TRIZ /
ТРИЗ.

TRIZ PROBLEM CARD

System:
  ...

Useful function:
  ...

We want to improve:
  ...

But when we improve it:
  ...

It becomes worse because:
  ...

Hard constraints:
  ...

Why existing compromise is unsatisfactory:
  ...

Success would mean:
  both requirement A
  AND requirement B
  without unacceptable harm.

BAD:
  “Нам нужен отдельный AI critic,
   который проверяет каждый ответ.”

IFR:
  “Ошибки, которые реально важны,
   обнаруживаются до выдачи результата,
   без заметного увеличения latency/cost
   для обычных запросов.”

NOW SOLUTION SPACE OPENS:
  schema checks
  deterministic validators
  source checks
  selective verification
  risk routing
  precomputation
  cached checks
  stronger model only on uncertain claims

Qualitative TRIZ ideality:

                   useful effects
IDEALITY  ≈  ─────────────────────────
             costs + harmful effects

For AI architecture:

useful:
  verified task success
  coverage
  quality
  autonomy
  time saved

costs / harms:
  latency
  model spend
  human review
  failure risk
  security surface
  complexity
  maintenance
  privacy exposure

COMPONENT A
   │
   ├── useful function ──→ COMPONENT B
   │
   ├── insufficient function ──→ B
   │
   └── harmful function ──→ C

EXAMPLE:

Retriever
   ├── provides evidence → Context
   ├── may miss rare evidence → Quality
   └── adds latency → Request

Model
   ├── synthesizes answer → User
   ├── may hallucinate → User
   └── consumes tokens → Budget

CLASSICAL IDEA:

S1 ── F ──▶ S2

S1 / S2 = substances / interacting objects
F       = field / interaction that produces effect

If interaction is:
  missing
  insufficient
  harmful
  uncontrolled

TRIZ asks how to:
  add / change field
  add mediator
  modify substance
  transform interaction
  remove harmful path

PROBLEM:
  We want fewer incorrect factual outputs.

CURRENT MOVE:
  Run a strong critic/judge on every response.

TECHNICAL CONTRADICTION:
  deeper verification ↑
    → error detection ↑
    → latency/cost ↑

PHYSICAL CONTRADICTION:
  verification must be DEEP
  AND verification must be LIGHT.

SEPARATION BY CONDITION:
  deep only when:
    high-risk domain
    external side effect
    weak evidence
    source conflict
    low confidence
    high business impact

RESOURCES:
  schema validator
  source metadata
  deterministic calculations
  cached source checks
  existing risk class
  cheap verifier
  strong model only as escalation

PRINCIPLES:
  01 segmentation
  03 local quality
  10 preliminary action
  15 dynamics
  23 feedback
  35 parameter change

SOLUTION CONCEPT:
  verification ladder:

  L0 schema / deterministic
  L1 source/evidence checks
  L2 cheap semantic verifier
  L3 strong critic
  L4 human

  choose minimal level by risk.

RESULT:
  system is simultaneously
  “fast” for ordinary tasks
  and “deep” for dangerous tasks.

WANT:
  knowledge index always fresh.

HARM:
  continuous full re-index is expensive.

CONTRADICTION:
  sync must be frequent
  AND sync must be infrequent.

SEPARATION:
  by data class / change signal / time.

RESOURCES:
  provider updated_at
  webhook/change token
  ETag/hash
  source version
  usage popularity
  quiet batch window

CONCEPT:
  event-driven incremental sync
  + cheap metadata polling
  + periodic reconciliation
  + full rebuild only when parser/index version changes.

TRIZ PRINCIPLES:
  segmentation
  preliminary action
  periodic action
  feedback
  dynamics

OUTCOME:
  “always fresh” is approximated where
  freshness matters without constant full rebuild.

TECHNICAL CONTRADICTION:
  browser agent gets more authority
  → task coverage improves
  → security risk rises.

PHYSICAL CONTRADICTION:
  agent must have authority
  AND must not have authority.

SEPARATION:
  by action class
  by domain
  by time
  by exact approved operation

RESOURCES:
  typed actions
  allowlist
  principal identity
  approval token
  operation_id
  read-only API
  secret broker

CONCEPT:
  broad perception
  + narrow execution capability.

MODEL MAY:
  see many options
  reason about them

HOST MAY AUTHORIZE:
  only exact bounded action.

RESULT:
  autonomy of cognition
  without autonomy of authority.

CONTRADICTION:
  remember more
  → personalization ↑
  → privacy/staleness/context pollution ↑

IFR:
  system remembers exactly what is useful
  at the moment it is useful
  and irrelevant/stale/private data
  creates no burden.

SEPARATION:
  episode vs durable memory
  short-term vs long-term
  verified vs unverified
  user fact vs secret
  active vs expired

RESOURCES:
  source/provenance
  confidence
  TTL
  user corrections
  task context
  governance classification

CONCEPT:
  memory consolidation
  after task:
    extract candidate
    verify
    classify
    attach source
    TTL
    store only eligible fact

NOT:
  save every conversation forever.

USER PROBLEM
   ↓
1. RESTATE SYSTEM + USEFUL FUNCTION
   ↓
2. IDENTIFY:
   desired improvement
   worsening effect
   constraints
   current workaround
   ↓
3. PROPOSE 2–4 CONTRADICTION FORMULATIONS
   ↓
4. USER / EVIDENCE SELECTS MOST ACCURATE
   ↓
5. FORMULATE IFR
   ↓
6. RESOURCE INVENTORY
   ↓
7. CHOOSE TRIZ PATH:
   technical contradiction → principles
   physical contradiction → separation
   interaction problem → function/Su-Field
   hard problem → ARIZ-like deepening
   ↓
8. GENERATE 5–10 MECHANISM CONCEPTS
   not slogans
   ↓
9. CHECK:
   constraints
   secondary problems
   cost
   risk
   feasibility
   ↓
10. PROTOTYPE / SIMULATE / EVAL
   ↓
11. KEEP ONLY SOLUTIONS
    THAT ACTUALLY MOVE THE METRIC

TRIZ WORKSHEET
========================================================

1. SYSTEM
What system/process/component are we changing?
→

2. USEFUL FUNCTION
What must it actually achieve?
→

3. CURRENT PROBLEM
What is insufficient/harmful?
→

4. TECHNICAL CONTRADICTION
If we ____________________,
then ____________________ improves,
but ____________________ becomes worse.

If we do NOT ____________________,
then ____________________ remains good,
but ____________________ is insufficient.

5. PHYSICAL CONTRADICTION
The same element/property must be:
→ ____________________
AND
→ NOT ____________________

6. IFR
Ideally:
the useful function ____________________
happens
without ____________________
and without significant new ____________________.

7. RESOURCES
Information:
→
Time:
→
Space/topology:
→
Existing components:
→
Environment/supersystem:
→
Harmful effects that can become signals:
→

8. SEPARATION
Can contradiction be separated:
[ ] in time
[ ] in space
[ ] by condition
[ ] by system level

9. PRINCIPLES TO TRY
[ ] segmentation
[ ] taking out
[ ] local quality
[ ] universality
[ ] preliminary action
[ ] cushioning
[ ] inversion
[ ] dynamics
[ ] periodic action
[ ] blessing in disguise
[ ] feedback
[ ] intermediary
[ ] parameter change
[ ] composite approach

10. CONCEPTS
A →
B →
C →
D →

11. SECONDARY CONTRADICTION
Best concept creates new conflict:
→

12. TEST
What smallest prototype/eval can prove
that A improves without unacceptable B?
→

13. DECISION
Baseline:
→
Candidate:
→
Measured delta:
→
Keep / revise / reject:
→

You are assisting with TRIZ-style problem solving.

Do not generate solutions immediately.

First:
1. Restate the system and useful function.
2. Identify what we want to improve.
3. Identify what worsens as a direct consequence.
4. Propose 2–4 precise technical contradiction formulations.
5. If possible, reformulate the selected one as a physical contradiction.
6. Write an Ideal Final Result that does not assume a specific technology.
7. Inventory existing resources:
   information, time, space/topology, components,
   environment/supersystem, harmful effects.
8. Choose only relevant TRIZ directions:
   separation strategies,
   inventive principles,
   function analysis/trimming,
   ARIZ-style deepening if needed.
9. Generate multiple concrete mechanisms, not slogans.
10. For every concept state:
    - how it resolves the contradiction;
    - which resources it uses;
    - what secondary problem it creates;
    - feasibility assumptions;
    - smallest experiment/eval.
11. Do not treat a TRIZ principle as evidence.
12. Do not claim a concept works until verified.

PROBLEM
  ↓
DEFINE SYSTEM
  useful function
  constraints
  harmful / insufficient function
  ↓
TECHNICAL CONTRADICTION

IF action X:
  A improves
  B worsens

IF no X:
  B remains acceptable
  A is insufficient
  ↓
CAN WE FORMULATE PHYSICAL CONTRADICTION?

element/property must be:
  X
  AND
  NOT-X
  ↓
IDEAL FINAL RESULT

useful function happens
without previous harm
with minimal added burden
  ↓
RESOURCE INVENTORY

information
time
space/topology
existing components
environment
supersystem
harmful effects
unused capacity
  ↓
CHOOSE SEARCH PATH

simple contradiction
  → separation
  → inventive principles

interaction/function issue
  → function analysis
  → trimming
  → Su-Field analogy if useful

hard unresolved contradiction
  → ARIZ-style deeper analysis
  ↓
GENERATE MECHANISMS

not:
  “use segmentation”

but:
  “split verification by claim risk;
   deterministic validators handle
   cheap cases, strong verifier only
   sees uncertain material claims”
  ↓
CHECK SECONDARY PROBLEMS
  new cost?
  new latency?
  new security risk?
  new maintenance burden?
  ↓
SIMULATE / PROTOTYPE / EVAL
  ↓
DID A IMPROVE
WITHOUT UNACCEPTABLE B?

NO
  → reformulate contradiction
  → inspect resources again
  → generate new concept

YES
  → implement
  → measure in production
  → save as reusable case

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

TRIZ IN AI ARCHITECTURE:

quality vs latency
  → adaptive verification

autonomy vs safety
  → broad cognition / narrow authority

freshness vs ingestion cost
  → event-driven incremental sync + reconcile

personalization vs privacy
  → selective governed memory

model quality vs cost
  → conditional escalation

context richness vs token budget
  → relevance-based context builder

reliability vs complexity
  → simple synchronous path
     + durable mechanisms only where needed

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

THE MOST IMPORTANT MOVE:

DO NOT ASK FIRST:

“WHICH OF THE 40 PRINCIPLES
SHOULD WE APPLY?”

ASK:

“WHAT EXACTLY IS THE CONTRADICTION?”

THEN:

“WHY DO THE TWO REQUIREMENTS
HAVE TO CONFLICT
IN THE SAME PLACE,
AT THE SAME TIME,
UNDER THE SAME CONDITION,
AND AT THE SAME SYSTEM LEVEL?”

THAT QUESTION
IS WHERE TRIZ
STARTS TO BECOME USEFUL.