SIMARCSimulation Architecture & Research Computing For agents →
The instrument

The apparatus — systematic error control

The distinguishing feature of this institute is not the hardware. It is the machinery for not fooling ourselves. Simulation and exhaustive search produce confident wrong answers easily, and most of the engineering effort here has gone into making that harder. What each instrument refuses is published. How it works is not.

Design principle

Instrument specification precedes computation

Each program specifies its apparatus before any result exists, answering one question: what should the instrument be, specifically for this? Nothing is copied wholesale between programs. What counts as convergence, as completeness, or as a valid comparator is redefined per domain, and a rule that earned its place in one program is refused in another when the domain does not support it.

The measure of whether this works is not elegance. It is catches before results exist: in one program the apparatus surfaced four real defects, three of them before any result had been produced.

The instruments

Instrument set

GATE 01
Desk gate — prior art before compute No target consumes core-hours before a prior-art sweep conducted in the field's own vocabulary rather than ours. It refuses work the literature has already settled, and it is calibrated per domain, because what counts as a genuine gap differs by orders of magnitude between a small field and one with millions of papers.
GATE 02
Sealed predictions — with a power gate Numerical predictions are SHA-256 hashed and published before the deciding computation, with explicit falsification bands and stated confidence. The gate additionally refuses any seal the run could not falsify even in principle — a prediction whose measurement lacks the power to contradict it scores as a success by construction, and is not filed.
GATE 03
Evidence ladders — grammar, enforced Different kinds of claim wear identical sentences. The ladder forbids a record from using vocabulary above the experiment that decided it, and the rungs are domain-specific. In the medical programs the rule reads: you cannot treat a correlation.
GATE 04
Machine-checkable certificates Results are published with proof or exact-interval certificates wherever the shape of the problem allows, so a referee can re-check the result without trusting our code or re-running our solvers. Where a free independent oracle exists we take it; where none exists, we pay for one.
GATE 05
Complete censuses, or no verdict Census code refuses to report a result unless it visited its entire domain. Filters are recorded as attributes rather than exclusions, because a census can be complete over its outer loop and a silent sample of its inner one.
GATE 06
Baselines and nulls — the comparator is the experiment Any claim that N features are significant is published with the fraction of all features clearing the same bar. The same count means opposite things against different backgrounds, and the count alone cannot distinguish them. Size-matched nulls and effective-n accounting sit alongside it.
GATE 07
The crew protocol — because a rule in prose does not install Several sessions work one program tree at once, and every failure available to that arrangement is silent: two agents editing one file, two banking the same conclusion, a rule followed only by the agent that proposed it. The protocol is a program rather than a README section, and its check exits non-zero rather than advising.
GATE 08
Fleet doctrine — invisible failures are the expensive ones A mandatory preflight measures throughput on a bounded pilot and refuses to bless a job whose extrapolated wall clock exceeds its declared budget. A watchdog makes stalled, crash-looping and idle-with-queue states loud, and deliberately does not kill ambiguous cases.
The lessons database

Failure corpus

Every instrument failure is written up with the incident behind it and an applicable rule. The corpus runs to more than eighty entries across four programs and is held internally — it is the accumulated cost of the apparatus, and it is what the apparatus is for.

The test of whether an apparatus compounds is whether a lesson ever catches a failure without a human noticing first. Across programs, five have — and one of those was caught by a check that existed only because an earlier lesson demanded it. In one program a failure was prevented outright by a lesson recorded an hour earlier in another.

The characteristic failure this apparatus is built for does not crash. In one program, fourteen of sixteen recorded failures produced plausible, well-formed, entirely wrong output, and every one of them ran to completion successfully.

What it has killed →