Evidence → adversarial analysis → probability → decision → audit
Decision intelligence that has to earn its answer.
KAHRELUM OS is an evidence-to-decision system for forecasting, research and high-stakes analysis. It separates fact from inference, attacks its own conclusions, and is designed to block rather than manufacture certainty when the evidence does not support an answer.
Current formal release · v2.3.4 FINALCommercial status: Private Pilot / Prospective Validation
Request a PilotReview the Evidence
External adversarial audit
Identity and arithmetic hold. The limitations stay attached.
The delivered audit-package SHA-256 matched exactly. Registered forecast metrics, paired differences, bootstrap intervals, analytical arithmetic and composites reproduced from the frozen inputs.
Forecast evidence
Two benchmarks. Two different readings.
| Benchmark | KAHRELUM | GPT | Comparator |
|---|---|---|---|
| RP100 mean Brier | 0.222593 | 0.321955 | Gemini 0.189073 |
| GX50-v2 mean Brier | 0.146176 | 0.204458 | Grok 0.203929 |
RP100 KAHRELUM−GPT paired Brier difference: -0.099362, 95% CI [-0.144114, -0.058405]. GX50-v2 KAHRELUM−GPT and KAHRELUM−Grok primary 97.5% intervals both include zero.
Where it fits
Four ways to test KAHRELUM.
Unresolved questions, explicit probabilities, calibration and resolution without hindsight rewriting.
Evidence → alternatives → adversarial challenge → probability → action, with revision conditions preserved.
Frozen side-by-side evaluations under declared protocols, scoring rules and custody boundaries.
Test whether incomplete or unsupported conclusions are blocked before they become decisions or published claims.
What an output looks like
Not just an answer — an inspectable decision state.
QUESTION
Will a defined event occur before the frozen horizon?
FACT
Verified evidence available at the declared cutoff.
INFERENCE
Mechanism connecting that evidence to the event.
FORECAST
62% — example format only, not a live forecast.
BLACKGLASS
Strongest countercase and evidence that would invalidate the thesis.
COMMAND
MONITOR — no action until the declared gate clears.
BLOCK CONDITION
If the required evidence object cannot be verified, do not manufacture closure.
Canonical pipeline
ROUTER → SCOUT → SOURCEGRID → K-ALIGN → IPR → BLACKGLASS-I → CRF → COMMAND → BLACKGLASS-II → RECORD LOCK
MCRI compiles explicit requirements into a mission contract. AAIK governs evidence, instability and exposure. RECORD LOCK preserves what was known, inferred and unresolved instead of rewriting history.
Release identity
Hash-bound, not hand-waved.
v2.3.4 release archive SHA-256:c6caca3290abbee7121fb743fabb1210b59b7536aa46dd6d6874cd02962bc1e1
External-audit package SHA-256:084f0d1348342e707e1fc953b1337d3a805f9890301fa54ba01983aa1a46876f
The older v2.3.3 exact-byte promotion record remains preserved separately; its hashes are not presented as v2.3.4 identity.
Independent prospective validation
You control the next test.
The next evidence-changing step is an externally controlled prospective benchmark. The preferred protocol gives the evaluator control over case selection, cutoff dates, preregistration, resolution criteria, gold outcomes and scoring.
A negative result stays in the record. The goal is not to design a test KAHRELUM is supposed to win; it is to create a result that is credible either way.