Problem → evidence → solution

Engineering Solutions

Start with the problem you are seeing. Find the evidence to collect, the chapter that explains the mechanism, and the AI diagnostic that applies it to your own repository.

Programmer.ie solutions are organized around engineering failure modes rather than publication date.

A solution is considered covered when the site has a concrete diagnostic or application workflow tied back to the underlying learning material. Missing entries remain visible so coverage gaps drive what gets built next.

42Known problems
25Covered
17Coverage gaps
4Engineering domains
6 of 10 covered

PyTorch

Framework, tensor, training and GPU failures

Covered Transformers

Attention Q, K, V or Masks Have the Wrong Shape

  • mask mismatch
  • head dimension error
  • wrong attention broadcasting
Run diagnostic → Learn why →
Coverage gap Training

Training Improves but Validation Gets Worse

  • validation divergence
  • memorization
  • unstable generalization

Tracked explicitly so the next content wave can close this gap.

Coverage gap Evaluation

Evaluation Looks Better Than Real Performance

  • data leakage
  • train evaluation mismatch
  • suspiciously high metrics

Tracked explicitly so the next content wave can close this gap.

Coverage gap Tensors Shapes

Batch and Sequence Dimensions Are Being Confused

  • silent transpose error
  • wrong batching semantics
  • sequence axis confusion

Tracked explicitly so the next content wave can close this gap.

Coverage gap Training

Training Does Not Resume Correctly From a Checkpoint

  • optimizer state lost
  • scheduler reset
  • reproducibility drift

Tracked explicitly so the next content wave can close this gap.

4 documented coverage gaps remain in this domain.

6 of 10 covered

Models

Objectives, representations and architecture decisions

Covered Objectives

A Learned Score Does Not Mean What the System Thinks It Means

  • miscalibrated score
  • target leakage
  • downstream misuse
Run diagnostic → Learn why →
Covered Objectives

Multiple Prediction Heads Are Fighting Each Other

  • gradient conflict
  • incompatible losses
  • unstable multi-task learning
Run diagnostic → Learn why →
Covered Architecture

Separate Learned Components Do Not Justify Their Boundaries

  • duplicated representation
  • unclear ownership
  • unnecessary learned modules
Run diagnostic → Learn why →
Covered Architecture

Fast and Slow Model State Do Not Have Distinct Roles

  • state collapse
  • redundant recurrence
  • unclear timescales
Run diagnostic → Learn why →
Covered Architecture

Recursive Compute Adds Cost Without Better Decisions

  • repeated latent updates
  • no equal-budget gain
  • unstable convergence
Run diagnostic → Learn why →
Covered Model Selection

The Repository May Be Using a More Complex Model Than It Needs

  • architecture inflation
  • missing baseline
  • complexity without evidence
Run diagnostic → Learn why →
Coverage gap Objectives

The Training Objective Is Only a Proxy for the Real Decision

  • metric gaming
  • offline online mismatch
  • proxy optimization

Tracked explicitly so the next content wave can close this gap.

Coverage gap Evaluation

Model Confidence Is Not Calibrated

  • overconfidence
  • unreliable thresholds
  • probability mismatch

Tracked explicitly so the next content wave can close this gap.

Coverage gap Evaluation

Architecture Components Have Never Been Ablated

  • unknown component value
  • no baseline
  • unjustified complexity

Tracked explicitly so the next content wave can close this gap.

Coverage gap Representations

Learned Representations Are Collapsing

  • low embedding variance
  • identical outputs
  • weak separation

Tracked explicitly so the next content wave can close this gap.

4 documented coverage gaps remain in this domain.

6 of 10 covered

Agents

Planning, tools, memory, search and verification

Covered Verification

The Agent Declares Success Without Independent Evidence

  • self-verification
  • claimed side effects
  • false completion
Run diagnostic → Learn why →
Coverage gap State

The Agent Treats Its Own Context as Authoritative State

  • stale context
  • invented state transitions
  • state divergence

Tracked explicitly so the next content wave can close this gap.

Coverage gap Execution

Retrying an Agent Action Can Duplicate Side Effects

  • duplicate writes
  • duplicate sends
  • repeated mutations

Tracked explicitly so the next content wave can close this gap.

Coverage gap Control Flow

Agent Budgets Exist Only in the Prompt

  • unbounded tokens
  • unbounded tools
  • unenforced cost limits

Tracked explicitly so the next content wave can close this gap.

Coverage gap Tools

Tool Results Are Not Converted Into Reliable Observations

  • raw tool ambiguity
  • parser drift
  • missing error semantics

Tracked explicitly so the next content wave can close this gap.

4 documented coverage gaps remain in this domain.

7 of 12 covered

Advanced Agents

Reliability, authority, durability and production control

Covered Execution

Speculative Work Can Escape Before It Is Accepted

  • uncommitted side effects
  • branch leakage
  • speculative mutation
Run diagnostic → Learn why →
Covered Reliability

One Failing Dependency Can Take Down the Agent Platform

  • retry storm
  • cascading timeout
  • worker exhaustion
Run diagnostic → Learn why →
Covered Provenance

A Historical Agent Run Cannot Be Reconstructed

  • missing versions
  • incomplete observations
  • unreplayable side effects
Run diagnostic → Learn why →
Covered Reliability

An Agent Incident Has Symptoms but No Defensible Root Cause

  • unclear divergence
  • weak timeline
  • remediation not proven
Run diagnostic → Learn why →
Covered Authority

The Agent Has More Authority Than Its Evidence Justifies

  • unsafe autonomy
  • rubber-stamp approval
  • missing escalation
Run diagnostic → Learn why →
Covered Durability

Plans, Commitments and Executable Work Are Conflated

  • stale commitments
  • cancellation errors
  • unclear obligation state
Run diagnostic → Learn why →
Covered Control Plane

The Model Controls the System That Is Supposed to Control the Model

  • policy in prompts
  • unenforced authority
  • model-owned budgets
Run diagnostic → Learn why →
Coverage gap Distributed Execution

Distributed Agent Tasks Are Not Safe to Retry

  • duplicate task effects
  • retry races
  • inconsistent completion

Tracked explicitly so the next content wave can close this gap.

Coverage gap Reliability

Agent Work Arrives Faster Than the Platform Can Process It

  • queue growth
  • latency collapse
  • worker saturation

Tracked explicitly so the next content wave can close this gap.

Coverage gap Behavioral Engineering

Agent Behavior Changes Without an Explicit Release

  • prompt drift
  • model drift
  • policy drift

Tracked explicitly so the next content wave can close this gap.

Coverage gap Continuity

Long-Running Work Continues After the User Intent Has Changed

  • obsolete task execution
  • stale goal
  • superseded workflow

Tracked explicitly so the next content wave can close this gap.

Coverage gap Durability

A Long-Running Agent Cannot Resume Safely After Failure

  • lost progress
  • duplicate work
  • unrecoverable workflow state

Tracked explicitly so the next content wave can close this gap.

5 documented coverage gaps remain in this domain.