KEPNER–
TREGOE.

CONCERN PRIORITY

Concern: retrieval latency increased

Seriousness:
  medium — p95 misses internal target,
  but no user outage yet

Urgency:
  high — release rollout reaches 100% tomorrow

Growth:
  high — queue depth rising 18% / hour

Action:
  PA immediately
  owner: Search Platform

PROBLEM:
  Model response p95 rose from 700ms to 2.4s.

WHAT
  IS:
    model profile "standard-v4"
  IS NOT:
    "fast-v2" on same gateway

WHERE
  IS:
    EU deployment
  IS NOT:
    US deployment

WHEN
  IS:
    requests after 14:20 deployment
  IS NOT:
    requests before 14:20

EXTENT
  IS:
    long-context requests > 40k tokens
  IS NOT:
    requests < 10k tokens

DISTINCTIONS:
  standard-v4 + EU + new deployment + long context

CHANGES:
  new serving config at 14:20
  KV-cache policy changed only on EU standard pool

POSSIBLE CAUSE:
  new KV-cache configuration causes additional
  long-context prefill/eviction overhead.

TEST:
  explains WHAT? yes
  WHERE? yes
  WHEN? yes
  EXTENT? yes

NEXT:
  inspect metrics / reproduce / rollback one pool
  to verify causal hypothesis.

DECISION:
  primary vector platform

MUST:
  EU deployment available
  tenant filtering supported
  deletion supported
  max annual infra budget ≤ X
  benchmark correctness threshold ≥ Y

WANTS:
  operational simplicity           weight 10
  low p95 latency                  weight 9
  low total cost                   weight 8
  good hybrid filtering            weight 8
  mature backup/recovery           weight 7
  easy observability               weight 6
  low lock-in                      weight 5

PLAN:
  route 30% traffic from Model A to Model B.

CRITICAL AREA:
  structured tool-calling.

POTENTIAL PROBLEM:
  Model B emits more invalid tool arguments.

LIKELY CAUSE:
  schema adherence differs on long conversations.

PREVENTIVE:
  run tool-contract eval set;
  validate args deterministically;
  start at 5% canary;
  exclude high-risk tool intents initially.

TRIGGER:
  invalid_tool_args > 0.7%
  OR high-risk schema failure > 0.

CONTINGENT:
  disable Model B route;
  return affected cohort to A;
  preserve traces for failure analysis.

OWNER:
  Model Platform on-call.

SECOND POTENTIAL PROBLEM:
  Model A fallback capacity insufficient.

PREVENTIVE:
  reserve capacity / verify quota.

CONTINGENT:
  degrade optional DEEP work,
  rate-limit noncritical traffic.

DEVIATION:
  grounded-answer pass rate fell 91% → 78%.

IS / IS NOT

WHAT IS:
  product manuals corpus
WHAT IS NOT:
  policy documents corpus

WHERE IS:
  documents ingested after parser v3
WHERE IS NOT:
  old indexed documents

WHEN IS:
  after Aug 28 reindex
WHEN IS NOT:
  before reindex

EXTENT IS:
  tables-heavy PDFs
EXTENT IS NOT:
  text-heavy PDFs

DISTINCTIONS:
  parser v3 + new reindex + tables

CHANGES:
  table extraction logic changed
  chunk boundaries changed

CANDIDATE CAUSES:
  A) embedding model degraded
  B) parser table extraction loses row associations
  C) query router changed

TEST AGAINST FACTS:
  A fails WHEN/WHERE distinction:
    same embedding used old documents successfully.
  C fails WHAT:
    same routing across corpora.
  B explains all known IS/IS NOT facts.

VERIFY:
  compare Document IR;
  reconstruct affected table;
  reparse with v2;
  run eval.

LESSON:
  don't optimize vector search before diagnosing
  whether source representation is broken.

{
  "decision_statement": "...",
  "musts": [
    {
      "criterion":"EU residency",
      "threshold":"required"
    }
  ],
  "wants": [
    {
      "criterion":"operational simplicity",
      "weight":10
    }
  ],
  "alternatives": [
    {
      "name":"A",
      "ratings":[
        {
          "criterion":"operational simplicity",
          "rating":8,
          "evidence_ref":"..."
        }
      ]
    }
  ],
  "sensitivity_checks": [...],
  "adverse_consequences": [...],
  "selected":"...",
  "assumptions":[...],
  "review_triggers":[...]
}

KEPNER–TREGOE WORKSHEET
===========================================================

A. SITUATION APPRAISAL

Concerns:
1.
2.
3.

For each:
Seriousness:
Urgency:
Growth:
Owner:
Next process:
[ ] Problem Analysis
[ ] Decision Analysis
[ ] Potential Problem Analysis


B. PROBLEM ANALYSIS

Problem statement:
OBJECT:
DEVIATION:

WHAT
IS:
IS NOT:

WHERE
IS:
IS NOT:

WHEN
IS:
IS NOT:

EXTENT
IS:
IS NOT:

DISTINCTIONS:
1.
2.
3.

CHANGES:
1.
2.
3.

POSSIBLE CAUSES:
A.
B.
C.

TEST EACH CAUSE AGAINST:
WHAT:
WHERE:
WHEN:
EXTENT:

MOST PROBABLE:
→

VERIFICATION TEST:
→


C. DECISION ANALYSIS

Decision statement:
→

MUST OBJECTIVES:
1.
2.
3.

WANT OBJECTIVES:
criterion / weight
1.
2.
3.

ALTERNATIVES:
A.
B.
C.

SCORES:
→

SENSITIVITY:
→

ADVERSE CONSEQUENCES:
A →
B →
C →

SELECTED:
→

RATIONALE:
→

REVIEW TRIGGER:
→


D. POTENTIAL PROBLEM ANALYSIS

Plan / step:
→

Potential problem:
→

Likely cause:
→

Preventive action:
→

Trigger / early warning:
→

Contingent action:
→

Owner:
→

You are assisting with Kepner–Tregoe Problem Analysis.

Rules:
1. Do not propose a root cause immediately.
2. Restate the problem as OBJECT + DEVIATION.
3. Separate facts, inferences and unknowns.
4. Build IS / IS NOT across:
   WHAT, WHERE, WHEN, EXTENT.
5. IS NOT must be a meaningful comparable case.
6. Never invent a missing fact.
7. Identify distinctions between IS and IS NOT.
8. Identify known changes associated with distinctions.
9. Generate candidate causes from distinctions and changes.
10. For every candidate, test whether it explains
    every important IS / IS NOT fact.
11. Explicitly list facts the cause fails to explain.
12. Rank hypotheses only provisionally.
13. Define the smallest safe observation/experiment
    that could verify or falsify the top hypothesis.
14. Do not call a hypothesis a root cause
    until verification evidence exists.

You are assisting with Kepner–Tregoe Decision Analysis.

1. Write a precise decision statement.
2. Ask for the decision horizon and scope.
3. Separate objectives into:
   MUST — non-negotiable pass/fail thresholds;
   WANT — desirable weighted objectives.
4. Challenge every MUST:
   is it genuinely mandatory?
5. For every WANT:
   define measurable meaning where possible.
6. List alternatives without scoring yet.
7. Screen alternatives against MUSTs.
8. Score surviving alternatives against WANTs.
9. Cite evidence for each important rating.
10. Mark unsupported ratings as assumptions/unknowns.
11. Run sensitivity on weights and uncertain ratings.
12. Identify adverse consequences for top alternatives.
13. Recommend only after objectives, evidence and risks are visible.
14. Record assumptions and review triggers.

START
  ↓
WHAT KIND OF SITUATION IS THIS?

────────────────────────────────────────────

MANY CONCERNS / CHAOS
  ↓
SITUATION APPRAISAL

list concerns
  ↓
separate mixed issues
  ↓
clarify
  ↓
priority:
  seriousness
  urgency
  growth
  ↓
owner + next process

────────────────────────────────────────────

UNEXPECTED DEVIATION
CAUSE UNKNOWN
  ↓
PROBLEM ANALYSIS

EXPECTED
vs
ACTUAL
  ↓
OBJECT + DEVIATION
  ↓
IS / IS NOT

WHAT
WHERE
WHEN
EXTENT
  ↓
DISTINCTIONS
  ↓
CHANGES
  ↓
POSSIBLE CAUSES
  ↓
TEST EACH CAUSE
AGAINST ALL FACTS
  ↓
MOST PROBABLE HYPOTHESIS
  ↓
VERIFY WITH:
  telemetry
  experiment
  reproduction
  rollback
  direct observation
  ↓
CAUSE VERIFIED

────────────────────────────────────────────

CAUSE KNOWN / NEED TO CHOOSE ACTION
  ↓
DECISION ANALYSIS

decision statement
  ↓
MUST objectives
  pass / fail
  ↓
WANT objectives
  weighted
  ↓
alternatives
  ↓
MUST screening
  ↓
evidence-based scoring
  ↓
sensitivity analysis
  ↓
adverse consequences
  ↓
select alternative
  ↓
record assumptions
and review trigger

────────────────────────────────────────────

PLAN CHOSEN
  ↓
POTENTIAL PROBLEM ANALYSIS

critical step
  ↓
what could go wrong?
  ↓
likely cause
  ↓
PREVENTIVE ACTION
  reduce probability
  ↓
EARLY WARNING / TRIGGER
  ↓
CONTINGENT ACTION
  reduce impact
  ↓
owner

────────────────────────────────────────────

WITH TRIZ:

KT PA
  finds the real cause/conflict
        ↓
TRIZ
  invents stronger mechanism
        ↓
KT DA
  chooses between concepts
        ↓
KT PPA
  de-risks rollout
        ↓
EVAL / PROTOTYPE / PRODUCTION
  verifies real outcome

────────────────────────────────────────────

IN AI SYSTEMS:

OBSERVABILITY
  provides facts

KT
  organizes diagnosis

TRIZ
  helps invent non-obvious redesign

EVALS
  measure whether change helped

PRODUCTION ARCHITECTURE
  implements safely

────────────────────────────────────────────

THE MOST IMPORTANT KT QUESTION
IS OFTEN NOT:

“WHAT IS THE ROOT CAUSE?”

IT IS:

“WHAT EXACTLY IS DIFFERENT
BETWEEN WHERE THE PROBLEM
IS
AND WHERE IT LOGICALLY
COULD HAVE BEEN
BUT IS NOT?”

THAT CONTRAST
TURNS A WALL OF POSSIBLE CAUSES
INTO A MUCH SMALLER
TESTABLE SET.