledger, the agent program); the stronger form, tests deleted too, on logstat: 0.74 under the original tests, 1.0 under the transcripts | — |
| The bricks page (bricks-and-the-cost-of-zero-dependencies): the shelf boundary, the cost, five audit items and a cap per brick, the ordering rule, the fallback | done 17 Sep, afternoon; five rows; the crypto and TLS bricks are the steps before program 7 | 3 (crypto 1, TLS 2) |
The language after the rounds (spec/design-v0/10-language-after-the-rounds.md, Fable wrote it 15 Sep on the Mac, after P6): one section per candidate change the runtime’s evidence supports, each with the round row that motivates it, its cost, and code options for Robert: deadlines carried by the process or the message type; the failure model stated on the process declaration and checked at the caller; placement left out until a program needs it; the runtime surface as a capability; and the laws that cost loops without catching bugs removed. Zero new syntax stays the default; no change without a control-run row behind it. Robert’s locked language items fold into it | the page written 15 Sep: six sections, two with a change (a deferred reply token, §1; the restart budget asked for on every :always child, §2), the counted laws as settings (§3), no change to never and invariant (§4), MO0317’s rewrite (§5); §1 is the step that decides P6 for Mo, for Robert | one step per accepted section, §1 first |
Step 31, a deferred reply (interpreter-step-31): chapter 10 §1 built: Reply(T), reply_to, answer, in the checker and all three runtimes, a corpus file, the spec lines | done 16 Sep, 01:15, accepted: Reply(T), reply_to, answer, MO0411, all three runtimes, the corpus file; the standing rows unchanged, the deferred reply level with send-and-a-message-back at 8 askers and 0.46 at 128 under mo run (the parked fiber, for the placement step); Fable’s crash probe 8 of 8 Down then 8 answered | — |
| Step 33, the crash report freed, and the interpreter’s abort on a large log (interpreter-step-33): the report freed after printing in both runtimes; the high-water mark fixed | done 16 Sep, 13:05, accepted: 46 MiB a restart to under 0.3 as a binary, 28 to 0 under mo run; the abort fixed with a unit test; unwritable, restart, budget green both runtimes; carried: the 1.45 s restart at 100,000 jobs, the compaction copy per reference | — |
Step 32, crash reports apart from the ring, and the reopening store (interpreter-step-32): /crashes kept where load cannot evict it, both runtimes; a corpus file showing a restarted process reopening its store | done 16 Sep, 10:20, accepted: the last 16 reports in their own store, --crashes N and MO_CRASHES=N, newest first; crash-kept.mo and restart-reopens.mo; the probe lists the crash with the default ring; rows within noise; carried: the full reports both runtimes keep for a whole run, unbounded | — |
The language items Robert agreed to (now sections of the page above): the counted shape laws as project settings; a named function passed by name where an anonymous function goes; the grammar forms that cost loops in every program (return in a case arm, a qualified call, a split lambda body) as diagnostics that say what to write; the invariant construct reconsidered after round 8 | superseded: these became chapter 10 §§2, 3, 5 and sit in the Next table as row 3 | — |
| A compile benchmark at 5,000 generated modules (the agent-loop claim at scale) | queued, small | 1 |
| Program 7, a real open-source service reimplemented against its own test suite (a Redis subset with streams, persistence, auth, pub/sub, its operation set pre-registered from Redis’s own tests), once a program finishes with no new gap or bug note | Next, row 4 (M-3, 17 Sep): after the crypto and TLS bricks and the speed probe, before generation six’s build | 2 |
Tier 3 proving, mo prove | queued | 2–3 |
| The package registry | deferred until an outsider runs a real service | — |
| Program 8, the toolchain in Mo | late | — |
The Next table at the top is the order; the rows here are the phases’ record.
Done
| step | what | evidence |
|---|---|---|
| 44 | bounded owned Conn.chunks, existing source backpressure, exact binary input and half-close reply | interpreter-step-44; independent272/272,30 controls and20 larger line samples; accepted on Darwin20 Sep2026 |
| 1–3 (sessions 1–3) | the wiki, design v0, the grammar, 13 comparisons | mo-wiki/ |
| 4 (session 5) | the corpus, 52 programs, three-model bake-off | corpus, model-bakeoff |
| 5.1–5.4 (session 5) | lexer, parser, tier-1 checker, VM, tier-2 contracts, test runner, processes, supervisors, Mo.Sim scheduler: the milestone | interpreter-step-1 … interpreter-step-4, 08-milestone.md |
Step history (the overnight run of session 5 to step 33, every row done; newest first)
| step | what | measures |
|---|---|---|
| 5.5 | done: the formatter | interpreter-step-5 |
| 5.6 | done: main and Mo.Server | interpreter-step-6 |
| 6 | done: program 2, logstat, in Mo from a spec | program-2 |
| 6c | done: the control run in Go and Python | control-run |
| 7 | done: programs of many modules, memory, speed | interpreter-step-7 |
| 7b | done: the stdlib, 09-stdlib.md | interpreter-step-8 |
| 8 | done: Mo.Sim seeds and faults | interpreter-step-9 |
| 9, 10 | done: sidecar, mo fix, error catalog, README | interpreter-step-10 |
| 11 | done: Net, program 3 kv, the runtime under real programs, two defaults undone | interpreter-step-11, program-3, interpreter-step-12, interpreter-step-12b |
| 12 | done: control run round 2, Mo 11.8 min from 25.5 | control-run-2 |
| 13 | done: the C backend, native logstat 7× the interpreter | interpreter-step-13 |
| 14 | done: follow-ups, contracts in every build, formatter shapes, stdlib rows (interpreter-step-14) | round 3 of the control run |
| 15 | done: processes and Net in the C backend; native kv 1.4× the interpreter at half the memory (interpreter-step-15) | — |
| 16 | done: HTTP in the stdlib, native httpd 1.7× the interpreter (interpreter-step-16) | — |
| 16b | done: round 3 of the control run; Mo’s checks caught a real bug, six syntax loops, timing void (control-run-3) | — |
| 17 | done: round 3 follow-ups (interpreter-step-17) | — |
| 18 | done: the outside review’s no-compat fixes (interpreter-step-18, outside-review-2026-09-13-response); the failure model in chapter 3 | — |
| 18b | done: program 4, notes; native 23,692 gets/s with 32 clients; the recipes saved a design and exposed the conformance gap (program-4) | — |
| 20 | done: the runtime owns the loop, 11 fictional bounds → 0, every expected file unchanged (interpreter-step-20) | — |
| 20b | done: round 4, timing valid, Mo 16.1 min to Go’s 9.3 and Python’s 9.0, loops 5/1/0, the laws kept (control-run-4) | round 5 after program 1 |
| 21 | done: green threads in both runtimes, 65,530 idle connections from 8,000, a process at rest half its size, the four chapter 7 bets measured, Fs.fixture() refuses .., three diagnostics (interpreter-step-21) | program 1 |
| 19 | done: what program 4 found; 200,000 processes at 9 MB native, the deadlock a report, --recipe, mutation tests 9 of 10 (interpreter-step-19) | — |
| 33 | done: step 29b: memory bounded by what an update reaches, in every loop shape; a process region reserves address space and a walk compacts the frames waiting in calls; the real 1M log replays at 2.3 GB where it was killed past 9 (interpreter-step-29b) | step 30 |
| 32 | done: step 29: restart: :never honoured with its report, exit past a delayed send, replay compacting in generations (1M native 716 → 81 s), simulated time only when a test waits, invariants counted on the verified: line; Fable’s 1M probe on a real log passed 8 GB after the fold, to interpreter-step-29b (interpreter-step-29) | step 29b |
| 31 | done: program 6, ledger, 92 min, 14 modules, 1,168 transfers a second native; four invariants kept; a restart: :never bug (program-6) | round 8 |
| 30 | done: step 28: a map written in place by a move analysis in both backends, the tuple accumulator moved, the never rule through branches, regions giving pages back, six gaps (interpreter-step-28) | program 6 |
| 29b | done: round 7, pre-registered on Robert’s measure, held on all four (control-run-7) | step 28, program 6 |
| 29 | done: step 27: the file law gone, state/result/old as names, a never reads values at rest, \u{X}, fold_lines past a bad line (interpreter-step-27) | round 7 |
| 28b | done: round 6, pre-registered, failed on all four predictions: 1.58 times Go, a shape-law loop, no check caught a bug in any language, 0.74 of Go’s lines (control-run-6) | Robert’s reading of the laws |
| 28 | done: step 26: the one-line if in tail position is the value, the keyword sentence for state and old as names (interpreter-step-26) | round 6 |
| 27 | done: step 25: the one-line if as a value in both runtimes, state and old as field names, the if hole closed, agent narrowed to its spec, the suite’s first Linux run green (interpreter-step-25) | step 26, round 6 |
| 26 | done: step 24: the read-only Fs hole closed, a handle in a state field, delay: on send, Deadline.remaining (interpreter-step-24) | program 6 |
| 25 | done: program 5, agent, 77 min, 19 modules, 448.7 five-step runs a second native, verified by a 22-check session (program-5); the budget model works where an asker exists; an authority hole and the handle law’s cost found | step 24, round 6 |
| 24 | done: step 23, the runtime surface: platform.runtime, the event ring, mo run --surface; six of jobq’s nine questions answered in full (interpreter-step-23) | program 5 |
| 23 | done: step 22, the derived deadline reply_by in both runtimes, jobq’s hand-written sums gone, the --recipe line count, the recipe’s rewrite rule (interpreter-step-22) | program 5 |
| 22b | done: round 5, pre-registered, mixed: Mo 14.8 min to Go’s 11.6 (1.28, from 1.73), loops 5/2/3, no shape-law loop (control-run-5) | round 6 after program 5 |
| 22 | done: program 1, jobq, 40 min, nine modules, 4,051 lease-and-ack pairs a second native, verified by a 29-check session (program-1); literal deadlines lie where they nest; invariant kept none of eight candidates | round 5, step 22 |
| later | the package registry (d34-packages-are-recipes); mo prove | deferred, see the phases table |
Session 3 note
Robert: spec/design-v0 is a folder of files, one per chapter. Kept.