Score theatre
One number that sounds confident and explains nothing.
94%“match” · source unknown
- Title similarity mistaken for capability
- Missing data silently averaged into confidence
- No record of who decided, or why
Hiring decisions you can defend.
Timur is Human Talent Decision Intelligence. It tells recruiters who to review next — on what evidence, under what authority, and with which unknowns still open.
The intended journey is plain: bring a role brief and authorised CVs, inspect the evidence, review a shortlist and record the decision.
Illustration below uses synthetic records — illustrated behaviour, not a live product claim.JLreview
Jordan Lim
talent:01842 · 7 claims · 2 sources
Decision vector — no fused score
Capability evidencestrong
Enterprise AI launch · CV v4 p.2 · recruiter confirmed
Employer feasibilitypass
Hard gates executed before ranking
Candidate intentunknown
No current signal — stays visible, never imputed
Representation integrityopen
1 date contradiction between sources
Profiles find people. They don’t justify decisions. Timur turns evidence into a decision a recruiter can explain — what is known, what is contradicted, what is still unknown, and who is accountable.
The problem
A single score hides everything that matters.
Directories and job boards optimise for identity, attention and keyword relevance. Recruiters need evidence, feasibility, intent and consent — and a record of why a shortlist looks the way it does.
The Timur decision vector
Separate dimensions, each backed by claims you can open.
Capability evidencestrong
Employer feasibilitypass
Candidate intentunknown
Placement sustainabilitymoderate
- Hard constraints execute before ranking
- Unknowns and contradictions stay visible
- Every recommendation leaves an audit trail
The decision kernel
Five contracts. One accountable loop.
Sources are replaceable. The canonical evidence model and decision logic are not. Timur speaks a small, versioned language that every part of the product shares.
What the employer can actually hire
A recruiter-confirmed role requirement: role family, constraints, must-haves and location — not a pasted job ad.
role_family: "product-management" location: "SG" must_have: ["b2b-saas-launch"] confirmed_by: "recruiter" # required
Claims, not flattering profiles
Versioned assertions with resolvable provenance. Contradictions and staleness weaken an assessment instead of disappearing.
claim: "led enterprise AI launch" source: "cv:v4#p2" observed: 2026-08-14 verification: "recruiter_confirmed" contradictions: 1 # stays visible
Permission is a hard gate
Who may process or act, for what purpose, until when.
Process for assessment
purpose: placementgranted
purpose: placementgranted
Contact candidate
purpose: outreachblocked
purpose: outreachblocked
The record that survives review
Recommendation, uncertainty, policy versions and the human outcome.
next_actiongatesassessmentspolicy_versionschema_versionhuman_decision
Durable facts for audit and replay
Every state change is an event you can trace.
- twin.confirmed09:12
- claims.emitted09:14
- gates.evaluated09:14
- decision.recordedpending
How it works
From role brief to defensible shortlist.
Seven stages, one decision procedure. A processing grant is required at intake; action grants are checked again before contact or sharing. Models may help extract and explain; they may not waive a mandatory requirement or turn uncertainty into confidence.
A recruiter confirms the Opportunity Twin — the role the employer can actually hire for — under a purpose-scoped processing grant. No grant, no intake.
CVs, attested history and professional context become versioned claims with source, time and verification state.
Records across authorised sources resolve to one person, with a human review queue for uncertain merges.
Candidates are retrieved against the Twin. Retrieval surfaces who to consider; it never decides.
Deterministic hard constraints run before any ranking. A missing certification cannot be quietly waived.
Action-specific grants are checked again before share or contact. A processing grant at intake is not a licence to outreach.
A Decision Envelope records what was recommended, what stayed unknown and what the human decided.
Stage 01 — Intakedeterministic
intake→evidence→identity→retrieve→gates→authority→decision
- Role
- Senior Product Manager
- Location
- Singapore
- Must-have
- B2B SaaS launch
- Processing grant
- placement · granted
Stage 02 — Evidencedeterministic
intake→evidence→identity→retrieve→gates→authority→decision
- claims emitted
- 7
- sources
- cv:v4 · agency_history
- verified
- 4 / 7
- contradictions
- 1
Stage 03 — Identitydeterministic
intake→evidence→identity→retrieve→gates→authority→decision
- Candidate records
- 2 → 1 person
- Merge confidence
- High
- Uncertain merges
- 1 in review queue
Stage 04 — Retrievedeterministic
intake→evidence→identity→retrieve→gates→authority→decision
- Considered
- Synthetic cohort
- Surfaced for gates
- Top candidates
- Decides?
- never
Stage 05 — Gatesdeterministic
intake→evidence→identity→retrieve→gates→authority→decision
- Location eligibility
- pass
- Work-pass feasibility
- pass
- Mandatory certification
- unknown → validate
Stage 06 — Authoritydeterministic
intake→evidence→identity→retrieve→gates→authority→decision
- Assess for placement
- still granted
- Share with employer
- expires in 12 days
- Contact candidate
- blocked · no grant
Stage 07 — Decisiondeterministic
intake→evidence→identity→retrieve→gates→authority→decision
- Recommendation
- Review next
- Open unknowns
- Intent, certification
- Next action
- Validate with candidate
- Human decision
- recorded
Canonical evidence
Store claims, not polished assertions.
A profile is readable and searchable — and insufficient for a responsible recommendation. A hiring decision needs source, time, verification, legal basis and contradiction history.
JL
Jordan Lim
Senior Product Manager · Singapore
- Experience
- 9 years
- Availability
- Not stated
- Compensation
- Not stated
- Source & date
- —
Keywords and titles. No provenance, no verification, no basis for use.
- subject
- talent:01842
- claim
- led enterprise AI launch
- source
- cv:v4 · page 2
- valid_period
- 2023 — 2025
- verification
- recruiter_confirmed
- legal_basis
- candidate_consent
- contradictions
- 1 date discrepancy
Eligible for capability assessment. The date contradiction remains visible to the recruiter.
Design principles
Built on rules, not vibes.
These aren’t aspirations on a slide. They are written into the canonical schemas and the release gates. Illustrated behaviour on this page is not the same as implemented or verified capability — every milestone is measured against that distinction.
Evidence over score fusion
Inspectable claims and gates are the product truth — not a single opaque match number.
Provenance is mandatory
Every assertion carries its source, time and verification state. Unknowns stay visible.
Tenant isolation
Data and jobs are scoped by tenant with row-level security. Cross-tenant access is a defect.
Purpose and action authority
Processing and outreach require explicit grants. Missing authority is designed to block the action.
Human accountability
Timur assists. Rejection and outreach remain decisions a person makes and owns.
Synthetic before personal
Deterministic behaviour on synthetic records until controlled-data acceptance is earned.
Accuracy & trust
Measure decisions, not eloquence.
A model can sound certain while being wrong. Timur’s release gates are defined up front — labelled cases, provenance, consistency and real outcomes. These are targets the pilot must meet, not results we claim today.
Provenance coverage
Every claim used in a recommendation points to source, time and extraction version.
release gateUnauthorised outreach
No platform-triggered contact without the required permission state. Designed to enforce — a release gate, not yet demonstrated.
release gateIdentity merge precision
A duplicate is annoying. Merging two different humans is dangerous.
initial targetPurpose and legal basis
Why each data item may be collected, used, disclosed and retained travels with it. Public availability is not a licence for unlimited reuse.
Candidate rights and correction
Access, correction, withdrawal and representation status — with a complete audit trail. Withdrawn permissions don’t come back on restore.
Model accountability
Models, schemas and decision policies are versioned. The evidence behind each recommendation and every human override is preserved.
Security boundary
Tenant isolation with row-level security, least privilege, encryption, private data paths and separate POC and production identities.
Employer identity
Opportunity Twins attach to Singapore-credible employer records such as UEN, so the role — and the hiring entity — are unambiguous.
Feasibility as versioned rules
Work-pass and COMPASS-style checks inform the recruiter as policy-versioned rules. They never replace MOM’s official assessment.
Roadmap
The path to a provable pilot.
Six engineering milestones turn the foundation into a demonstrated decision loop — for one Singapore role family, on synthetic data, measured against adjudicated truth. No quality claims before the evaluation milestone.
Red lines
What Timur will never do.
Some shortcuts make demos prettier and hiring worse. Permanent red lines stay in the product. Current pilot restrictions are labelled separately — “never” and “until accepted” are different commitments.
Fused match scores
No single percentage pretending to be certainty.
Autonomous rejection
People are never screened out by the system alone.
Unauthorised outreach
No contact without a purpose-scoped grant.
Platform scraping
Only authorised, attributable sources become evidence.
Personal data to external models
Not a permanent red line. Talent V0.1 is deterministic. Sending personal data to an external model needs a named acceptance — it is off until then.
FAQ
Straight answers.
Still curious? Ask us directly — we’d rather show you the loop than describe it.
Is Timur an AI recruiter?
No. Timur is decision intelligence for recruiters. It organises evidence and records decisions. Gates and authority are designed to enforce; they are not yet demonstrated (M3). It does not reject candidates on its own or contact anyone without an explicit grant to act.
Why no match percentage?
A single score hides what is known, contradicted or missing. Timur shows a decision vector — capability, feasibility, intent, representation — each backed by claims you can open, with unknowns kept visible instead of averaged away.
Who is Timur built for first?
Recruitment agencies operating in Singapore. The first release focuses on one confirmed role family, authorised first-party agency history and two or three recruiters in a shadow workflow.
Where does candidate data come from?
Authorised sources only: CVs, the agency’s own attested history and permitted professional context. Every item enters as a claim with provenance. Enrichment vendors are inputs to claims, never decision authority.
Is Timur live with real candidate data today?
Not yet. Talent V0.1 runs on synthetic records with deterministic processing. Live personal data requires a separate, named controlled-data acceptance after the pilot trust controls are in place.
How does Timur handle Singapore requirements?
Purpose, legal basis, retention and policy versions travel with the data. Work-pass and COMPASS-style feasibility are modelled as versioned rules that inform the recruiter — they do not replace official MOM assessment.
For Singapore recruitment agencies
Bring one role family. We’ll show you the decision loop.
A walkthrough on synthetic records: Opportunity Twin, evidence claims, hard gates, blocked actions and the Decision Envelope — end to end.