Hugging Face's logo

Datasets:
ssslin
/
pricing-outer-loop-ablations

Modalities:
Tabular
Text
Formats:
csv
Languages:
English
Size:
< 1K
Tags:
digital-twins
survey-simulation
ablation-study
llm-agents
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
Dataset Viewer
Auto-converted to Parquet Duplicate
channel_id
stringclasses
6 values
persona_id
int64
3
51
n_anchors
int64
2
4
n_evidence
int64
2
4
n_numeric
int64
2
4
values
stringlengths
3
7
mean
float64
1
5.5
spending_restraint
3
4
4
4
7|3|2|3
3.75
spending_restraint
6
4
4
4
8|3|2|3
4
spending_restraint
8
4
4
4
6|3|2|2
3.25
spending_restraint
11
4
4
4
3|1|4|5
3.25
spending_restraint
12
4
4
4
7|2|1|2
3
spending_restraint
14
4
4
4
7|3|1|2
3.25
spending_restraint
16
4
4
4
6|3|3|3
3.75
spending_restraint
17
4
4
4
6|2|2|3
3.25
spending_restraint
21
4
4
4
3|2|3|4
3
spending_restraint
22
4
4
4
6|3|3|2
3.5
spending_restraint
23
4
4
4
7|2|3|4
4
spending_restraint
24
4
4
4
2|3|4|5
3.5
spending_restraint
28
4
4
4
4|2|4|5
3.75
spending_restraint
38
4
4
4
1|1|5|5
3
spending_restraint
40
4
4
4
7|3|3|4
4.25
spending_restraint
42
4
4
4
3|2|3|4
3
spending_restraint
44
4
4
4
6|1|2|3
3
spending_restraint
45
4
4
4
5|2|3|4
3.5
spending_restraint
46
4
4
4
6|2|3|4
3.75
spending_restraint
51
4
4
4
5|3|2|2
3
financial_capacity
3
2
2
2
2|7
4.5
financial_capacity
6
2
2
2
2|7
4.5
financial_capacity
8
2
2
2
5|2
3.5
financial_capacity
11
2
2
2
3|1
2
financial_capacity
12
2
2
2
5|1
3
financial_capacity
14
2
2
2
5|3
4
financial_capacity
16
2
2
2
3|3
3
financial_capacity
17
2
2
2
1|2
1.5
financial_capacity
21
2
2
2
5|1
3
financial_capacity
22
2
2
2
5|4
4.5
financial_capacity
23
2
2
2
5|1
3
financial_capacity
24
2
2
2
1|3
2
financial_capacity
28
2
2
2
3|2
2.5
financial_capacity
38
2
2
2
5|1
3
financial_capacity
40
2
2
2
5|1
3
financial_capacity
42
2
2
2
4|7
5.5
financial_capacity
44
2
2
2
1|3
2
financial_capacity
45
2
2
2
5|1
3
financial_capacity
46
2
2
2
2|1
1.5
financial_capacity
51
2
2
2
3|1
2
household_need
3
3
3
3
2|1|4
2.333
household_need
6
3
3
3
3|6|3
4
household_need
8
3
3
3
3|1|1
1.667
household_need
11
3
3
3
2|1|3
2
household_need
12
3
3
3
4|1|3
2.667
household_need
14
3
3
3
3|6|2
3.667
household_need
16
3
3
3
2|6|3
3.667
household_need
17
3
3
3
1|3|3
2.333
household_need
21
3
3
3
2|1|3
2
household_need
22
3
3
3
3|1|2
2
household_need
23
3
3
3
4|1|2
2.333
household_need
24
3
3
3
2|6|1
3
household_need
28
3
3
3
2|2|2
2
household_need
38
3
3
3
4|1|3
2.667
household_need
40
3
3
3
5|1|2
2.667
household_need
42
3
3
3
1|5|3
3
household_need
44
3
3
3
4|6|2
4
household_need
45
3
3
3
2|1|3
2
household_need
46
3
3
3
2|1|2
1.667
household_need
51
3
3
3
2|1|3
2
sustainable_consumption
3
3
3
3
3|2|4
3
sustainable_consumption
6
3
3
3
4|4|5
4.333
sustainable_consumption
8
3
3
3
4|4|5
4.333
sustainable_consumption
11
3
3
3
2|2|5
3
sustainable_consumption
12
3
3
3
5|5|5
5
sustainable_consumption
14
3
3
3
3|4|5
4
sustainable_consumption
16
3
3
3
2|2|4
2.667
sustainable_consumption
17
3
3
3
5|5|5
5
sustainable_consumption
21
3
3
3
3|3|4
3.333
sustainable_consumption
22
3
3
3
4|4|5
4.333
sustainable_consumption
23
3
3
3
2|3|4
3
sustainable_consumption
24
3
3
3
2|3|2
2.333
sustainable_consumption
28
3
3
3
3|3|4
3.333
sustainable_consumption
38
3
3
3
3|3|3
3
sustainable_consumption
40
3
3
3
3|4|4
3.667
sustainable_consumption
42
3
3
3
2|3|4
3
sustainable_consumption
44
3
3
3
5|5|5
5
sustainable_consumption
45
3
3
3
4|4|4
4
sustainable_consumption
46
3
3
3
1|1|1
1
sustainable_consumption
51
3
3
3
4|4|5
4.333
brand_distinctiveness
3
3
3
3
2|2|2
2
brand_distinctiveness
6
3
3
3
1|1|1
1
brand_distinctiveness
8
3
3
3
5|3|4
4
brand_distinctiveness
11
3
3
3
3|1|1
1.667
brand_distinctiveness
12
3
3
3
4|2|2
2.667
brand_distinctiveness
14
3
3
3
1|1|1
1
brand_distinctiveness
16
3
3
3
2|1|1
1.333
brand_distinctiveness
17
3
3
3
1|5|5
3.667
brand_distinctiveness
21
3
3
3
2|2|3
2.333
brand_distinctiveness
22
3
3
3
2|5|4
3.667
brand_distinctiveness
23
3
3
3
4|2|2
2.667
brand_distinctiveness
24
3
3
3
1|1|1
1
brand_distinctiveness
28
3
3
3
2|2|2
2
brand_distinctiveness
38
3
3
3
2|4|4
3.333
brand_distinctiveness
40
3
3
3
2|2|2
2
brand_distinctiveness
42
3
3
3
2|2|2
2
brand_distinctiveness
44
3
3
3
1|1|3
1.667
brand_distinctiveness
45
3
3
3
2|2|2
2
brand_distinctiveness
46
3
3
3
1|1|1
1
brand_distinctiveness
51
3
3
3
2|1|1
1.333
End of preview. Expand in Data Studio

Pricing AutoPipeline — outer-loop ablations (v2r, v4, v5, v6)

Full run artifacts for four component ablations of the pricing AutoPipeline described in MikeDeng2002/from-survey-histories-to-interpretable-digital-twins, branch outer-loop-ablations.

Each version freezes the entire procedure, changes exactly one component, and measures the effect on the same endpoint. The question is which part of an "interpretable mechanism" pipeline actually carries its accuracy.

Result

version component ablated Δ vs baseline (W3 / W4) verdict
v2r Layer 1B + Layer 2's selection judgement — keeps the two channels R1 ranked lowest instead of its two highest −.0069 / −.0208 no difference
v4 Layer 3's compiled runtime rule −.0023 / −.0023 no difference
v5 Layer 1A's choice of anchors — random draw from the answer-blind catalog, 3 seeds −.0625 / −.0563 worse
v6 Layer 1A's semantic frame — construct names, definitions, directional hypotheses, scopes +.0185 / +.0139 no difference

One component of four carries the effect. Swapping which channels are kept, deleting the compiled mechanism, and stripping every semantic label all leave the endpoint where it was. Replacing the chosen anchors with random source questions costs six points and falls below the raw-transcript baseline.

Removing the compiled rule does not even change the simulator's directional behaviour: mean absolute anchor-to-prediction correlation runs .297 → .284 → .276 across v0, v4 and v6, while the matching correlation in the human data is near zero and sign-unstable.

So in this testbed the pipeline performs evidence location and presentation, not mechanism learning. The minimal method that reproduces it is: pick roughly seven consumption-relevant survey items, extract this respondent's answers to them, show the simulator those answers.

Endpoint definition

Channel-only exact purchase accuracy on 18 validation respondents × 12 validation products = 216 cells, scored as the pair mean of two byte-identical repeats, judged at δ = .03 with a persona-clustered bootstrap CI (10,000 draws, resampling respondents rather than cells).

Power: SE ≈ .027. A "no difference" verdict rules out effects of about .05 or larger; it does not rule out smaller ones. Twelve versions have been compared against this validation set, so the pure-noise expected maximum gain is ≈ .060 — none of these results should be read as an improvement.

Models

role model
simulator gpt-5.4-nano, reasoning_effort=high, OpenAI Batch
discovery layers (1A / critic / 2 / 3) gpt-5.6-terra

Held fixed across every version so that cross-version comparisons stay meaningful.

⚠️ Sensitive content — read before use

This dataset contains material the GitHub release deliberately excludes:

  • full persona transcripts — the RAW+ADD arms carry each respondent's complete survey history (~96,000 characters per prompt) inside requests.jsonl
  • respondent identifiers — persona_id throughout
  • human answers — sidecar.json items carry wave3_answer and wave4_answer; cells_*.csv carries predicted plus correct_wave3/4, from which the human answer is recoverable
  • model outputs — batch_output.jsonl

These are pseudonymous respondents from a public research panel, not identified individuals, and this release inherits the CC BY 4.0 terms of the upstream dataset. Handle accordingly.

Provenance and licence

Derived from LLM-Digital-Twin/Twin-2K-500 (CC BY 4.0) by way of Mike-deng-2002/pricing_and_cognitive_bias_experiment, which holds the v0 baseline run these ablations are built from. Every prompt here is a v0 archived prompt with one component swapped or removed. Released under CC BY 4.0; please cite the upstream dataset.

Sealed final test

The split manifest reserves 12 respondents × 8 products as an untouched final test. No cell of it appears anywhere in this dataset. The guards were mechanical rather than by convention: builders assert on sealed ids, the reporting script exits if the validation sets intersect them, and the input bundle omits their labels entirely so no script had anything to score against.

If you evaluate on it, say so — once its results inform a change it stops being a final test.

Files 

shared/     SPLIT_MANIFEST_v2.json     the frozen three-way split (ids only, self-hashed)
            ENDPOINTS_v0/v1/v2.json    aggregate endpoint tables with bootstrap CIs
            channel_registry_v0.json   Layer 1A's six channels (also public on GitHub)
            evidence_bank_v0.json      20 development respondents' answers to the anchors
            evidence_states_dev.csv    parsed numeric values per respondent per anchor
            labels_prices.json         prices and both waves' answers, D_dev ∪ D_val only

v2r_v3r/    requests.jsonl  sidecar.json  cells_rerun.csv  batch_output.jsonl  batch_state.json
            l3_compile_result.json  ×2    terra's Layer 3 compiles for both channel pairs
v4/         requests.jsonl  sidecar.json  cells_v4.csv     batch_output.jsonl  batch_state.json
v5/         requests.jsonl  sidecar.json  cells_v5.csv     batch_output.jsonl  batch_state.json
            draws.json                    3 seeded anchor draws + terra's labels and compiles
v6/         requests.jsonl  sidecar.json  cells_v6.csv     batch_output.jsonl  batch_state.json

cells_*.csv is the scored unit: one row per respondent × product × arm, with correct_wave3, correct_wave4 and valid.

Reproducing

Code, protocol, per-version hypothesis cards, freeze specifications and ledgers live on branch outer-loop-ablations; start from runs/RUNNING.md. The builders are deterministic — it records the expected input_sha256 for each version, so a fresh checkout can be verified before spending anything on the API.

version input_sha256 requests
v2r + v3r e62a771f05823e5b48e598817b81433a477a29c3038c0a20daaeb98d3a8beb00 144
v4 b92ccc91d483bc6be2b58960a94127111d39a11a8e972ba5e025998a9d6ac658 72
v5 b0fa632e150025c0bda6e1e51ba01da24108e743dc765fffccfab29743a0f489 144
v6 3088329d096644f46e0dd5a34c3cdd1a8150a3259b040410c0698ba3ab61c7ee 72
Downloads last month
30
Number of rows:
120
Total file size:
39.4 MB