license: cc-by-4.0
language:
- en
tags:
- digital-twins
- survey-simulation
- ablation-study
- llm-agents
pretty_name: Pricing AutoPipeline Outer-Loop Ablations
size_categories:
- n<1K
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_idthroughout - human answers —
sidecar.jsonitems carrywave3_answerandwave4_answer;cells_*.csvcarriespredictedpluscorrect_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 |