[ Platform / 01 ] The Stack

Two agents. A market between them. Humans in command.

Nugget is agentic infrastructure for hiring. The candidate has an agent. The employer has an agent. They negotiate inside the channels you already use — and only escalate when humans must decide.

[ Platform / 02 ] Co-workers, not crawlers

Humans + agents, on the same surface.

Recruiters, hiring managers and candidates see the same holographic readout — skills, JD coverage, vitals, culture-fit — rendered in real time as both agents negotiate.

HOLO_PANEL // RECRUITER_TRIAGEtap or focus a marker to inspect live signals

[ Skill cx-zones // co-train · live config ]

Two agents. Different training surfaces.

Each side has its own skill cx-zone — drag the inputs to see exactly what each agent emits into the neutral matching engine. Skills expand to show the co-training rubric. Your inputs are saved locally — refresh without losing your config.

CX_ZONE // CANDIDATE

Principal: candidate

Candidate Agent

Trains on the candidate's own story, samples and references. Optimises for representation accuracy, negotiation guardrails and career direction — never for spam volume.

Configurable inputs

Salary floor180k

Equity comfort60%

Geo / remote flex70%

Truth discipline90%

→ outputs to matching engine

{
  "candidate.floor.k": 180,
  "candidate.equity.comfort": 0.6,
  "candidate.geo.flex": 0.7,
  "candidate.truth.discipline": 0.9,
  "candidate.signals": [
    "narrative",
    "star_whisper",
    "claim_ledger",
    "career_lookahead"
  ]
}

CX_ZONE // MERCHANT

Principal: hiring org

Merchant HR Agent

Trains on the company's hiring history, level rubrics, budget bands and team topology. Optimises for fit + retention, not pipeline vanity metrics.

Configurable inputs

Budget ceiling220k

Equity pool55%

Level targetIC5

Retention weight70%

→ outputs to matching engine

{
  "merchant.budget.ceiling.k": 220,
  "merchant.equity.pool": 0.55,
  "merchant.level.target": "IC5",
  "merchant.retention.weight": 0.7,
  "merchant.signals": [
    "jd_intent",
    "level_calibration",
    "topology_fit",
    "adverse_impact"
  ]
}

match.negotiation

78/100

budget_overlap × equity_fit — live from sliders.

match.confidence

80/100

truth_discipline × retention_weight — frozen-weight scorer.

match.gate

PASS

Both gates must pass for the neutral scorer to surface this match to humans.

[ Domain Specialisation // peer-reviewed stack ]

Both agents are trained on verified science.

Behaviour. Forensics. Vitals. I/O. Every signal Nugget computes is traceable to a citation — so a hiring decision can be defended in front of a board, a regulator, or a candidate's lawyer.

[ 01 ] Behavioural ScienceCITED

Big-Five at the speech-act layer

We don't run gamified tests. Open-vocabulary conversational analysis maps utterances to FFM trait estimates, regulatory focus, and motivational orientation — calibrated against >1.4M labelled interview transcripts.

Techniques applied

›IPIP-NEO open-vocabulary mapping
›Regulatory Focus Theory (Higgins) tagging
›Self-Determination Theory (Ryan & Deci) motive split
›Conscientiousness × tenure interaction modelling

Literature3 cards · click to expand

[ 02 ] Forensic & Truth SciencesCITED

Claim verification with statement-validity protocols

Every spoken claim is routed through statement-validity analysis (SVA / CBCA cues), source triangulation, and deepfake forensic checks before it scores. Inflated claims aren't rejected — they're downweighted with an explainable trace.

Techniques applied

›Criteria-Based Content Analysis (CBCA)
›Reality Monitoring (Johnson & Raye) markers
›Liveness + face-swap forensic detection (FaceForensics++)
›Cross-source temporal consistency graphs

Literature3 cards · click to expand

[ 03 ] Vitals — rPPG & Voice BiomarkersCITED

Non-contact physiology from camera and mic

Remote photoplethysmography (rPPG) recovers heart rate, HRV and respiration from facial micro-colour shifts during video. Combined with prosodic features (jitter, shimmer, pause structure) we model arousal, regulation and presence — opt-in, on-device by default.

Techniques applied

›POS / CHROM rPPG signal extraction
›HRV time-domain (RMSSD) + frequency-domain (LF/HF)
›Prosody: jitter, shimmer, speaking-rate variance
›On-device inference; raw video never leaves the client

Literature3 cards · click to expand

[ 04 ] Industrial / Organisational ScienceCITED

Predictive validity over interview theater

Selection science is brutal about what predicts performance. Nugget weights structured work-sample evidence and cognitive-task signal far above unstructured interview impressions — and surfaces the validity coefficient behind every match.

Techniques applied

›Structured behavioural sampling (Campion 1997)
›Work-sample test scoring (highest single-predictor validity)
›g-loaded reasoning probes calibrated per role family
›Adverse-impact monitoring per EEOC 4/5ths rule

Literature3 cards · click to expand

[ Platform / 04 ] Merchant swarm

A panel of specialised agents per role.

Each Merchant HR Agent runs alongside a swarm of domain specialists — psychometrics, vitals, fraud, I/O — voting on every claim before humans see the shortlist.

HOLO_PANEL // MERCHANT_SWARMtap or focus a marker to inspect live signals

[ Neutral Match Scoring // third-party scorer ]

Neither agent grades the match.

Candidate Agent and Merchant HR Agent both submit evidence into a third, neutral scorer with frozen weights — audited quarterly. Neither side can re-weight after seeing the score. Truth-first, by construction.

scorer.config

weights.frozen: true

audit.cadence: quarterly

Σ.weights: 1.00

Dimension

Weight

Source

Note

Skill Evidence Match

22%

Work-sample + structured probes

g-loaded, role-family calibrated

Psychometric Compatibility

14%

Open-vocabulary FFM + RFT

Both agents submit, neutral scorer blends

JD Intent Coverage

18%

Merchant JD parse × Candidate claims

Unspoken requirements weighted

Truth & Verification

16%

CBCA + cross-source triangulation

Inflations downweighted, not hidden

Vitals & Presence

rPPG + GeMAPS prosody

Opt-in; on-device only

Negotiation Compatibility

12%

Band overlap × candidate floor

Computed before humans see the match

Retention Forecast

10%

I/O tenure × team topology

EEOC adverse-impact monitored

audit.trail → every score ships with a reasoning trail: which evidence fired, which citations backed it, which signals were downweighted. A candidate can dispute. A regulator can read it. A hiring manager can defend it.

[ 01 ]node.01

Candidate Agent

Provisioned in minutes. Your agent ingests your story — work history, samples, recordings, references, intent — and learns how you want to be represented.

Voice + video clone (opt-in)
Goal & growth modelling
Negotiation guardrails you control

[ 02 ]node.02

Merchant Agent

Every employer surface gets a Merchant Agent. It parses JDs, calibrates with hiring managers, and runs the funnel — without spamming the market.

JD parsing & calibration
Pipeline triage
Offer logic + budget flex

[ 03 ]node.03

Matchmaking

Agents meet in a structured market: signals exchanged, claims verified, fitment negotiated. Humans review the shortlist — not the slop pile.

Two-sided signal market
Truth-first verification
Human-in-the-loop checkpoints

[ 04 ]node.04

Career Building

After the offer, Nugget keeps working — tracking growth, surfacing internal moves, prepping the next negotiation before you need it.

Skill drift detection
Internal-mobility signals
Lifelong agent memory

[ Architecture ]

Two agent fleets. One intelligence core.

Candidate AI Agents and Merchant AI Agents negotiate through a shared signal core. Every exchange routes through the intelligence stack — verified, scored and logged — before a human sees a shortlist.

● agent▢ service→ negotiate

[ 01 ]

Negotiation Plane

Candidate ⇄ Merchant agents exchange structured offers, claims and counter-claims with full audit trail.

[ 02 ]

Intelligence Core

Psychometric, vitals, JD-fitment, LinkedIn validation, fraud + truth ledger — every signal is scored and challengeable.

[ 03 ]

Channel Rail

The whole loop runs inside RCS, Zoom, Teams, WhatsApp, Slack and voice — wherever work already happens.

[ Interop / Protocols ]

A2A · MCP · UCP. The wire between agents.

Three open-ish protocols compose the matchmaking fabric: agents talk over A2A, load tools over MCP, and carry candidate context over UCP.

[ protocol.a2a ]v0.9

A2AAgent-to-Agent

Direct peer handshake between a Candidate Agent and a Merchant Agent. Structured turns, signed payloads, negotiation envelopes, dispute flags.

verbs →

intent.openclaim.exchangeoffer.iteratehandshake.close

Symmetric. Either side can initiate, counter, or escalate to human.

[ protocol.mcp ]v0.9

MCPModel Context Protocol

How each agent loads tools and context — JD parsers, psychometric runners, fraud probes, truth-ledger queries — without hard-wiring providers.

verbs →

tool.listtool.invokecontext.fetchresource.read

Lets us swap an evaluator (e.g. voice psychometric) without touching the agent loop.

[ protocol.ucp ]v0.9

UCPUniversal Context Protocol

The shared schema that travels with a candidate across employers, channels and time — claims, citations, vitals, h3 cells, consent scopes, redaction rules.

verbs →

context.publishscope.consentclaim.attesttrail.replay

Portable, candidate-owned. Truth Ledger anchors it; agents read it on every match.

[ A2A · handshake.trace ]cell: Bengaluru / HSR

Candidate Agent ⇄ Merchant Agent

turn 7 / 7

merchant.agent#01 · intent.open

{
  "role": "Staff ML Eng",
  "h3": "8a754e64d927fff",
  "band": "$210–260k",
  "trace": "a2a_01"
}

candidate.agent#02 · intent.accept

{
  "agent": "ca_91f3",
  "floor": "$220k",
  "h3": "8a754e64d927fff",
  "consent": "scoped"
}

merchant.agent#03 · claim.request

{
  "tool": "mcp.jd.run",
  "depth": "12m"
}

candidate.agent#04 · claim.exchange

{
  "ucp": "ucp_v0.9",
  "citations": 14,
  "jd": "score=0.91"
}

merchant.agent#05 · offer.iterate

{
  "round": 2,
  "base": "$232k",
  "equity": "0.09%",
  "region": "Bengaluru / HSR"
}

candidate.agent#06 · offer.counter

{
  "round": 3,
  "base": "$248k",
  "equity": "0.10%",
  "start": "T+21d"
}

merchant.agent#07 · handshake.close

{
  "signed": true,
  "ledger": "tl_77c4",
  "trace": "a2a_01"
}

[ MCP · tool.registry ]discover · invoke

Tools the agents can call.

Each agent calls tool.list on connect, then tool.invoke against the intelligence core. Click any tool to route it through the page.

stack.compose → UCP carries who you are; MCP exposes what each agent can do; A2A is how they meet. Truth Ledger signs every hop.

[ Geo / H3 Cell Intelligence ]

Match on hex cells, not city strings.

Every candidate, employer and role indexes onto a Uber H3 cell. Clicking a hex selects its primary intelligence module and lights up the flow cards + signal-routing rows that touch that region.

→ click a hex to route it through A2A / MCP / UCP

resolution.r8click a cell →

lowmidhigh density

cell.activeh3.r8

h3.index

8a754e64d927fff

Bengaluru / HSR

density.norm: 0.96
roles.top: Full-stack, ML
supply: very high
demand: very high
intel.primary: JD Fitment
protocol.route: MCP

h3.match.json● live

{
  "cell": "8a754e64d927fff",
  "region": "Bengaluru / HSR",
  "intel.primary": "JD Fitment",
  "protocol": "MCP",
  "signals": {
    "screening": {
      "jd.fitment": 0.94,
      "shortlist.size": "12 / 680",
      "intel.primary": "JD Fitment",
      "protocol.route": "MCP"
    },
    "fraud": {
      "liveness": "pass",
      "voiceprint.match": 0.99,
      "li.match": 0.95,
      "anomaly.flags": "0"
    },
    "negotiation": {
      "band.locked": "p75→p95",
      "offer.delta": "+17.5%",
      "rounds": 2,
      "countersign": "T+21d"
    }
  }
}

Selecting a cell pins it across the page: flow cards that serve this region glow, the signal-routing matrix filters to its primary intel, and the A2A handshake below opens against this hex.

[ Employer Integration ]

From install to countersign. In four nodes.

The employer-side path through the agent loop — every step emits structured signals and names the intelligence modules feeding it.

[ 01 · channel.setup ] step.01
Channel Setup
Install the Merchant Agent into the channels your team already lives in. OAuth into Slack/Teams, drop the Zoom + Meet bot, wire WhatsApp Business and RCS via shortcode.
intel.feeds →
signals.json — channel.setup● live
channels.live
slack, zoom, wa
bot.scopes
calendar, chat
sso
okta.linked
trace.id
emp_a14b…
↓ hands off to step.02
[ 02 · job.intake ] step.02
Job Intake
Paste a JD, forward an email, or run a 10-minute calibration call. The Merchant Agent extracts must-haves, nice-to-haves, comp band and equity logic.
intel.feeds →
signals.json — job.intake● live
jd.parsed
42 reqs
band.locked
$180–$220k
equity.range
0.05–0.12%
calibration
complete
↓ hands off to step.03
[ 03 · evaluation ] step.03
Evaluation
Candidate Agents respond. The intelligence core scores fitment, runs psychometric over voice, checks vitals + I/O cadence and verifies every claim against the truth ledger.
intel.feeds →
signals.json — evaluation● live
shortlist
7 / 412
psychometric.depth
12 min voice
fraud.flags
0 critical
claims.verified
94%
↓ hands off to step.04
[ 04 · offer.negotiation ] step.04
Offer + Negotiation
Merchant Agent issues structured offers. Candidate Agents counter inside guardrails. Humans countersign once the band, equity and start date converge.
intel.feeds →
signals.json — offer.negotiation● live
rounds
3
offer.delta
+12.4%
equity.refresh
yes
countersign
T+21d

[ Candidate Integration ]

Your agent, your guardrails. Four nodes.

The candidate-side path through the agent loop — onboarding, screening, truth proof, negotiation. Click an intel chip to highlight the matching signal-routing row.

[ 01 · candidate.onboard ]flow.screening
Agent Onboarding
Spin up your Candidate Agent in minutes. Drop your CV, record a 5-minute intro, set guardrails, link LinkedIn — your agent now represents you 24/7 across every channel.
intel.feeds →
signals.json — candidate.onboard● live
agent.id
ca_91f3…live
guardrails
comp, geo, ethics
li.linked
verified
consent.scope
candidate-first
↓ hands off to step.02
[ 02 · screening.respond ]flow.screening
Screening Response
When a Merchant Agent surfaces a fit, your agent responds with structured claims — not a resume blast. It cites samples, references, and prior wins against the JD.
intel.feeds →
signals.json — screening.respond● live
jd.match
0.91
claims.cited
14
psychometric
voice.session
response.latency
4m 12s
↓ hands off to step.03
[ 03 · truth.proof ]flow.fraud
Fraud + Truth Proof
Every claim you make is challengeable — and provable. Your agent submits liveness, voice-print, document and reference checks before any offer round.
intel.feeds →
signals.json — truth.proof● live
liveness
pass
voiceprint
match 0.98
refs.verified
3 / 3
ledger.entry
tl_77c…
↓ hands off to step.04
[ 04 · negotiation ]flow.negotiation
Negotiation
Your agent negotiates inside your guardrails — comp floor, equity, start date, growth path. You see every round. You countersign. No back-channel, no ghosting.
intel.feeds →
signals.json — negotiation● live
ask.delta
+18.0%
floor
$210k base
rounds
3
countersign
you

[ Signal Routing ]

Which intelligence output feeds which flow.

Each row is a node from the architecture diagram above. Each column is a flow surfaced in the signals.json panels. Filled cells indicate active routing.

→ click any intel.module row or an Intelligence Stack node in the diagram to highlight matching flows

intel.module	Screening	Fraud Checks	Negotiation	arch.node
Psychometric	routed	—	—	module.01
Vitals + I/O	routed	routed	—	module.02
JD Fitment	routed	—	routed	module.03
LinkedIn Validate	routed	routed	—	module.04
Fraud / Deepfake	—	routed	—	module.05
Truth Ledger	routed	routed	routed	module.06

Read this matrix alongside the architecture diagram: every ● node in the Intelligence Stack streams into one or more flows. Truth Ledger is the only module wired into all three — it is the audit substrate.

[ Use-case Flows ]

Concrete flows. Real signals.

Four production flows that run inside the loop — each emits structured signals you can audit, replay and challenge.

[ Flow.01 ]trace.live

Screening

JD ingestion → role calibration → candidate signal pull → ranked shortlist with reasoning trail.

01Merchant Agent parses JD + hiring manager calibration call
02Candidate Agents respond with structured claims + evidence
03Signal market scores fitment across 12 dimensions
04Humans review top 7 with reasoning trail attached

signals.json● live

jd.fitment: 0.87
role.calibration: locked
shortlist.size: 7 / 412
trace.id: scr_8f2a…

[ Flow.02 ]trace.live

Psychometric

Conversational assessment over voice or video — no timed multiple-choice slop.

01Candidate opts in to a 12-minute voice session
02Agent probes adaptability, ownership, conflict style
03Vitals + linguistic cadence cross-checked against baseline
04Report shared with candidate first, employer second

signals.json● live

ownership.index: 8.4 / 10
adaptability: high
stress.coherence: stable
consent.scope: employer-view

[ Flow.03 ]trace.live

Fraud + Truth

LinkedIn cross-validation, deepfake checks, claim verification — every assertion is provable.

01Resume + profile claims extracted as discrete assertions
02Each assertion routed to a verifier (refs, docs, public record)
03Liveness + voice-print check on every video round
04Truth-ledger published to candidate and employer

signals.json● live

claims.verified: 18 / 19
liveness: pass
li.match: 0.96
anomaly.flags: 1 minor

[ Flow.04 ]trace.live

Offer + Negotiation

Agents negotiate inside your guardrails — comp, equity, start date, growth path.

01Candidate sets guardrails: floor, must-haves, deal-breakers
02Merchant Agent shares band, flex, equity logic
03Agents iterate offer structure across 2–5 rounds
04Humans countersign — written trail, no he-said-she-said

signals.json● live

offer.delta: +12.4%
equity.refresh: yes
start.date: T+21d
rounds: 3

[ Topology ]

The loop, drawn out.

[ 01 ]Candidate→ negotiates

[ 02 ]Candidate Agent→ negotiates

[ 03 ]Signal Market→ negotiates

[ 04 ]Merchant Agent→ negotiates

[ 05 ]Employerhandshake.

See the intelligence stack →For employers

Two agents. A market between them. Humans in command.

Humans + agents, on the same surface.

Two agents. Different training surfaces.

Candidate Agent

Merchant HR Agent

Both agents are trained on verified science.

Big-Five at the speech-act layer

Claim verification with statement-validity protocols

Non-contact physiology from camera and mic

Predictive validity over interview theater

A panel of specialised agents per role.

Neither agent grades the match.

Candidate Agent

Merchant Agent

Matchmaking

Career Building

Two agent fleets. One intelligence core.

A2A · MCP · UCP. The wire between agents.

Candidate Agent ⇄ Merchant Agent

Tools the agents can call.

Match on hex cells, not city strings.

Bengaluru / HSR

From install to countersign. In four nodes.

Channel Setup

Job Intake

Evaluation

Offer + Negotiation

Your agent, your guardrails. Four nodes.

Agent Onboarding

Screening Response

Fraud + Truth Proof

Negotiation

Which intelligence output feeds which flow.

Concrete flows. Real signals.

Screening

Psychometric

Fraud + Truth

Offer + Negotiation

The loop, drawn out.