Validate final routing policies and harden artifact loading

CHANGELOG.md

@@ -5,7 +5,29 @@ versioning.
 
 ## Unreleased
 
+These entries describe repository source. They do not announce a new PyPI or
+npm release; check the installed package version before relying on them.
+
+### Documentation
+- Document direct student predictions, teacher deferral and the OOD guard.
+- Add an [overview](docs/system1.md) covering the teacher/student contract,
+  meaningful quality and cost measurements, free self-hosting, and the boundary
+  between this library and the hosted application's managed features.
+- Clarify that OSS OOD rejection is currently reported as `deferred`, that
+  acceptance scores are not certified per-request probabilities, and that
+  fixed-label fitting does not train a general zero-shot model or its encoder.
+- Correct stale integration examples and unsupported latency, savings and
+  monotonic-coverage claims. Keep historical paper results identified as such.
+
 ### Fixed
+- Make direct `fit()` transactional as well as `update()`: partial reports,
+  indexes, or manifests cannot replace a previously working generation.
+- Fail loading when a present or required OOD guard is unreadable or missing;
+  preserve legacy unguarded artifacts without inventing a certificate.
+- Reserve final certification rows before fitting, label discovery or balancing.
+  Check the selected serving policy, including its development-only OOD guard,
+  once using a one-sided agreement lower bound; never retry other candidates
+  using these outcomes. Zero accepted examples cannot certify a policy.
 - Preserve the previous artifact generation when an update fails validation,
   fitting, or publication. Honor explicit update configuration without mutating
   the caller, retain saved configuration by default, and increment the fit count.
@@ -26,6 +48,15 @@ versioning.
 - Correct update examples to use `new_embeddings=`.
 
 ### Changed (breaking)
+- New fits reserve 20% of rows by default (`certification_fraction`) and use
+  `certification_alpha=0.10`. Small or uncertain fits may now decline deployment.
+  `manifest.certification` records the result; legacy-named `coverage_cal` and
+  `teacher_agreement_cal` now measure this final partition for new artifacts.
+  The confidence statement requires independent representative examples and
+  does not cover distribution shift, correlated sessions, or adaptive reuse.
+- Persist `ood_reference.npy` separately from all embeddings used for updates.
+  Existing artifacts keep their previous behavior unless their guard is broken.
+- Remove unsupported universal coverage, savings and monotonic-growth claims.
 - The prediction server binds to `127.0.0.1` by default and no longer sends a
   wildcard CORS header. Explicitly configure a remote bind address when needed;
   browser deployments should configure CORS at their authenticated proxy.
@@ -85,6 +116,10 @@ versioning.
 
 ## 0.2.0 (2026-06)
 
+Historical release notes. The scan remains a diagnostic; the calibration
+procedure described here is superseded by the independent final-policy check
+in Unreleased. These entries do not establish current deployment guarantees.
+
 ### Added
 - `tracer.scan()`: a fast, conservative day-one read of a traces file, before any
   training. It groups traffic by similarity and measures, on a held-out slice it
@@ -106,11 +141,11 @@ versioning.
   `embeddings=`, or select a local sentence-transformers model with `model=`.
 
 ### Changed
-- The parity gate now certifies on an exact held-out lower bound instead of an
-  in-sample point estimate, so a policy cannot clear the target by in-sample luck
-  and then break the contract on real traffic. Coverage is now monotonic in the
-  target, and a hybrid select-then-verify procedure recovers coverage at strict
-  targets that a plain held-out split discarded.
+- The parity gate introduced held-out lower-bound checks and a hybrid
+  select-then-verify procedure. Later review found that selection and final
+  certification must use separate data; see Unreleased. This historical
+  procedure does not guarantee agreement on future traffic or monotonic
+  coverage across targets.
 - The HTML report is restyled to the light Tracer theme, and the word "audit" is
   dropped across the report and docs.
 

CONTRIBUTING.md

@@ -2,12 +2,13 @@
 
 Thank you for your interest in contributing.
 
-TRACER is a routing library with formal parity guarantees, and those guarantees
-only mean something if they hold up under measurement. So for anything that can
-change routing behavior, the most useful thing you can bring is evidence. The
-standards below exist to keep contributions focused and to make the best use of
-your time and ours. None of this is meant to gatekeep good ideas. it's meant to
-help us evaluate them fairly.
+TRACER builds task-specific classifiers: classical or
+shallow neural classifiers make direct fixed-label decisions, while a learned
+policy leaves other requests to the caller's teacher. The final check bounds
+teacher agreement under stated sampling assumptions. It does not guarantee
+ground-truth accuracy, per-request correctness or robustness to distribution
+shift. Contributions should preserve those distinctions and provide evidence
+for changes in behavior.
 
 ## Contribution standards (read first)
 
@@ -24,8 +25,8 @@ for us to act on when they come with numbers. Three guidelines follow from that.
 
 2. **Changing a core parameter, threshold, or gate: bring an ablation.** The
    acceptor threshold, the parity gate, the calibration logic, the default
-   `FitConfig` values, and the split strategy are load-bearing for the parity
-   guarantee. A change to any of them is much easier to accept with before/after
+   `FitConfig` values, and the split strategy determine the statistical contract.
+   A change to any of them is much easier to accept with before/after
    numbers on the eval showing the frontier does not regress (and, ideally,
    improves). If you think a parameter should be user-tunable, the most
    convincing case is one where a different value measurably wins.
@@ -45,14 +46,23 @@ Run against a fixed, public eval so results are reproducible by a maintainer:
 - Use a known dataset (the repo's test fixtures, or a public set such as
   Banking77) and a fixed `seed`.
 - Report the metric that matters: **coverage at the target teacher agreement**
-  (how much traffic the surrogate can take while holding parity), plus the
-  realized teacher agreement on a held-out split (not the calibration split).
+  together with accepted counts, realized agreement, the final lower bound and
+  confidence level. Use an untouched evaluation split, not the examples used
+  for classifier/threshold selection. Where ground truth exists, report its
+  accuracy separately, including per-class or worst-group behavior.
+- Measure full-path latency and cost when claiming efficiency: embeddings,
+  student inference, teacher deferrals and relevant serving overhead. A gate's
+  score is not a calibrated per-request correctness probability.
+- Separate sessions/workflows/time where the dataset permits; random row splits
+  do not establish independence. Preserve historical benchmark results and label
+  new results with their exact implementation and evaluation protocol.
 - Compare against the current behavior on `main`, so the delta is clear.
 - Include the exact command and config you ran so the numbers can be checked.
 
-If a change does not move the frontier (or regresses it), we'll usually hold
-off on merging even when the code itself is clean, since the numbers are what
-we're optimizing for.
+Performance proposals should show their tradeoff. Correctness fixes, truthful
+contracts and reproducible failure handling are valuable even when they reduce
+previously overstated coverage; they do not need an invented benchmark win.
+Never retune a policy against its final certification outcomes to make it pass.
 
 ## Setup
 
@@ -68,6 +78,8 @@ pip install -e ".[dev]"
 ## Running tests
 
 ```bash
+TOKENIZERS_PARALLELISM=false RAYON_NUM_THREADS=1 OMP_NUM_THREADS=1 \
+OPENBLAS_NUM_THREADS=1 MKL_NUM_THREADS=1 VECLIB_MAXIMUM_THREADS=1 \
 pytest tests/ -v
 ```
 
@@ -97,6 +109,8 @@ src/tracer/
   types.py               <- TraceRecord, QualitativeReport, ArtifactManifest, ...
   fit/
     pipeline.py          <- global / L2D / RSB pipeline construction + calibration
+    certification.py     <- untouched final-policy teacher-agreement check
+    ood.py               <- distance guard, included in final serving check
     surrogate.py         <- model zoo (LogReg, SGD, MLP, RF, ET, DT, GBT, XGB) + selection
   analysis/
     qualitative.py       <- XAI report: slices, boundary pairs, examples, deltas
@@ -134,17 +148,30 @@ further. It's usually quick to run, and it's the thing that lets us say yes.
 
 ## Changing the gate, calibration, or default config
 
-These are the parts that back the parity guarantee. A PR here must include an
-ablation: the frontier on `main` versus the frontier with your change, on the
-same eval and seed. State clearly which guarantee you are strengthening and
-show it does not regress coverage or realized agreement.
+These parts define selection and final certification. Performance changes need
+an ablation on the same external evaluation and seed. Contract/correctness fixes
+need focused regressions that demonstrate the broken behavior and its repair.
+Verify that no final outcomes influence training, threshold selection, OOD
+fitting or retries. Keep no-policy results, exceptions and legacy-artifact
+behavior explicit; a completed failed check and a failed write are different.
 
 ## Adding a new pipeline family
 
 Implement a `build_<name>(split, target_ta) -> dict` function in
 `src/tracer/fit/pipeline.py` following the same structure as `build_global`,
 `build_l2d`, and `build_rsb`. Register it in the `builders` dict inside
-`fit_frontier`. Include a benchmark showing where the new family wins.
+`fit_frontier`, preserving the builders' keyword options. Include a benchmark
+showing where the new family wins. The full selected policy, including residual
+stages and OOD checks, must receive the final check and match runtime behavior.
+
+## Documentation and integration boundaries
+
+Check examples against current function signatures and saved artifact fields.
+The bundled HTTP server takes vectors and does not call teachers. Generic
+embedding/export adapters do not imply a turnkey provider integration or
+standards-compliant OTLP transport. The current package exposes Python APIs,
+not a packaged CLI; hosted catalogs, credentials and Echo wallets are separate.
+Document proposed features as proposals rather than existing OSS capabilities.
 
 ## Submitting a PR