Sage-A1.5 · CONDITION OPTIMIZATION

From 40 historical conditions,
narrowing exploration to 89%

A causal-guided replay for AC Friedel–Crafts conditions: target definition, an eight-variable space, a mechanism graph, and five measured update rounds.

Play case replay View best condition
AC waveform Solvent droplet Causal graph node Recorded feedback 89% R⁺ EXP-038 · BEST OBSERVED89%DCM · TBABF6
40Target-reaction historical records
88 variables · mixed types
5 × 4Measured candidates in replay
89%Highest observed yield
01 · OPTIMIZATION FILM

Use 20 seconds to see optimization happen

The film follows data filtering, variable definition, causal reasoning, five candidate-feedback rounds, and the EXP-038 high point.

Sage-A1.5 · OPTIMIZATION REPLAY
00:00 / 00:20

Animated run replay. Recorded data, variables, causal structure, five rounds, and EXP-038 provide the visual material.

02 · END-TO-END FLOW

Every step has inputs,evidence, and writeback

Run data, the variable space, causal structure, and five candidate updates form one optimization chain.

PIPELINE REPLAY project-048fd560 / revision 0Historical-data driven · 5 completed batches
01 · INGESTRaw experiment dataPPT + CSV · EXP-001–074scope: EXP-001–040 02 · NORMALIZETarget definitionMaximize 2-3aa yield0–100% · trace kept nonnumeric 03 · DESIGN SPACE8-variable space5 continuous · 3 discrete10 × 8 = 80 evals 04 · REASONINGLit. → causal graph16 papers · 12 effects24 V-nodes · 35 edges 05 · LOOP5-round loop20 measured pointsrank → update BEST OBSERVEDHFIP anchorEXP-038 · yield 89% MODEL FEEDBACKRank error → editr1 edge − · r3 time + · r5 synergy +
01Raw experiment data40 target-reaction records
02Target and variable space8 variables · type preserved
03Literature and causal graph16 papers · 35 edges
04Five candidate replay rounds20 measured points
05Best observed conditionEXP-038 · 89%
02 · SEARCH HISTORY

The model follows the results,rewriting hypotheses each round

Select a round to inspect its four measured candidates, high point, and feedback into the next round.

Yield trajectory of 20 measured points

isolated_yield · historical CSV / expert-submitted

5 BATCHES COMPLETE
7 → 32%

Round 1 challenges the high-frequency hypothesis

The 10 Hz candidate reached only 7%; 0.2 Hz / 1.0 mmol reached 32%, removing a wrong positive high-frequency edge.

62 → 89%

The HFIP condition has the highest observed yield

In round 4, the result rises from 62% without additive to 89% with HFIP at 96 min, the highest observation in the recorded data.

0.89

Target parsed on a 0–100% scale

The run configuration uses 2-3aa main-product yield as isolated_yield; percentages are shown while trace and n.d. remain nonnumeric.

03 · OBSERVATION BOARD

Turn 20 observations into readable evidence

The distribution, variable weights, batch ledger, and legend are organized from records already present in this replay.

Yield distribution and quartiles

20 MEASURED OBSERVATIONS

Variable importance

REPORT WEIGHTS · 8 VARIABLES

Round-by-round observation ledger

FIVE BATCHES · FOUR MEASURED CANDIDATES EACH

04 · MODEL REFLECTION

More than parameter changes,mechanistic hypotheses are revised

Each round feeds observed ranking back into the graph: edit edges when possible and add effect nodes only for new mechanisms.

ROUND 01
32%

Low frequency beats the model’s initial high-frequency preference

10 Hz gives only 7%; 0.2 Hz and 1.0 mmol give 32%.

REMOVE_V→E · Incorrect high-frequency edge
ROUND 02
30%

Electrolyte loading is absent from the graph

The graph cannot distinguish contributions from 0.1–0.3 versus 1.0 mmol.

FLAG · Tier-3 node required
ROUND 03
43%

Add the time-driven overreaction penalty

The graph ranked times equally, but measurements differ by 23% versus 43%.

ADD_EFFECT · reaction_time sink
ROUND 04
89%

HFIP anchor region produces a high-quality point

The highest recorded yield motivates a hypothesis about cationic transition-state stabilization.

INCUMBENT · EXP-038
ROUND 05
83%

Link long half-cycles with solvent quenching

Without HFIP, low frequency extends R⁺ residence time, so quench risk needs a separate interaction.

ADD_E→E · negative synergy
CAUSAL SNAPSHOT · ROUND 0

From variable subspaces to explainable chemical effects

The causal graph does not replace experiments; it makes ranking assumptions explicit, then observations determine which connections are weakened, added, or interact.

24Variable nodes
12Effect nodes
35Total edges
frequency_hz
0.05–10 Hzcurrent_ma
2.5–20 mAadditive / solvent
HFIP · DCM · … AC half-cycle and
intermediate lifetimeC–X reduction /
electron fluxCationic transition-state
stabilizationSolvent proton quench
negative sink isolated_yield
2-3aaSolid: variable → effect → target · Dashed: round-5 interaction
05 · SEARCH SPACE

Preserve variable types,then search

Discrete values are chosen only from observed historical members; continuous bounds use the target reaction’s observed endpoints.

Weights and domains

importance weights from report

current_ma

CONTINUOUS

2.5–20 mA
Constant AC current

importance0.20

solvent

DISCRETE

MeCN · DCM · DCE · DME · THF

importance0.18

frequency_hz

CONTINUOUS

0.05–10 Hz
AC frequency

importance0.17

electrolyte

DISCRETE

Six historical electrolyte members
including TBABF6

importance0.10

substrate2_volume_ul

CONTINUOUS

44–1000 µL
Arene substrate volume

importance0.10

additive

DISCRETE

none · iPrOH · HFIP

importance0.10

reaction_time_min

CONTINUOUS

30–240 min
Electrolysis / reaction time

importance0.08

electrolyte_amount_mmol

CONTINUOUS

0.1–1.0 mmol
Supporting-electrolyte loading

importance0.07
BEST OBSERVED CONDITION

Round 4 captures
89% isolated_yield

The replay includes a 62% observation without additive and an 89% observation with HFIP. The model proposes reduced quenching and transition-state stabilization as hypotheses; these historical comparisons do not establish causality.

89%
EXP-038 · 2-3aa main-product yield · CSV record: 89%