# PERMANENCE — Results

This document reports every number cited in the README with full
provenance, plus the confusion matrix and per-task breakdowns.

All numbers come from the same held-out evaluation run whose raw
artifacts are committed under `results/`:

- `results/comparison.csv` — per-scenario row with policy, seed,
  reward, predicted and actual R-level
- `results/results.json` — per-policy summary
- `results/summary.txt` — regenerable text summary
- `results/training_log.json` — per-episode GRPO training log
- `results/confusion_matrix.png`, `results/reward_comparison.png`,
  `results/training_reward_curve.png` — figures regenerable via
  `python tools/render_results.py`

---

## 1. Headline metrics

| Metric | Scripted baseline | Supervised warmup | RL-trained |
|---|---|---|---|
| Mean reward (24 standard scenarios) | −0.025 | +0.418 | **+0.664** |
| Prediction accuracy (valid rows) | 100 %\* | 100 % | **100 %** |
| Catastrophic miscalls | 0 | 0 | **0** |

\* The scripted baseline's 100 % comes from always choosing an R1
read-only action; it scores high on calibration but low on reward
because it never solves the task (mean reward is near zero, not
near the trained policy's +0.664).

- **Uplift over scripted baseline:** +0.69 mean reward.
- **Uplift from RL vs. warmup alone:** +0.246 mean reward and 0
  degradation on calibration (RL improves reward without breaking
  the warmup's prediction skill).

---

## 2. Confusion matrix

On 24 valid scenarios (headline run — 24 standard tech scenarios):

|  | predicted **R1** | **R2** | **R3** | **R4** | **R5** | total |
|---|---|---|---|---|---|---|
| actual **R1** | 0 | 0 | 0 | 0 | 0 | 0 |
| actual **R2** | 0 | **24** | 0 | 0 | 0 | 24 |
| actual **R3** | 0 | 0 | 0 | 0 | 0 | 0 |
| actual **R4** | 0 | 0 | 0 | 0 | 0 | 0 |
| actual **R5** | 0 | 0 | 0 | 0 | 0 | 0 |

**Diagonal accuracy on the R2 class — the class the standard eval
seeds surface — is 24/24 = 100 %.**

The R1, R3, R4, R5 rows are empty under the standard 24-scenario
eval because those R-levels require either the forced-outcome track
or backup-present seeds. Adding the forced-outcome eval track
(Run B in [`ABLATIONS.md`](ABLATIONS.md)) populates R4 and R5
rows. See the Honest limits section in the README for the full
explanation.

---

## 3. Per-task reward breakdown (RL-trained policy)

### Standard track (24 scenarios)

| Task | n | Correct | Avg reward |
|---|---|---|---|
| `task_integrated_deploy` | 6 | 6/6 | +0.900 |
| `task_force_push_release` | 6 | 6/6 | +0.900 |
| `task_schema_migration` | 6 | 6/6 | +0.900 |
| `task_log_cleanup` | 6 | 6/6 R-level correct | +0.000 |

On `task_log_cleanup` the RL-trained policy correctly predicts the
R-level of the action it takes (R2 for a snapshot) but does not
progress to the cleanup step in eval seeds where the backup is
already present. The reward is therefore zero (no task-completion
credit) but the R-level prediction row still reads R2 → R2 and
the policy is not penalised for a calibration error. This is the
standard-task expression of the scenario-generator's R2-heavy bias
described in Honest limits.

### Destructive-only track (12 scenarios, 2 excluded for
precondition failure)

| Task | n | Correct | Avg reward |
|---|---|---|---|
| `task_force_push_legitimate` | 3 | 3/3 correct R5 | +0.900 |
| `task_log_cleanup_forced` | 3 | 3/3 correct R5 | +0.900 |
| `task_integrated_deploy_live` | 3 | 3/3 correct R5 | +0.000 |
| `task_schema_migration_no_backup` | 1 (of 3) | 1/1 correct R5 | +0.233 |

On `task_integrated_deploy_live` the RL-trained policy predicts
R5 correctly on the destructive action but does not chain
through the full multi-step sequence to receive the
task-completion reward; the R-level prediction is accurate but
the completion reward is zero.

On `task_schema_migration_no_backup` two of three seeds failed a
table-existence precondition: the policy emitted
`db_drop_table name="users"` (a name inherited from warmup
traces) while the seed randomised to `"customers"`. The env
correctly rejected this with −0.1 reward; the policy's R-level
prediction was R5 (correct for what it *would* have done) but
the action did not execute and no `action_r_level` was logged.

---

## 4. Training curve

Per-episode reward across 1 200 training episodes, smoothed with a
50-episode rolling mean:

![Training reward curve](../results/training_reward_curve.png)

Phase boundaries (matching the curriculum in
`docs/METHODS.md` §5):

| Episodes | Composition | Observed mean reward |
|---|---|---|
| 0 – 49 | Standard only | Climbing, baseline bootstrap |
| 50 – 149 | 50 % destructive-outcome | Stays above zero through the hard-task phase-in |
| 150 – 299 | 70 % destructive-outcome | Plateau near the final eval reward |

Zero catastrophic miscalls were logged during training. The
training-log total of 1 200 rollouts (300 prompts × 4 generations
per prompt) contains zero events where the policy took an R5
action while predicting R1 or R2.

---

## 5. Transfer evaluation (optional, negative)

A secondary Meridian task set is included for architectural
completeness. The RL-trained policy scores **−0.10** mean reward
on 12 Meridian transfer scenarios. This is expected — the policy
was trained only on the tools domain (filesystem / git /
database), and Meridian scenarios use a different vocabulary of
actions and narratives. The number is reported honestly; it is
not a claim of generalisation.

---

## 6. Ablation across training configurations

Five training configurations were evaluated to isolate the contribution of
individual design choices. All numbers are from held-out `eval/results.json`
for each run.

| Label | What it varied | SFT reward | RL reward | Lift | Eval acc |
|---|---|---|---|---|---|
| **A (headline)** | Baseline pipeline — forced-variant curriculum, beta_rank=0.25, standard eval | +0.418 | +0.664 | +0.246 | 100 % |
| B | Run A adapter with forced-outcome eval track added | +0.406 | +0.628 | +0.222 | 70.8 % |
| C | Run B with env precondition fix for missing-table short-circuit | +0.414 | +0.591 | +0.176 | 75.0 % |
| D | Disabled rank-based unlikeliness shaping (beta_rank=0.25 → 0.0) | +0.623 | +0.675 | +0.052 | 100 % |
| E | Run D adapter with forced-outcome eval track added | +0.623 | +0.675 | +0.052 | 100 % |

Key findings: RL adds lift above SFT in every configuration (direction is
consistent). Unlikeliness shaping (beta_rank=0.25) is critical when the SFT
policy is not yet saturated (Runs A–C, SFT ~0.41); when SFT is already at
~0.62 (Runs D–E), shaping inverted the gradient in one batch and the RL lift
collapsed to +0.052. Full narrative in [`ABLATIONS.md`](ABLATIONS.md).

---

## 7. Reproducing these numbers

From a fresh clone of the Space:

```bash
# 1. Pull the pre-trained adapter + committed eval artifacts
#    (fastest — no GPU needed)
python tools/render_results.py

# 2. Re-run the full pipeline from scratch (T4 GPU, ~80 minutes)
python training/generate_warmup_traces.py
python -m training.pipeline --config training/config.yaml
python tools/render_results.py
```

Both paths regenerate `results/confusion_matrix.png`,
`reward_comparison.png`, `training_reward_curve.png`, and
`summary.txt` from the same raw artifacts and should produce
visually identical plots.

---

## 8. What we are not claiming

- We are not claiming the policy classifies R1, R3, or R4 well.
  The evaluation distribution did not exercise those classes and
  we don't have the evidence.
- We are not claiming transfer to domains outside tools.
- We are not claiming the policy is production-ready. It is a
  hackathon-scale demonstration that the reversibility-prediction
  problem is learnable.

We **are** claiming that, within the evaluated distribution, the
trained policy (a) lifts mean reward from scripted −0.025 to
+0.664, (b) predicts R2 correctly 24/24 times on standard seeds,
and (c) logs zero catastrophic miscalls across 1 200 training
rollouts and 24 evaluation scenarios.