b378c796…a1c890
hamz.ai · The effort lattice · the tour Read the evidence page Learn how this works
Scroll
hamz.ai / lab / lattice · the guided tour

Does turning up reasoning effort replace skills?

One experiment answered that, and this page walks you through it the way it ran: scroll to watch the lattice get built, run, judged, and settled.

17
tasks frozen first
16
conditions
2,652
marks judged blind
Every number here is read from the published matrix · Hamza Malik · hamz.ai
Chapter 01 · The freeze

Seventeen tasks, locked before anything ran

Seventeen real tasks, each with a pass checklist, frozen and hash-stamped before a single run. Seventeen skills, one task each.

suite sha256 b378c79644280bb93fb8ac71d0cadcfe301fd15b226dfaec852b619a2aa1c890

The hash is the commitment: the tasks could not quietly change after the results started coming in.

Chapter 02 · Sixteen conditions

Three models, five efforts, and Haiku

Three Claude models at five effort levels each, plus Haiku with no effort dial, make the sixteen conditions of the grid.

One corner stays empty: Fable 5 at max never finished, it got too expensive. The slot is drawn open rather than papered over.

Chapter 03 · 32 arms

Every condition runs twice

Each condition ran the same seventeen tasks twice, once cold and once with the skill library installed, which makes thirty-two arms.

Solid · the loaded arm, skills installed
Hollow · the cold arm, no skills
Chapter 04 · Blind judging

Two judges, one adjudicator

Two blinded judges scored every checklist mark without knowing which arm they were reading, and a fixed adjudicator settled what they disagreed on.

4.6% of marks disagreed · 122 of 2,652
2.5% adjudicated · 135 of 5,304 blinded slots, two per mark
122 disputes settled by the adjudicator · none left unresolved
Chapter 05 · The verdict

The matrix crystallizes

Effort raises the floor, and the skills still add on top of it.

Lowest to highest effort, no skills: Fable +3.5 · Opus +3.9 · Sonnet +4.9 points, each past the pre-registered 3-point bar
Skills at each model's highest effort: still +2.4 to +9.8 points on top
Haiku, no dial: +26.5 points, the largest gap in the grid

Bar heights are the real with-skills rates on the 17 tasks every condition completed, read from matrix.json. The three lines above quote the pre-registered low-versus-top figures from MATRIX.md; Fable's pair is low versus high, its amended confirmatory endpoint, since max never finished. Small cells are directional, not inferential.

Chapter 06 · The lattice, finished

Turn it over yourself

Drag to rotate the finished lattice, or pick a cell and read its real numbers.

View
Model
Effort
Pick a model. An effort level is optional.
hamz.ai / lab / lattice · the guided tour · static version

Does turning up reasoning effort replace skills?

One experiment answered that. This is the same story the animated tour tells, as plain sections with identical facts. The full record lives on the evidence page.

17
tasks frozen first
16
conditions
2,652
marks judged blind

Every number here is read from the published matrix · Hamza Malik · hamz.ai

Chapter 01 · The freeze

Seventeen tasks, locked before anything ran

Seventeen real tasks, each with a pass checklist, frozen and hash-stamped before a single run. Seventeen skills, one task each. The full slate:

suite sha256 b378c79644280bb93fb8ac71d0cadcfe301fd15b226dfaec852b619a2aa1c890

The hash is the commitment: the tasks could not quietly change after the results started coming in.

Chapter 02 · Sixteen conditions

Three models, five efforts, and Haiku

Three Claude models at five effort levels each, plus Haiku with no effort dial, make the sixteen conditions of the grid. One corner stays empty: Fable 5 at max never finished, it got too expensive.

Modellowmediumhighxhigh (extra high)max
Haiku 4.5no effort dial · one condition, ran
Sonnet 5ranranranranran
Opus 4.8ranranranranran
Fable 5ranranranrannot run · too expensive to finish
Chapter 03 · 32 arms

Every condition runs twice

Each condition ran the same seventeen tasks twice, once cold and once with the skill library installed, which makes thirty-two arms.

The loaded runs get the full seventeen-skill library installed, from multi-model-adversarial-review to pre-merge-validation-gate.

cold
loaded
Chapter 04 · Blind judging

Two judges, one adjudicator

Two blinded judges scored every checklist mark without knowing which arm they were reading, and a fixed adjudicator settled what they disagreed on.

Chapter 05 · The verdict

The matrix crystallizes

Effort raises the floor, and the skills still add on top of it.

91.2
no dial
Haiku 4.5
96.1
low
99.0
medium
99.0
high
98.5
xhigh
99.5
max
Sonnet 5
95.6
low
97.6
medium
98.0
high
99.0
xhigh
100.0
max
Opus 4.8
99.0
low
100.0
medium
100.0
high
100.0
xhigh
 
max · not run
Fable 5

Each bar is one condition's with-skills rate; the gray base is that condition's cold rate and the green top is what the skills added. Rates use the 17 tasks every condition completed, read from matrix.json. The three lines below quote the pre-registered low-versus-top figures from MATRIX.md; Fable's pair is low versus high, its amended confirmatory endpoint, since max never finished. Small cells are directional, not inferential.

Chapter 06 · The lattice, finished

The cells, in full

Every cell's cold rate, loaded rate, and skill delta, on the 17 tasks every condition completed.

ModelEffortColdLoadedSkill deltaRuns
Haiku 4.5no dial64.7%91.2%+26.5 pointssingle run per arm
Sonnet 5low84.8%96.1%+11.3 pointsmean over 3 runs per arm
Sonnet 5medium90.7%99.0%+8.4 pointsmean over 3 runs per arm
Sonnet 5high87.8%99.0%+11.3 pointsmean over 3 runs per arm
Sonnet 5xhigh87.3%98.5%+11.3 pointsmean over 3 runs per arm
Sonnet 5max89.7%99.5%+9.8 pointsmean over 3 runs per arm
Opus 4.8low87.3%95.6%+8.3 pointsmean over 3 runs per arm
Opus 4.8medium85.8%97.6%+11.8 pointsmean over 3 runs per arm
Opus 4.8high88.2%98.0%+9.8 pointsmean over 3 runs per arm
Opus 4.8xhigh91.2%99.0%+7.8 pointsmean over 3 runs per arm
Opus 4.8max91.2%100.0%+8.8 pointsmean over 3 runs per arm
Fable 5low94.1%99.0%+4.9 pointsmean over 3 runs per arm
Fable 5medium92.6%100.0%+7.4 pointssingle run per arm
Fable 5high97.6%100.0%+2.4 pointsmean over 3 runs per arm
Fable 5xhigh97.1%100.0%+2.9 pointssingle run per arm
Fable 5maxnot run · it got too expensive to finish

The full record lives on the evidence page. The harness, tasks, and judge inputs are in the repo.