Outside review, 13 Sep 2026: Fable’s response and the baselines

Fable read outside-review-2026-09-13 and its evidence the afternoon it landed, re-ran three of its probes on the toolchain after step 17 (recursion still exits 134 with a Zig trace; a closure still writes through a captured Out and the dropped map value is not flagged; the grammar has no grouped patterns), and made the calls below. Robert’s decision the same hour: the laws and the no-while rule stay as they are, “we need to test and see what happens and can reevaluate later”, and every finding and decision is documented so the numbers serve as baselines.

The calls

review itemFable’s callwhere
invariant polarityagree; flip to “stays true” now, no compat coststep 18
handles are not capabilities; capture in closuresagree; Handle(T) is authority, capture refusedstep 18
recursion crashes with a Zig traceagree; a depth limit and a Mo crash reportstep 18
grouped patternsagreestep 18
map and set equality by orderagree; by contentstep 18
the sidecar’s cache keyagree; transitive bodiesstep 18
vacuous fault testsagree; --faults --until F, then rewrite kv’sstep 18, program 4
self-reported never witnessesagree; mutation-test the contract machinerya later step
update as a transaction; the failure modelagree the promise overreaches; Fable writes the failure modelbefore program 4’s --sim claims
deadlines as budgetsagree; program 4’s experimentprogram 4
governance of worker defaultsagree in part: a semantic tag and an expiry, not an approver before acceptancethe log, from step 18
shape laws as policy with an overridedisagree; Robert: stay, measurecontrol runs
no whiledisagree; rows that drive their own loops first; Robert: stay, measure. Program 4 made it three programs with a fictional bound; Robert (evening) chose the runtime owning the loop over a loop keywordinterpreter-step-20
default parametersRobert rejected them on sight; stays—
recipes as securitythe review says what direction 34 says: unproven; program 4 tests itprogram 4
spec altitude replacing reviewagree on a risk rule for persistence, auth, concurrency; the altitude stays the default viewthe failure model
Roc’s platform precedentagree; Mo.Server ships first-party and so will the restroadmap

The baselines on the day of the review

These are the numbers a later re-evaluation compares against.

measurevalue
control run round 3, loops to greenMo 9 (6 syntax diagnostics, 1 real bug, 2 test mistakes), Go 2, Python 1
control run round 2, wall-clockMo 11.8 min, Go 8.2, Python 7.9 (round 3’s timing void, the machine slept)
checks that caught a real bug, all roundsMo 2 (round 3: a test, a never), Go 0, Python 0
lines per function, round 3Mo median 6 (bodies 4), Go 9.5, Python 5
program and test lines, round 3Mo 851, Go 1,387, Python 1,064
corpus66 modules and programs, 142 files under test, suite 22 s
native against interpretedlogstat 5×, kv GETs 1.4×, echo 1.5×, http 1.7×
kv after 50k SETs19.2 MiB native, 37.4 interpreted
decision-log rows218, of which 4 overturned

What changes because of the review

Step 18 (interpreter-step-18) makes the no-compat fixes. The decision log gains the semantic tag. Fable writes the failure model before program 4’s simulation claims are read. Program 4 tests budgets and recipes. The laws are re-evaluated after round 4 of the control run, against this page.