v1 · In Build · Internal Pilot Q2 2026
Axiom Zenith · System 05
CRUCIBLE
Job Description · Pricing · Reasoning
Price any role, any company, any location — in seconds. Crucible forges defensible compensation bands from government filings, pay transparency data, and live market signal, then synthesizes the reasoning so every number can be defended in a client conversation.
Run a price Sign in for production →
01 · Engine

Four-stage reasoning pipeline

Each price moves through four stages — decompose, retrieve, ensemble, synthesize. Cheap models do the structural work; the synthesis model only sees what matters.

~5–10s end-to-end
$0.08 · avg cost
STAGE 01
Decompose
Claude Haiku · JSON
JD text → structured role, seniority, scope, skills, SOC code, and Crucible canonical level. Strict schema, zero preamble.
ModelHaiku 4.5
Cost$0.003
P501.4s
→
STAGE 02
Retrieve
pgvector + SQL · top-30
Three parallel queries: semantic neighbors via text-embedding-3-large (Azure OpenAI), company-specific comps, geo-role BLS baseline. Weighted by recency, employer, MSA.
Enginepgvector HNSW
Cost$0.001
P500.6s
→
STAGE 03
Ensemble
Statistical estimators
Trimmed-mean comp statistics, geo-adjusted BLS baseline, company-tier multiplier. Three independent estimators, surfaced contradictions noted.
MethodRobust stats
Cost$0.000
P500.1s
→
STAGE 04
Synthesize
Claude Sonnet · final band
JD + decomposition + top-10 comps + estimator ensemble → final low/mid/high band with reasoning, confidence drivers, and explicit adjustments.
ModelSonnet 4.6
Cost$0.076
P504.2s
02 · Run

Try a live price

Three sample roles preloaded. Click “Price this role” to watch the pipeline execute and the band assemble in real time. This is a sandbox simulation — for real pricing, sign in.

Sandbox · Mock data
Production at crucible.52-160-145-218.nip.io
Input · Job Description
Preset
Company
Location
JD Text
Output · Pipeline Trace
Idle
Press “Price this role” to run the pipeline
1.4M
Comp records
H-1B + PERM + pay-transparency, last 36 months
5.2s
P50 latency
JD in → band out, end-to-end
$0.08
Cost per price
Haiku decompose + Sonnet synthesize
±9%
Band tightness
Backtested vs realized offers, n=312
03 · Sources

Trusted, defensible data

No scraping of Levels.fyi, Glassdoor, or LinkedIn. Crucible runs on government filings, legally compliant pay-transparency postings, and a feedback loop from realized offers.

v1 sources · 4
v1.1 sources · +3
DOL H-1B + PERM
0.90
Source · 01 · Government
3 years of LCA disclosures and PERM filings. Verified employer, worksite, SOC code, and prevailing wage rate. The defensible spine of every band.
~860K rowsLIVE
BLS OEWS
0.85
Source · 02 · Government
Occupational Employment & Wage Statistics. Per SOC × MSA percentiles (10/25/50/75/90). Used as the geographic baseline and sanity check.
~410K cellsLIVE
Pay Transparency
0.95
Source · 03 · Compliance
CA, CO, WA, NY, IL, MD, HI, DC, NJ. Legally-required disclosure tied to actual JD text. Highest-confidence source; powers the semantic retriever.
~140K rowsLIVE
SEC EDGAR
0.95
Source · 04 · Filings
DEF 14A proxy statements. Named executive officer comp. Anchors the top of band for public companies and reveals comp philosophy.
~28K NEOsv1.1
O*NET Taxonomy
—
Source · 05 · Ontology
SOC code normalization, skill ontology, occupation tasks. Not a comp source — the scaffolding that lets every other source speak the same language.
1,016 occsLIVE
Feedback Loop
1.00
Source · 06 · Internal
Every recruiter validation, every realized offer, every "your band was off" feedback. Closes the loop. Drives calibration and drift detection.
Compoundingv1.1