# Stage-3 hold-out skip patterns -- generation report

Generated by `simulate_patterns.py` (this directory). Inputs are read-only:

- Extension logic: `S:\30_OriginalApps\16_HareSkip\forge-neo-Anima-HareSkip\hareskip\{skip_pattern,probability_models}.py`
- Trajectory: `S:\30_OriginalApps\16_HareSkip\experiment-HareSkip\analysis-phase1-02\outputs\tables\master_long.csv`
- Exclusion sets: `S:\30_OriginalApps\16_HareSkip\experiment-HareSkip\patterns\stage2-interaction-patterns.txt`, `S:\30_OriginalApps\16_HareSkip\experiment-HareSkip\patterns\layer1-single-skip-sweep-30steps.txt`

## Configuration

| item | value |
|---|---|
| num_steps | 30 |
| probability model | `sigmoid_band_v0.1` |
| skip_window | (0.05, 0.95) (extension default) |
| zone_boundaries | (-4.0, 0.0) (extension default) |
| zone max streak | danger 1 / middle 2 / safe 3 |
| skip_seed | `sha256(f"{image_seed}|hareskip|{offset}")` mod 2**63 |
| image seeds | 3000995193, 2455776111, 2767019676 |

## Verification

Two independent implementations were compared on **every candidate drawn**
(not just the 9 selected ones):

1. `self_generate` in this script -- written from scratch by reading the
   extension source (probability formula, window eligibility, z proxy,
   zone/streak trimming with the `(p, z, index)` argmin flip rule, and the
   RNG consumption order: `rng.random()` is drawn **only for eligible
   steps**, because in `_draw_pattern` the call sits inside the
   `if eligible` conditional expression).
2. `hareskip.skip_pattern.generate_skip_pattern` imported directly from the
   extension repo (`sys.path.insert`, read-only; nothing installed or
   modified).

Compared per candidate: `params`, `z_by_step`, `p_by_step`, the boolean
`skip` vector and `skipped_steps`. Any mismatch aborts the script.

- Candidates cross-checked: **10** -- all identical.
- `derive_skip_seed` also cross-checked per candidate (SHA-256, not `hash()`).
- The extension's own suite `tests/test_skip_pattern.py` was run against the
  same modules: **42 passed** (including the pinned literal
  `derive_skip_seed(12345, 0) == 3695650839502921262`).

Validation of the final 10 patterns: **no problems (all steps in 2..30, no step 1, no duplicates, no collision with stage-1/2, pairwise Hamming >= 3 among the 9 simulated)**

## z(step) correspondence

`master_long.csv` columns `z` / `tnow` follow **mapA**
(`traj_step = skip_step - 1`), the default in analysis-phase1-02/REPORT.md.
For the ERSDE-Beta sampler all **15 conditions carry identical `z`/`tnow`**
for every `skip_step` 2..30 (verified: zero conflicting values).
Each recorded `z` reproduces `logsnr_proxy_from_t_now(tnow)` to < 1e-9, so
the measured `tnow` column is fed straight into the extension as
`t_now_by_step`.

Step 1 has no measured trajectory row (skip_step starts at 2); its z is
extrapolated linearly backwards to z = -14.8536. It is irrelevant to the
result: idx 0 has progress 0.0 < 0.05 so it is outside the skip window,
and at inference the first call has no residual, so it is never skipped.

| step | t_now | z | zone |
|---:|---:|---:|---|
| 1 (extrap.) | 0.999405273 | -14.85363 | danger |
| 2 | 0.997992039 | -12.41725 | danger |
| 3 | 0.993243277 | -9.98088 | danger |
| 4 | 0.986301303 | -8.55332 | danger |
| 5 | 0.978079259 | -7.59631 | danger |
| 6 | 0.967707634 | -6.80020 | danger |
| 7 | 0.955337703 | -6.12587 | danger |
| 8 | 0.941548824 | -5.55867 | danger |
| 9 | 0.925727427 | -5.04568 | danger |
| 10 | 0.907582879 | -4.56895 | danger |
| 11 | 0.887755156 | -4.13603 | danger |
| 12 | 0.865482271 | -3.72318 | middle |
| 13 | 0.840949059 | -3.33061 | middle |
| 14 | 0.813186824 | -2.94170 | middle |
| 15 | 0.783667684 | -2.57434 | middle |
| 16 | 0.750749290 | -2.20523 | middle |
| 17 | 0.713836491 | -1.82818 | middle |
| 18 | 0.674917459 | -1.46102 | middle |
| 19 | 0.630938530 | -1.07249 | middle |
| 20 | 0.584249079 | -0.68048 | middle |
| 21 | 0.533505082 | -0.26844 | middle |
| 22 | 0.479591817 | 0.16336 | safe |
| 23 | 0.420863301 | 0.63846 | safe |
| 24 | 0.360182405 | 1.14915 | safe |
| 25 | 0.298076898 | 1.71294 | safe |
| 26 | 0.233108103 | 2.38169 | safe |
| 27 | 0.170212761 | 3.16824 | safe |
| 28 | 0.113678373 | 4.10741 | safe |
| 29 | 0.060460649 | 5.48679 | safe |
| 30 | 0.020710060 | 7.71242 | safe |

## Patterns

| # | a | image_seed | offset | skip_seed | n | skipped steps | danger | middle | safe | max streak | E[skips] |
|---:|---:|---:|---:|---|---:|---|---:|---:|---:|---:|---:|
| 1 | 0.3 | 3000995193 | 0 | 4101374585893130546 | 5 | 15, 17, 23, 24, 25 | 0 | 2 | 3 | 3 | 7.09 |
| 2 | 0.3 | 2455776111 | 0 | 4945151993939216742 | 5 | 19, 22, 24, 25, 27 | 0 | 1 | 4 | 2 | 7.09 |
| 3 | 0.3 | 2767019676 | 0 | 9040233029051876849 | 7 | 12, 18, 19, 22, 23, 24, 27 | 0 | 3 | 4 | 3 | 7.09 |
| 4 | 0.6 | 3000995193 | 0 | 4101374585893130546 | 7 | 11, 17, 18, 20, 23, 24, 25 | 1 | 3 | 3 | 3 | 11.15 |
| 5 | 0.6 | 2455776111 | 0 | 4945151993939216742 | 8 | 13, 15, 17, 19, 23, 24, 25, 27 | 0 | 4 | 4 | 3 | 11.15 |
| 6 | 0.6 | 3000995193 | 1 | 5743156414877519378 | 10 | 9, 12, 13, 15, 17, 18, 22, 23, 26, 28 | 1 | 5 | 4 | 2 | 11.15 |
| 7 | 0.9 | 3000995193 | 0 | 4101374585893130546 | 5 | 9, 12, 23, 24, 25 | 1 | 1 | 3 | 3 | 15.94 |
| 8 | 0.9 | 2455776111 | 0 | 4945151993939216742 | 9 | 7, 11, 13, 15, 17, 23, 24, 25, 27 | 2 | 3 | 4 | 3 | 15.94 |
| 9 | 0.9 | 2767019676 | 0 | 9040233029051876849 | 8 | 6, 8, 14, 15, 23, 24, 25, 27 | 2 | 2 | 4 | 3 | 15.94 |
| 10 | 0.9 | 2767019676 | 0 | 9040233029051876849 | 10 | 3, 5, 6, 8, 14, 15, 23, 24, 25, 27 **(+manual 3, 5)** | 4 | 2 | 4 | 3 | 15.94 |

Rows 1-3: a=0.3. Rows 4-6: a=0.6. Rows 7-9: a=0.9. Row 10: extreme-stress.

### Row 10 (extreme stress)

Row 10 is **not** a plain simulator output. It takes the realised a=0.9
pattern of row 9 (image_seed %d, offset %d) and **manually adds steps 3
and 5**. Those two steps sit in the danger band (z = -9.98 and -7.60, both
z < -4), which the probability model almost never selects and where the
danger-zone streak cap is 1 -- so this row is a deliberate out-of-sample
probe of the prediction formula's limit, not a reachable extension output.
Its max streak (3) may therefore exceed what the extension would emit.

## Output grammar

`stage3-holdout-patterns.txt` -- 10 lines, one pattern per line,
comma-separated 1-based step numbers in ascending order; same grammar as
`stage2-interaction-patterns.txt` and the stage-1 sweep file.