MathSphere Nexus
When a model writes code for a person, who checks it — and what may the checker do about what it finds? The Ladder is the house’s answer, in three rungs:
Rung 1 decides. Deterministic checks — compile, lint, a bounded run judged by exit code — and nothing else. No model sits on this rung. Its verdict is the decision.
Rungs 2 and 3 advise. A model reviews the candidate (Rung 2), and a model checks it against the task’s stated requirements, line by line (Rung 3). Their verdicts are shown to the person. They are never a veto: advice informs the person’s consent, and the gate — the prototype’s only write path — acts on the person’s word, not the model’s.
The budget lives at the grant. Review costs are drawn against a grant the person created, and the check stands before the model is called — an exhausted grant refuses the call, receipted, instead of discovering the cost afterward. And both ways out of a repair loop hand work back honestly: success returns the passing candidate; exhaustion returns the last candidate labeled failed, with the full record of every try.
The prototype was built to test five claims: (1) Rung 1 is deterministic. (2) Reviewers cannot leak into decisions — structurally, and in behavior. (3) A same-seat verdict (the generator’s own model, reviewing) differs measurably from a different seat’s. (4) The budget is enforceable at the grant, not the loop. (5) Both exits hand back correctly.
In slices, each built, run live, and receipted before the next began: Slice A — Rung 1 alone, with check sets for Python and Go, and a Rust set run as an extension. The gate — the only write path, proven on six scenarios. Slice B — the reviewing rung. Slice C — the fit rung. The trials — 30 candidates written by one model (qwen2.5-coder:7b) over a fixed ten-task set, each judged by Rung 1, by a different seat (llama3.2:3b), and by the generator’s own model in a fresh call. Trial slice 2 — the same 30 candidates re-judged by a larger different seat (llama3.1:8b), with every candidate hash, verdict hash, and prompt hash asserted against the first run before a single verdict was taken. The ground-truth check — every candidate actually run, its output tested against what its task asked for. Slice D — budgets at the grant, and both exits.
Determinism (Claim 1). Each fixture was judged 100 times: one verdict hash per fixture, every time. The Python verdict hashes matched across two machines and two Python versions. Slice A was closed at 10 of 10 by the owner’s ruling; the Rust set stands beside the close as an extension run — 15 fixture receipts in two chains, the gate’s chain a third.
Isolation (Claim 2, structural half). The review prompt is assembled from exactly three inputs — the task, the candidate, and Rung 1’s verdict sheet. The generator’s explanation has no parameter, no field, and no side channel into it; the isolation is a shape, not a policy. A verdict that is not exactly PASS or FAIL: reason is receipted as no verdict, with the model’s raw answer preserved. Slice C produced the proof of that discipline live: the seat answered by echoing the template’s placeholder instead of the requirement number, and the contract held it as no verdict — the receipt (read aloud and checked, field by field) shows both contract breaks, preserved, not smoothed over.
The trials (Claims 2 and 3, behavioral half). Rung 1 passed 28 candidates and refused 2. Against Rung 1, the different seat agreed 22 times and falsely failed 5; the same seat agreed 30 of 30. The five false fails were then adjudicated one by one against the tasks: all five were false alarms — including two different invented defects for two candidates that were byte-for-byte identical. The same seat’s perfect agreement, read carefully, is concordance with the decider — redundant, not useless; the different seat’s disagreements were the signal worth examining, and on examination they were noise.
The second seat. The larger seat (llama3.1:8b) told a different story: zero false fails, the identical-bytes pair judged consistently and correctly — and weaker contract discipline, with 6 no-verdicts against the small seat’s 3. Size removed the invented-defect noise and loosened the format discipline. Both facts are in the record; neither is rounded away.
Ground truth. The deterministic check ran all 30 candidates and tested their output: 30 correct, 0 defective — the generator had written 30 correct programs, so no reviewer catch was possible, and none is claimed. Scored against truth instead of against the decider, the table reads:
Budgets and exits (Claims 4 and 5). Slice D passed on the stub and then live against a real seat: a grant refused the second review before the call was made, receipting the refusal with the remedy; the repair loop handed back the passing candidate on success, and on exhaustion the last candidate labeled failed, with the full nine-try record; the oracle loop kept the same discipline over five rounds. 29 receipts, chain verified.
The verdict lines and tables from the runs’ own output, November 15–16, 2026, on the owner’s machine:
SLICE A: closed at 10/10 by the owner’s ruling (Rust run as extension) STEP 3 (THE GATE): PASS — twice, six scenarios, chain verified SLICE B: PASS — chain over 6 receipts SLICE C: PASS — chain over 7 receipts TRIALS RUNNER: PASS — 30 receipts, chain OK SLICE D: PASS — stub run, then live run; 29 receipts, chain OK TRIALS 2 RUNNER: PASS — 30 receipts, chain OK, all evidence hashes asserted GROUND TRUTH CHECK: PASS — 30 receipts, chain OK ground truth: 30 correct, 0 defective Rung 1 vs truth: misses []; false refusals [8, 27] different seat (llama3.2:3b): catches=0 false_alarms=6 correct=21 abstentions=3 same seat (qwen2.5-coder:7b): catches=0 false_alarms=2 correct=28 abstentions=0 different seat (llama3.1:8b): catches=0 false_alarms=1 correct=23 abstentions=6 [30 of 30 trials scored from the named second chain]
The parse — how each proved thing is used. To have code decided, Rung 1 runs it: compile, lint, a bounded run — the verdict is the decision, and its hash is reproducible by anyone holding the candidate (determinism, Slice A). To have it reviewed, Rung 2 reads the task, the candidate, and the verdict sheet — never the generator’s account of itself (isolation, Slice B). To check fit, Rung 3 answers requirement by requirement, and any answer off the contract is recorded as no verdict with the raw text kept (Slice C). To spend review, a grant is drawn down before the call, and exhaustion refuses in the open (budget at the grant, Slice D). To leave a repair loop, you carry out either the passing candidate or the failed one, labeled, with every try on record (both exits, Slice D).
Receipts, hash-chained — every slice, trial, review, refusal, and handback above is a receipt in a chain that verifies; earlier chains are moved aside, timestamped, never deleted. Standard library only — the prototype imports nothing outside Python’s standard library; the models are reached over the local model server’s HTTP interface. Advise, never veto — no model’s verdict can refuse, pass, or write anything; the gate writes only on the person’s act. Demo discipline — this page is the documentation half; the slices’ own test runs are the live half.
Honesty about scope. The trial set contained no defective candidates, so a reviewer’s ability to catch a real defect is untested — not failed, and not proven; a set with planted defects is the instrument that would test it. Reviewer quality beyond these seats, tasks, and this machine is unmeasured. And the prototype is not the REPL: it is advance evidence for the REPL build (roadmap Phase 10), which will stand or fall on its own runs, in its order.
A command is what you type. The name is what the project is.