GLEE / AI RESEARCH LAB
We investigate dependable AI systems and how intelligence, engineering and Canada's resources could expand productive capability. Read the methods, results and questions still open.
CANADIAN CENTURY / NATIONAL PRODUCTIVE TRANSFORMATION
Bring resources, human expertise, AI, robotics and industry together at national scale. Determine what is feasible, what it costs and how benefits reach Canadians.
First note · 15 October 2026: Canadian industry baselines and an executable manufacturing model. Achievable national gains remain unestimated.
Read the first model and explore its assumptions ↗NATIONAL RESEARCH UPDATE / 16 OCTOBER 2026
How large could Canada's productive transformation be? Read the magnitude requirements, engineering limits, competing strategies and unfinished household account.
NEW NOTE / 16 OCTOBER 2026
Electricity for industrial expansion: reported generation, construction costs and continuous-load requirements.
REPRODUCIBLE RESEARCH / 16 OCTOBER 2026
Construction capacity: measured panel assembly, historical costs and a runnable research package.
RESEARCH TOOLKIT / 16 OCTOBER 2026
From process improvements to national prosperity: a research prospectus and runnable production-data toolkit.
MINING RESEARCH / 16 OCTOBER 2026
Autonomous haulage in Canada: historical operating claims and an interactive example of process bottlenecks, costs and ordinary alternatives.
CANADIAN CENTURY / 17 OCTOBER 2026
A selected accounting comparison reconciled; manufacturing source claims examined; a complete production-and-cost comparison still missing. Achievable national and household gains remain unestimated.
Read the research update and primary sources ↗01 / Follow a real question
OBSERVED 12 OCTOBER 2026 · 8 TASKS · 32 RUNS
No accepted repairs in either arm. This trial could not tell us whether guidance improves repair quality.
We compared an agent reviewing its own work with an agent receiving evidence guidance. Most attempts stopped before a proposed repair could be tested.
The useful result is a problem with the experiment: we need to separate interface recovery from code-repair quality before drawing an efficacy conclusion.
Read the method and limits ↓Sixteen matched pairs across eight tasks in one repository. Each arm had 16 runs and four decisions per run. Self review produced zero candidate evaluations; evidence guidance produced one. Neither arm produced an accepted repair.
The runner intentionally exposed an older source hash while a current read was available. Twenty-three runs exhausted that recovery protocol; eight responses failed the exact response format. The one evaluated candidate failed six assertions. Thirty-one runs never reached candidate evaluation.
This was a negative end-to-end engineering pilot. Steering efficacy remains inconclusive. No general ability, economic advantage, or live adoption follows. A successor would need a separately frozen, interface-qualified design with positive and negative controls.
GLEE produced the original trial and a later AI-assisted audit. That audit reconciled the original 32 terminal runs, source and oracle hashes, and captured calls without restarting subjects. This public edition was prepared with AI assistance (Codex) and retains the source-report hash and observation date. Private trajectories and tests are not available here; independent reproduction and scientific validation are not established.
Inspect this study's public record ↗02 / Learn from an observed failure
These dated engineering cases show how an apparently sensible mechanism can answer the wrong question. Their observations apply to the named cases and windows.
ENGINEERING CASE · 2026-07-24
A health field answered a different question than the one callers were asking, and the label was accurate the whole time.
Change: Gate on capability rather than the label: reachable, swap-capable, models registered, queue not saturated. Fail toward attempting, since a wrong guess costs one failed request while the old behaviour cost every local job for hours.
Limit: Down remains correct as a description of residency. The fix is on our side of the contract, not theirs. Any system that answers is-it-warm when you asked can-it-serve has the same gap.
Read the original dated record ↗ENGINEERING CASE · 2026-07-24
A constraint a weak generator can satisfy without being correct will be satisfied without being correct.
Change: Add a check for figures appearing with no claim attached, and keep generated drafts human-reviewed rather than auto-published.
Limit: This catches recitation, not falsehood. A confident lie in clean prose still passes. Choose invariants by what their violations look like, not only by what they forbid.
Read the original dated record ↗03 / What remains unresolved
Follow-up questions raised by these records. They are directions for investigation, not claims that a new experiment is running.
Establish positive and negative controls for reading and applying a current source before comparing repair strategies.
Separate reported state, available capability, and a measured result under actual operating conditions.
Check what a figure supports, including the surrounding proposition and evidence, rather than its presence alone.
04 / Keep the evidence close
The study and cases above keep methods, observations, decisions, sources, and limits together. An engineering result is not automatically a scientific finding.
Archive / Earlier engineering records
The earlier board journal and proof-system paper remain dated records. They describe their own prototype state.