Research agenda, Sep 2026: Fable’s response to the contradictions

Robert merged the research agenda on 13 Sep with twelve concept pages from its runs R1, R3, and R6: two on the LLM literature, two on the safety and reliability traditions, and eight author profiles. Each page ends with a table marking ideas “contradicts Mo”. Fable read all twelve the same night, after round 4 of the control run, and made a call on each contradiction: agree (Mo changes, or already did), disagree (Mo stands, with the reason), or test (a named experiment decides). Rows for anything that changes something are in decision-log under this date.

Three contradictions were already settled before the pages were written. Step 20 removed the fictional loop bounds Dijkstra warns against (interpreter-step-20). Chapter 3’s failure model answers Armstrong’s R6 with the store recipe, though not by that name. Round 4 re-evaluated the shape laws against three rounds of loops (control-run-4). The pages were written from the vault as it stood at noon; the calls below are against the tree as it stands at midnight.

The calls

contradictionsource pagecall
Invariant synthesis is cheap, so verifiability is a weak reason to ban whileagents-and-verification-2026agree on the premise. The ban was never argued from verifiability in chapter 2, and after step 20 the runtime is the loop. The rationale is legibility and the removal of fictional bounds; round 5 tests it.
Numeric shape laws as model-performance levers; complexity metrics do not predict LLM performancelanguage-design-for-llms-evidenceagree, no change. The shape laws were never a model-performance claim. Their justification is human review economics and one declaration per ID edit; rounds 2 to 4 tripped none. Round 5 keeps counting.
Shape numbers are “should” rules at JPL; MISRA has deviationssafety-critical-coding-standardsdisagree. Robert has twice chosen no override (Q16; 13 Sep). MISRA’s tighten-only ratchet is already Mo’s. A deviation record is an escape hatch by another name. What Mo takes is a ledger: any program a law blocks, as opposed to a missing row, is recorded as a Q16 event. Zero across four programs and four rounds.
A pure, no-while shape carries a measured error-rate risklanguage-design-for-llms-evidencetest. Round 4’s loops (Mo 5, Go 1, Python 0) are consistent with the risk and every one was a grammar first guess. Round 5 counts loops by cause after the ergonomics step.
Agents may write a Mo generator instead of Molanguage-design-for-llms-evidenceagree. Round 5 records whether each worker wrote Mo directly. So far every round did; the worktrees hold the evidence.
Deadlines should be inherited and tightened, not restated per callreliability-and-testing-philosophiestest. The law stays: every call that can wait carries within:. Program 1 counts the deadlines it writes and how many are derived from an enclosing one. If literals lie, an inherited budget is designed then.
Annotated non-terminating loop instead of a bansafety-critical-coding-standardsagree, already in Mo. Ravenscar’s “all tasks are non-terminating” is a Mo process under step 20’s runtime: the runtime loops, the process never does.
Drop the nine-nines claimreliability-and-testing-philosophiesagree. erlang cited it until 17 Sep 2026, when the number was struck there; no other Mo page outside the raw runs does. Rule: Mo’s materials cite Erlang’s practice and scale, never the number.
Bounded loops need a real termination argument; while ban satisfied by a fake constantauthor-edsger-dijkstraagree, done in step 20 for loops. For recursion chapter 2 still promised “structural only, proved terminating”, which the toolchain never did: it bounds depth at 10,000 and crashes. The chapter now says what is true; termination is mo prove’s obligation.
The verified: line implies proof from testsauthor-edsger-dijkstradisagree. The line’s vocabulary already names each obligation and prints proven: not run beneath it.
R6 stable storage is unansweredauthor-joe-armstrongagree that it must be named. Chapter 3 already answers it: restart recovers the service, never the state; a process whose state must outlive a crash writes through a capability and replays on start; the store recipe is the pattern. Chapter 3 now names R6. Program 1 is the test.
Message passing is unreliable in Erlang; Mo’s sends crash on overflowauthor-joe-armstrongdisagree. Inside one program a send is reliable or a bug. Across a socket it is unreliable, and --faults injects Closed, Timeout, and Idle. Mo has no distribution (d08-beam-qualities-without-the-beam).
A 500-line file law re-imposes the module placement problemauthor-joe-armstrongdisagree. The unit of identity is the declaration (d29-edit-by-declaration-id); the file is a formatter concern. No round or program moved a function for the law.
Source-level scrutiny cannot establish trust; Mo inherits Zig and Cauthor-ken-thompsonagree. d03-source-carries-its-evidence is about programs, not the toolchain. Reproducible builds of mo and of mo build’s binaries become a release requirement, measured tonight (below), before any registry.
Large-system help is interfaces and dependency structure, not capsauthor-dennis-ritchieagree, already in Mo. expose lines, capability parameters, no import cycles, and recipes are the structure; the caps bound review, not systems. Program 1 is the first service large enough to tell.
”There is no escape”author-brian-kernighandisagree, Robert’s call. The platform is the escape. The Q16 ledger above is what would reopen it.
Only the user should introduce inefficiency; intent at every levelauthor-tony-hoaredisagree on the first, Robert’s call; step 21 re-measures the contract cost. On the second, bodies may hold assert and every module may hold never; the altitude fixes what a human reads, not what an agent may write.

What changed tonight

  • Chapter 2’s recursion line: “structural only, proved terminating” became a depth bound of 10,000 and a crash, with termination named as mo prove’s obligation. The toolchain had done this since step 18; the spec was the lie.
  • Chapter 3 names Armstrong’s R6 beside “What restart means”.
  • Three rules, each a decision-log row: the Q16 ledger, the nine-nines rule, and reproducible builds as a release requirement.
  • Round 5’s protocol gains two columns: loops by cause, and whether the worker wrote Mo directly.
  • Program 1’s brief gains a count: within: literals written, and how many derive from an enclosing deadline.

Reproducible builds, measured

Two builds of mo from the same tree with separate Zig caches and global caches, then kv built once by each mo, compared by SHA-256.

artifactresult
kv.c, mo_rt.c emitted by the two mo binariesidentical
the kv binary from eachidentical
the two mo binariesdiffer in 114 bytes: the Mach-O LC_UUID and the ad-hoc code signature it feeds; no path is embedded

So mo build is reproducible today, and mo itself is reproducible up to the macOS linker’s UUID, which is not the toolchain’s doing. A Linux static build is expected to be byte-identical and is measured at the registry step, when reproducibility becomes a release gate.

What the pages strengthen, taken without a row

The dependency closure of a declaration as the unit of context an agent gets, not the file (the composition frontier at two to five dependent functions is exactly program 1’s shape); hand-written diagnostics beating generated ones; the reachability obligation for fault tests, which MO0327 and --until already are; shrinking a failing --sim seed as a tier 3 item; the feature-cost ledger Wirth and Hoare both ask for, which the decision log’s semantic rows are becoming.