Parent: "Did you do your homework?" Child: "I was just about to do my homework, mom!"
— Phantom Pinpoint Effect, canonical anecdote (English translation).
A subject (S) holds an abstract goal region (G \subset \Theta) — for example, the diffuse intent "I should do my homework today". When an external trigger (E) (a parent walking in, an unforeseen market event, a journalist's question) happens to graze (G), the subject claims — post-hoc — that they were targeting a specific point (p^* \in G) all along. The specificity is fabricated; the agent simply collapses the region to a point the moment the region is touched.
PP is the identifiable conjunction of:
- Texas Sharpshooter projection — (p^* = \Pi_G(E)).
- Hindsight bias — claimed pre-trigger confidence > actual.
- Confabulation — sharpness fabricated post-hoc.
- Audience-driven cheap-talk — claim sharpness scales with (\gamma).
phantom-pinpoint is a fully-typed, MIT-licensed numpy implementation
that simulates PP, supplies a clean negative control ablation, and
ships pre-registered hypotheses with bootstrap confidence intervals.
git clone https://github.com/hinanohart/phantom-pinpoint
cd phantom-pinpoint
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
# reproduce every figure and result table
bash scripts/reproduce_all.shfrom phantom_pinpoint import PhantomPinpointModel, bootstrap_ci
# A modestly vague goal region under a public audience
model = PhantomPinpointModel(audience_size=2.0)
result = model.simulate(n_runs=1000, seed=42)
print(bootstrap_ci(result.delta_pp).as_dict())
# {'statistic': 16.78, 'ci_lo': 15.37, 'ci_hi': 18.23, ..., 'is_significant': True}# CLI
phantom-pinpoint baseline --n-runs 1000 --output results/baseline.parquet
phantom-pinpoint ablations --n-runs 1000 --output results/ablations.parquet
phantom-pinpoint figure ablation --input results/ablations.parquet --output figures/ablations.pdfThe log Bayes factor
[ \Delta_{PP}(p^{}) = \log,\mathcal N(p^{};,\text{anchor},,\sigma_{conf}^{2}I);-; \log,\mathcal N(p^{*};,\mu_{\text{post}},,\sigma_{\text{post}}^{2}I) ]
cleanly separates the Strategic (Phantom Pinpoint) and Bayesian generative models.
| mean ΔPP (nats) | 95 % bootstrap CI | post-hoc fit rate | |
|---|---|---|---|
| Strategic baseline | +16.78 | [+15.37, +18.23] | 0.164 |
| Bayesian neg-control | −25.78 | [−28.21, −23.42] | 0.198 |
Pre-registered acceptance criteria AC1 (strategic CI > 0) and AC2 (Bayesian CI < 0) both PASS.
| Ablation | Mean ΔPP | post-hoc fit | BH-FDR rejected |
|---|---|---|---|
| baseline | +16.78 | 0.164 | — |
| A1 (α≈0) | +17.16 | 0.133 | no |
| A2 (β≈0) | +9.34 | 0.246 | yes |
| A3 (γ=0) | +16.83 | 0.156 | no |
| A4 (shuffle G) | +25.87 | 0.095 | yes |
| A5 (shuffle E) | +97.36 | 0.024 | yes |
| A6 (Bayesian) | −25.78 | 0.198 | yes |
| A7 (σ→0) | +35.34 | 0.169 | yes |
Interpretation:
- A6 Bayesian is the negative control: ΔPP CI strictly below zero by construction, sealing the discriminative claim against the "PP is just hindsight bias" critique.
- A2 β=0 halves ΔPP because the strategic anchor reduces to (\Pi_G(E)) without any defensive pull, leaving more room for the Bayesian model to explain data drawn from the strategic model.
- A4/A5 shuffles drop the post-hoc-fit rate dramatically (−42 % / −85 %) and re-confirm DP4: shuffling the target or trigger destroys the Texas-Sharpshooter signature.
- A7 σ→0 sharpens the strategic claim to a δ-spike, inflating the log-Bayes-factor.
The claim kernel width (\sigma_{conf}) shrinks monotonically with audience size (\gamma) (Spearman (\rho = -1.0), (p = 1.4\times10^{-24})), matching the prediction that public claims sharpen. Unlike traditional hindsight-bias accounts, the PP framework predicts that this sharpening is audience-mediated, not data-driven.
We registered six hypotheses (docs/preregistration.md) before the data
collection. Two failed and we report both:
- H2b (audience drives ΔPP up monotonically) — failed, because the log Bayes factor is dominated by location mismatch and γ only modulates width. H2a (audience sharpens (\sigma_{conf})) passes cleanly, so we interpret the failure as a clean decomposition: γ governs specificity, β and the geometry of (G) govern direction.
- H6 (sharper trigger distribution → larger ΔPP) — failed with sign reversal. Wider (g_\sigma) leaves more triggers far from (G), at which point anchor (on the (G) boundary) and (\mu_{\text{post}}) (near the trigger) diverge more. This is consistent with the PP framework but contradicts the naive intuition.
Both failures are now explained analytically by the v0.2.0 ΔPP decomposition (see "What's new in v0.2.0" below): H2b's audience signal lives entirely in (\Delta_{PP}^{\text{width}}) (H9 PASS, ρ=+1.0, p=1.4×10⁻²⁴) and H6's sign reversal is fully captured by the (\Delta_{PP}^{\text{loc}}) term (H10 PASS).
Each grey segment connects a trigger (E) (yellow) to its strategic claim (p^*) (purple); the red circle is the goal-region boundary (\partial G). The Texas-Sharpshooter projection is visible as the radial collapse of external triggers onto (\partial G).
docs/theory.md— full mathematical derivationdocs/preregistration.md— pre-registered H1–H10 across confirmatory families C1 (v0.1.0) and C2 (v0.2.0), falsification criteria F1–F5, and the full honest-failure recordCHANGELOG.md— Keep a Changelog 1.1.0 release notesdocs/examples.md— anecdote inventory ("homework excuse", politician, trader, "あざと可愛い", sibling fight, SNS 匂わせ, sports commentary)
phantom-pinpoint/
├── src/phantom_pinpoint/ # library (pure numpy + scipy)
│ ├── core.py # 5 governing equations, PhantomPinpointModel
│ ├── agents.py # Subject / Audience / Trigger pedagogical wrappers
│ ├── simulation.py # run_condition / run_grid / parameter_sweep
│ ├── statistics.py # bootstrap_ci / permutation_test / BH-FDR
│ ├── ablations.py # A1–A7 pre-registered grid
│ ├── decomposition.py # ΔPP = ΔPP_width + ΔPP_loc closed-form (v0.2.0)
│ ├── effect_size.py # Cohen d / Hedges g / Cliff δ + bootstrap (v0.2.0)
│ ├── sensitivity.py # ±50 % univariate sweep, AC10 gate (v0.2.0)
│ ├── identifiability.py # geometric degeneracy diagnostic, AC9 (v0.2.0)
│ ├── visualization.py # 6 figures, atomic save
│ └── cli.py # typer-based CLI
├── tests/ # pytest, 110+ tests, ≥75 % coverage gate (92 % actual)
├── experiments/ # 9 reproducible scripts (01–05 v0.1.0, 06/07/09/10 v0.2.0)
│ ├── 01_baseline.py
│ ├── 02_audience_effect.py
│ ├── 03_vague_vs_sharp.py
│ ├── 04_ablations.py
│ ├── 05_repeated_trigger.py
│ ├── 06_sensitivity_sweep.py # AC10
│ ├── 07_identifiability.py # AC9
│ ├── 09_location_decomposition.py # H2b/H6 root-cause via decomposition
│ └── 10_effect_sizes.py # AC11/AC12 effect-size sheet
├── docs/ # theory / preregistration / examples / figures
├── CHANGELOG.md # Keep a Changelog 1.1.0
├── scripts/reproduce_all.sh
├── pyproject.toml # hatchling, py>=3.11
└── LICENSE # MIT
- Every experiment seeds
numpy.random.default_rngexplicitly. - All Parquet outputs are written atomically (tmp → rename) and carry a manifest JSON with git commit, Python version, platform, timestamp.
scripts/reproduce_all.shregenerates every figure and Parquet from a clean checkout in under one minute on a laptop CPU.- CI matrix runs Python 3.11 and 3.12 on Ubuntu, gating merges on
pytest --cov-fail-under=80.
@software{hinanohart_phantom_pinpoint_2026,
author = {hinanohart},
title = {{Phantom Pinpoint: Agent-Based Simulation of Post-hoc Specificity Confabulation}},
year = {2026},
version = {0.2.0},
url = {https://github.com/hinanohart/phantom-pinpoint},
license = {MIT}
}MIT. Contributions, replications, and human-subjects validations welcome — please open an issue or PR.
The v0.1.0 Δ_PP signature was dominated by location mismatch and therefore failed to detect audience-driven sharpening (H2b). v0.2.0 adds the closed-form decomposition
[ \Delta_{PP};=; \underbrace{\tfrac{d}{2}\log!\Bigl(\tfrac{\sigma_{\text{post}}^{2}}{\sigma_{conf}^{2}}\Bigr)}{\Delta{PP}^{\text{width}}} ;+; \underbrace{\tfrac12\Bigl(\tfrac{|p^{}-\mu_{\text{post}}|^{2}}{\sigma_{\text{post}}^{2}}-\tfrac{|p^{}-\text{anchor}|^{2}}{\sigma_{conf}^{2}}\Bigr)}{\Delta{PP}^{\text{loc}}}. ]
The width component recovers the audience-driven prediction that v0.1.0 missed — Spearman ρ = +1.0, p = 1.4 × 10⁻²⁴ (H9 PASS). The location component analytically explains v0.1.0's H6 sign reversal (H10 PASS).
| ID | Criterion | Status |
|---|---|---|
| AC8 | width + location identity holds within 1e-9, H9/H10 PASS | ✅ PASS |
| AC9 | identifiability degeneracy diagnostic | ❌ FAIL (honest, see preregistration §"Honest failures") |
| AC10 | ±50 % univariate sweep preserves sign in ≥ 90 % of cells | ✅ PASS (100 %) |
| AC11 | every primary contrast reports Cohen's d + bootstrap 95 % CI | ✅ PASS |
| AC12 | ρ-only reporting forbidden | ✅ ENFORCED |
from phantom_pinpoint import (
PhantomPinpointModel, decompose_simulation, bootstrap_effect_size,
)
model = PhantomPinpointModel(audience_size=4.0)
result = model.simulate(n_runs=1000, seed=42)
# decomposition needs prior_means + betas — see experiments/09_*.pyv0.2.0 (2026-04-27). AC1, AC2, AC4, AC5, AC6, AC8, AC10, AC11, AC12
pass. AC9 (geometric identifiability diagnostic) honestly fails — the
geometric degeneracy region does not align with the statistical
un-identifiability region; v0.3.0 will sharpen with a Mahalanobis
variant. See CHANGELOG.md and
docs/preregistration.md for the full record.