KATLAS / Cierge Foundation
CAR · Custody · Authority · Receipts
KATLAS node active

PromptCierge

Bounded prompts · LLM budget discipline · signed decision receipts.

AI may draft, compare or recommend. Authorised roles decide. KATLAS signs the receipt.

Custody

Custody controls the evidence.

Evidence stays at source. Only metadata and refs travel.

Authority

Authority governs the action.

Authorised roles decide. AI may draft, compare or recommend.

Receipts

Receipts prove what happened.

KATLAS signs the governed action — not the raw prompt corpus.

Step 1

Prompt intake

Read-only summary of the submitted request
Submitted prompt route
Summarise four supplier proposals and surface trade-offs to support a budget allocation route for the Civic Works Framework.
Requesting team
Strategic Procurement
Project context
Q3 supplier consolidation — Civic Works Framework
Intended decision support
Budget allocation route — advisory, not autonomous
Prompt purpose
Bounded summary + budget rationale draft
Risk level
Medium
Estimated budget impact
£420,000
Requested model route
GPT-class large model, single-shot, full proposal corpus
Submitted at
2026-02-14T09:12:00Z
Step 2

Evidence boundary

What the prompt route may see · what stays at source
Available to prompt route
Shared evidence · 5
metadata + refs only
  • Approved project summary
    ev:approved_project_summary
  • Role-approved budget context
    ev:role_approved_budget_context
  • Supplier scoring summary (rolled-up)
    ev:supplier_scoring_summary
  • Procurement policy reference
    ev:policy_reference
  • Prior approved prompt pattern ref
    ev:prior_approved_pattern
Withheld / internal only
Withheld evidence · 6
stays at source
  • Raw supplier files
    wh:raw_supplier_files
  • Personal data of bidders
    wh:personal_data
  • Confidential pricing detail
    wh:confidential_pricing_detail
  • Private board notes
    wh:private_board_notes
  • Unbounded prompt transcripts
    wh:unbounded_prompt_transcripts
  • Confidential vendor credentials and provider call logs
    wh:vendor_credentials_and_call_logs
Step 3

Model route and budget rationale

Reuse approved pattern + low-cost model + human approval
Recommended route
Reuse approved pattern + low-cost model + human approval before decision

Pattern PR-SUPPLIER-SUMMARY-v3 covers this workflow at ~12% of the cost of a fresh large-model call. Human approval is required before any budget allocation is recorded.

Alternatives considered
  • No LLM needed — deterministic scoring suffices
  • Reuse approved prompt patternpreferred
  • Small model sufficient (summarisation only)
  • Specialist model required (procurement-tuned)
  • Human review required before any model output
  • Refuse / escalate — risk threshold exceeded
Step 4

Authority gate

Approve · Refuse · Use existing pattern
Acting as
Prompt Governance Reviewer
Authority gate · KATLAS signs the chosen action
Step 5

KATLAS-signed receipt

Action outcome · signer · node timestamp · CAR block
Step 5 · Awaiting authority gate

A KATLAS-signed receipt will appear here once a governed action is taken.

LLM budget discipline

When can PromptCierge avoid another LLM call?

As decision routes consolidate, PromptCierge can recommend fewer LLM calls — replacing repeated prompting with approved patterns, deterministic checks and governed reuse.

Pattern already approved

Repeated request matches an approved prompt pattern — reuse, do not regenerate.

Deterministic rule is sufficient

If a rule, lookup or scoring formula answers it, an LLM call adds cost without value.

Prior approved route answers it

If a prior governed route already answered the workflow need, link to it.

Human approval first

If human approval is required before model output is useful, gate before calling.

Smaller model is enough

Low-risk summary work usually runs on a small model at a fraction of cost.

Refuse or escalate

High-risk requests should be refused or escalated — not pushed to a model.

Approved pattern library

Reusable, governed prompt patterns

PromptCierge learns from approved prompt patterns, outcomes and receipts — not from uncontrolled private prompts.

Procurement
Supplier proposal summary
PR-SUPPLIER-SUMMARY-v3

Bounded summary of supplier proposals into approved fields only. Confidential pricing and personal data remain at source.

Uses
142
Cost index
0.12×
Policy
Policy diff explainer
PR-POLICY-DIFF-v2

Plain-language diff between policy revisions; cites clauses only.

Uses
87
Cost index
0.08×
Risk
Risk triage rationale
PR-RISK-TRIAGE-v1

Rules-first triage; LLM only used when rules abstain.

Uses
63
Cost index
0.04×
Public Records
FOI response draft
PR-FOI-DRAFT-v4

Bounded FOI draft; withholds Schedule-2 exempt material.

Uses
211
Cost index
0.18×
Funding
Grant application scoring brief
PR-GRANT-SCORE-v2

Scores grant applications against published criteria only.

Uses
54
Cost index
0.09×
Operations
Governed meeting minutes
PR-MEETING-MIN-v6

Public-safe minutes from approved recording sources only.

Uses
318
Cost index
0.05×
PromptCierge · KATLAS / Cierge FoundationReceipts prove what happened.Evidence stays at source.