Update readme

EVAL.md

@@ -1,8 +1,12 @@
 # eval.py: Floor sanity check
 
-A minimal check. Use this to confirm your predictions parse and have the
-right shape. For full Floor + Mid + Rich scoring, see the companion
-evaluator in the same Hugging Face repository.
+A minimal smoke test. It confirms your predictions parse and have the
+right shape. It does NOT score recommendation quality.
+
+For deeper scoring (keyword matching, multi-tier reasoning, fixture
+citations), bring your own evaluator that compares each prediction
+against the `handcrafted_recommendation.json` in the matching scenario
+folder.
 
 ## Usage
 
@@ -24,13 +28,15 @@ python eval.py --predictions sample_predictions.json
 
 - Whether the recommendation engages with the right evidence.
 - Whether multi-tier scenarios cite both tiers.
-- Whether named fixtures are referenced.
-- Whether cost or projection numbers are present.
+- Whether named fixtures from the metadata are referenced.
+- Whether cost or projection numbers are reasonable.
 
-Those checks live in the companion evaluator.
+Those checks are quality assessments. They depend on what you want to
+score for and how strict you want to be. The dataset ships the gold
+answers; the scoring method is up to you.
 
 ## Exit codes
 
-- `0`: all predictions passed.
+- `0`: all predictions passed the Floor check.
 - `1`: at least one prediction failed at least one check.
 - `2`: usage error (missing file, malformed JSON).

README.md

@@ -37,7 +37,7 @@ recommendation that says what to change and what the impact will be.
 
 The dataset is synthesized. Telemetry was generated procedurally to match
 each scenario's narrative. Gold recommendations were hand-crafted and
-verified against per-scenario expectations.
+verified.
 
 The dataset uses AWS vocabulary throughout. Instance types, service names,
 and field names match AWS. This makes the scenarios concrete instead of
@@ -46,6 +46,11 @@ vendor-neutral.
 ## Folder layout
 
 ```
+README.md                                # this file
+LICENSE                                  # MIT
+EVAL.md                                  # what eval.py checks
+eval.py                                  # Floor sanity check (smoke test)
+sample_predictions.json                  # worked example of submission shape
 scenarios_summary.jsonl                  # one row per scenario (viewer table)
 scenarios/
   01/
@@ -92,7 +97,10 @@ Columns in the summary table:
 | `savings_monthly_usd`          | gold.cost_impact.savings_monthly_usd           |
 | `current_monthly_usd`          | gold.cost_impact.current_monthly_usd           |
 | `projected_monthly_usd`        | gold.cost_impact.projected_monthly_usd         |
-| `sla_availability_preserved`   | gold.projected_state.sla_availability_preserved|
+
+Some scenarios have negative `savings_monthly_usd`. That is expected. For
+those scenarios the right action increases cost to fix a performance or
+reliability problem (for example, adding a read replica).
 
 ## Schema
 
@@ -173,9 +181,8 @@ You can use this dataset two ways.
 **Train or fine-tune.** Treat each scenario's telemetry plus metadata as
 input. Use the `handcrafted_recommendation.json` as the target output.
 
-**Score predictions.** Use the companion evaluator in the same Hugging
-Face repository. It scores predictions across three tiers: Floor, Mid,
-Rich.
+**Evaluate AI agents.** Run your agent on the scenario inputs. Compare its
+output to the hand-crafted recommendation in that scenario's folder.
 
 ## Quick sanity check
 
@@ -183,43 +190,37 @@ Rich.
 python eval.py --predictions sample_predictions.json
 ```
 
-This runs the Floor competency check. It confirms your predictions parse,
-have the required fields, and use allowed category values. For full
-scoring, see the companion evaluator.
+This runs the bundled Floor sanity check. It confirms your predictions
+parse, have the required fields, and use allowed category values. It does
+NOT score recommendation quality. See `EVAL.md` for what is checked.
 
 ## Prediction shape
 
 See `sample_predictions.json` for a worked example. Required fields per
 prediction: `scenario_id`, `finding_type`, `specific_change`,
-`primary_tier`, `action_category`. Optional but used by Mid and Rich:
-`secondary_tier`, `reasoning`, `evidence`, `projected_state`,
+`primary_tier`, `action_category`. Optional but useful for deeper
+scoring: `secondary_tier`, `reasoning`, `evidence`, `projected_state`,
 `cost_impact`, `risk_assessment`.
 
-## How to evaluate
-
-Use the companion evaluator in the same Hugging Face repository. It scores
-predictions across three tiers.
-
-- Floor: structure and category checks. Any reasonable agent passes 18 out
-  of 18.
-- Mid: action keywords plus multi-tier reasoning.
-- Rich: fixture citations plus quantified projections.
+## How to score beyond the Floor check
 
-A typical run:
+The dataset ships gold answers and a Floor sanity check. It does not ship
+a quality scorer. Beyond the Floor check, the scoring method is up to
+you. Common options:
 
-```bash
-python <evaluator-folder>/eval.py \
-    --predictions your_predictions.json \
-    --dataset .
-```
+- Exact match on the enum fields (`finding_type`, `primary_tier`,
+  `action_category`).
+- Keyword or substring checks on `specific_change`.
+- Semantic similarity on the prose fields.
+- A custom rubric per scenario, comparing prediction fields against the
+  matching `handcrafted_recommendation.json`.
 
 ## Intended uses
 
 - Train or fine-tune a model that maps cloud telemetry to an optimization
   recommendation.
 - Evaluate AI agents on cloud-optimization reasoning.
-- Compare single-shot vs orchestrated agent designs. The Rich tier is
-  calibrated so that orchestration matters.
+- Compare single-shot vs orchestrated agent designs.
 
 ## License
 

eval.py

@@ -1,14 +1,18 @@
 """Floor sanity check for the synthesized cloud-optimization recommendations dataset.
 
-This is a minimal check. It reads your predictions file and confirms:
+This is a minimal smoke test. It reads your predictions file and confirms:
     1. Each prediction parses as JSON.
     2. Required fields are present.
     3. finding_type is one of the three allowed values.
     4. primary_tier is one of the allowed tier names (or null).
-    5. specific_change is a non-empty string of reasonable length.
+    5. action_category is one of the allowed values.
+    6. specific_change is a non-empty string of reasonable length.
 
-For full Floor + Mid + Rich scoring, see the companion evaluator in the same
-Hugging Face repository.
+This script does NOT score recommendation quality. It only confirms that
+predictions are well-formed and on-topic. For deeper scoring (keyword
+matching, multi-tier reasoning, fixture citations), bring your own
+evaluator that compares each prediction against the
+handcrafted_recommendation.json in the matching scenario folder.
 
 Usage:
     python eval.py --predictions sample_predictions.json
@@ -98,7 +102,7 @@ def main():
     print()
     if total_errors == 0:
         print(f"  All {len(preds)} prediction(s) passed the Floor sanity check.")
-        print("  For full Floor + Mid + Rich scoring, see the companion evaluator.")
+        print("  Quality scoring beyond this is up to you.")
         sys.exit(0)
     else:
         print(f"  {total_errors} problem(s) across the predictions file.")