cleanup(enhancements): move retrieval_fusion, flip Beta-LB gate, add invariant + obfuscation tests#163
cleanup(enhancements): move retrieval_fusion, flip Beta-LB gate, add invariant + obfuscation tests#163
Conversation
…B gate, add invariant + obfuscation tests Council v4 verdict (council_2026-05-02T11-59-00.md) and v4-rerun (council_2026-05-02T11-59-00.md) flagged a small set of production-readiness items that don't depend on the larger dual-write work. This commit lands those independently so dual-write atomicity can ship as its own reviewable PR. What - Move src/gradata/enhancements/retrieval_fusion.py into enhancements/scoring/retrieval_fusion.py and update importers. Council vote 5/7 — RRF is a ranking primitive, lives more naturally with scoring/ than as a sibling. - Flip GRADATA_BETA_LB_GATE default ON in enhancements/self_improvement/_graduation.py. The 2026-04 ablation documented in the file showed ~15-20% of RULE-tier graduations miscalibrated by format-not-content; shipping the fix opt-in was textbook silent regression (council 5/7). GRADATA_BETA_LB_GATE=0 preserves the override-off escape hatch. - New tests/test_initial_confidence_invariant.py — locks the INITIAL_CONFIDENCE / PATTERN_THRESHOLD = 0.60 boundary that almost promoted every fresh lesson before strict-> was wired in. - New tests/test_score_obfuscation_gate.py — CI gate that fails the build if any raw confidence float in [0,1] leaks into the <brain-rules> prompt-bound payload. middleware/_core.py build_brain_rules_block() updated to obfuscate. Why - Each item is independently testable, low-risk, and clears the runway for the dual-write atomicity PR. - Beta-LB default-on closes a known correctness hole that ships every release until flipped. - Obfuscation gate converts a comment-level guarantee (security/score_obfuscation.py) into an enforced one. Test plan - pytest tests/test_initial_confidence_invariant.py tests/test_score_obfuscation_gate.py tests/test_retrieval_fusion.py tests/test_rule_pipeline.py tests/test_middleware_core.py — 63 passed. - pyright src/ — 0 errors, 27 warnings (unchanged baseline). - ruff on changed files — clean. Layering check - No Layer 0 → 2 imports introduced. Risk - Beta-LB flip changes the default for graduation calibration. Users relying on miscalibrated PATTERN→RULE behavior will see fewer graduations until they set GRADATA_BETA_LB_GATE=0. This is the intended fix. Council references - council_2026-05-02T11-08-25.md (initial v4 review) - council_2026-05-02T11-59-00.md (v4 rerun, all 7 lenses through fallback chain) - council_2026-05-02T12-24-08.md (PR sequencing decision)
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
There was a problem hiding this comment.
Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.
📝 WalkthroughSummary
WalkthroughThis PR reorganizes the retrieval fusion module structure, inverts the GRADATA_BETA_LB_GATE environment variable default from opt-in to enabled-by-default, obfuscates confidence values in brain rules output, and updates corresponding test fixtures and assertions. ChangesModule Reorganization & Output Obfuscation
Beta LB Gate Default Behavior Inversion
Sequence Diagram(s)Not applicable—these changes constitute module relocation, logic inversion, and obfuscation rather than introducing new multi-component interactions or significantly altering control flow. Estimated code review effort🎯 3 (Moderate) | ⏱️ ~20 minutes Possibly related PRs
Suggested labels
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Review rate limit: 4/5 reviews remaining, refill in 12 minutes. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 3
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@Gradata/src/gradata/enhancements/self_improvement/_graduation.py`:
- Around line 110-115: The default-on gate (GRADATA_BETA_LB_GATE currently
defaulting to "1") causes lessons missing persisted alpha/beta_param to be
treated as Beta(1,1) in _passes_beta_lb_gate(), which denies legacy PATTERN
lessons; revert the compatibility break by making the gate default off (change
the default of GRADATA_BETA_LB_GATE to "0"/false) or alter
_passes_beta_lb_gate() to treat a Lesson lacking alpha or beta_param as passing
(i.e., skip the Beta(1,1) fallback and allow graduation) — update code
references to GRADATA_BETA_LB_GATE, _passes_beta_lb_gate(), Lesson.alpha, and
Lesson.beta_param accordingly.
In `@Gradata/tests/test_score_obfuscation_gate.py`:
- Around line 16-23: Replace the direct Brain.init(...) invocation in
test_apply_brain_rules_prompt_does_not_leak_raw_confidence with the test helpers
from conftest: either call brain = init_brain(tmp_path, name="ObfuscationGate",
domain="Testing", embedding="local", interactive=False) or switch the test to
use the fresh_brain fixture and adjust its parameters; this ensures BRAIN_DIR
and the _paths cache are handled the same way as other tests instead of calling
Brain.init directly.
- Around line 8-13: The regex _RAW_CONFIDENCE_FLOAT currently also matches
integers because the decimal portion is optional; update the pattern used by the
_RAW_CONFIDENCE_FLOAT constant so it requires an explicit decimal point and
digits (i.e., only match floats like 0.95 or 1.00), then keep
_assert_no_raw_confidence_float unchanged so it will only detect actual float
leaks and not plain integers like "1" or "0".
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: ASSERTIVE
Plan: Pro
Run ID: d6f52f71-6517-430a-b96c-70529a21ae9f
📒 Files selected for processing (14)
Gradata/src/gradata/enhancements/rule_pipeline.pyGradata/src/gradata/enhancements/scoring/retrieval_fusion.pyGradata/src/gradata/enhancements/self_improvement/_graduation.pyGradata/src/gradata/middleware/_core.pyGradata/tests/test_initial_confidence_invariant.pyGradata/tests/test_middleware_core.pyGradata/tests/test_retrieval_fusion.pyGradata/tests/test_rule_graduated_events.pyGradata/tests/test_rule_pipeline.pyGradata/tests/test_rule_to_hook.pyGradata/tests/test_rule_to_hook_promotion.pyGradata/tests/test_safety_assertion.pyGradata/tests/test_score_obfuscation_gate.pyGradata/tests/test_wiring_compound.py
📜 Review details
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (7)
- GitHub Check: pytest ubuntu-latest / py3.11
- GitHub Check: pytest windows-latest / py3.12
- GitHub Check: pytest macos-latest / py3.11
- GitHub Check: pytest macos-latest / py3.12
- GitHub Check: pytest windows-latest / py3.11
- GitHub Check: pytest ubuntu-latest / py3.12
- GitHub Check: pytest (py3.12)
🧰 Additional context used
📓 Path-based instructions (2)
Gradata/tests/**/*.py
📄 CodeRabbit inference engine (Gradata/AGENTS.md)
Gradata/tests/**/*.py: SetBRAIN_DIRenvironment variable viatmp_pathin conftest.py for test isolation — ensure_paths.pymodule cache refreshes when callingBrain.init()directly inside tests
Add unit tests intests/test_*.pyfor every CI push without LLM calls (deterministic); mark integration tests with@pytest.mark.integrationand skip them by default (they hit real LLM APIs)
Files:
Gradata/tests/test_retrieval_fusion.pyGradata/tests/test_middleware_core.pyGradata/tests/test_rule_to_hook.pyGradata/tests/test_initial_confidence_invariant.pyGradata/tests/test_rule_graduated_events.pyGradata/tests/test_wiring_compound.pyGradata/tests/test_score_obfuscation_gate.pyGradata/tests/test_rule_to_hook_promotion.pyGradata/tests/test_safety_assertion.pyGradata/tests/test_rule_pipeline.py
Gradata/src/**/*.py
📄 CodeRabbit inference engine (Gradata/AGENTS.md)
Gradata/src/**/*.py: Prefersentence-transformersfor local embeddings,google-genaifor Gemini embeddings,cryptographyfor AES-GCM encrypted system.db,bm25sfor BM25 rule ranking, andmem0aifor external memory adapters — guard all optional dependency imports withtry / except ImportErrorat the call site, never at module level
Maintain strict layering: Layer 0 (Primitives: _types.py, _db.py, _events.py, _paths.py, _file_lock.py; Patterns: contrib/patterns/) must never import from Layer 1 (Enhancements: enhancements/, rules/) or Layer 2 (Public API: brain.py, cli.py, daemon.py, mcp_server.py)
Never use bareexcept: pass— use typed exceptions or at minimumlogger.warning(...)withexc_info=Trueto avoid silent failure in a memory product
Never import from out-of-scope sibling directories../Sprites/or../Hausgem/withingradata/*code — that is a layering bug
Never leak private-sibling paths into public docs/code — no references to../Sprites/,../Hausgem/, email addresses, OneDrive paths, or Sprites-specific examples from insidegradata/*
Use atomic-write helper when writing JSON files to prevent corruption from mid-write crashes
Files:
Gradata/src/gradata/enhancements/rule_pipeline.pyGradata/src/gradata/middleware/_core.pyGradata/src/gradata/enhancements/self_improvement/_graduation.py
🔇 Additional comments (8)
Gradata/tests/test_initial_confidence_invariant.py (2)
12-20: Helper fixture is minimal and correct for graduation-path tests.Good construction of a deterministic
Lessonobject with only the fields needed for this invariant.
23-45: Boundary assertions for INSTINCT→PATTERN promotion are well covered.The tests correctly lock the tie case (
==threshold) as non-promoting and the above-threshold case as promoting, matching the strict comparison contract.Gradata/src/gradata/middleware/_core.py (1)
280-283: Obfuscation is applied at the correct boundary.Good change: Line 281 strips score suffixes at render time while preserving rule selection and XML structure.
Gradata/tests/test_middleware_core.py (2)
46-47: Assertion update matches new obfuscated marker format.These checks correctly enforce
[RULE]presence and confidence-suffixed marker absence.
59-59: Max-rules assertion stays aligned with obfuscated output.Counting
[RULE]here is the right adjustment for the new rendered format.Gradata/tests/test_score_obfuscation_gate.py (1)
38-60: Great middleware-level gate coverage.This test complements the core unit checks by asserting obfuscation on
build_brain_rules_block(...)output directly.Gradata/tests/test_rule_pipeline.py (2)
171-180: Gate override keeps this graduation test deterministic.Good update: disabling
GRADATA_BETA_LB_GATEin-test prevents default-flip drift and preserves intended threshold behavior coverage.
246-247: Missing-module mock now targets the correct relocated import.This aligns the test with the new
enhancements.scoring.retrieval_fusionpath and keeps optional-dependency failure handling properly covered.
| enabled = os.environ.get("GRADATA_BETA_LB_GATE", "1").lower() not in ( | ||
| "0", | ||
| "false", | ||
| "no", | ||
| "off", | ||
| ) |
There was a problem hiding this comment.
Default-on Beta-LB now hard-blocks lessons that lack posterior state.
Enabling the gate by default turns any Lesson without persisted alpha / beta_param into an automatic PATTERN→RULE deny, because _passes_beta_lb_gate() falls back to Beta(1,1), whose 5th-percentile lower bound is far below the default 0.85. That means legacy or handcrafted PATTERN lessons will stop graduating entirely unless the env override is set, which is a much stronger behavior change than “tighter calibration.”
Possible compatibility guard
def _passes_beta_lb_gate(
lesson: Lesson,
config: tuple[bool, float, int] | None = None,
) -> bool:
@@
- alpha = getattr(lesson, "alpha", 1.0)
- beta_param = getattr(lesson, "beta_param", 1.0)
+ alpha = getattr(lesson, "alpha", None)
+ beta_param = getattr(lesson, "beta_param", None)
+ if alpha is None or beta_param is None:
+ return True # keep legacy promotion behavior until posterior state is backfilled
+
from gradata.rules.rule_engine import _beta_ppf_05
return _beta_ppf_05(alpha, beta_param) >= thresholdAlso applies to: 151-155
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@Gradata/src/gradata/enhancements/self_improvement/_graduation.py` around
lines 110 - 115, The default-on gate (GRADATA_BETA_LB_GATE currently defaulting
to "1") causes lessons missing persisted alpha/beta_param to be treated as
Beta(1,1) in _passes_beta_lb_gate(), which denies legacy PATTERN lessons; revert
the compatibility break by making the gate default off (change the default of
GRADATA_BETA_LB_GATE to "0"/false) or alter _passes_beta_lb_gate() to treat a
Lesson lacking alpha or beta_param as passing (i.e., skip the Beta(1,1) fallback
and allow graduation) — update code references to GRADATA_BETA_LB_GATE,
_passes_beta_lb_gate(), Lesson.alpha, and Lesson.beta_param accordingly.
| _RAW_CONFIDENCE_FLOAT = re.compile(r"(?<![\w.])(?:0(?:\.\d+)?|1(?:\.0+)?)(?![\w.])") | ||
|
|
||
|
|
||
| def _assert_no_raw_confidence_float(prompt: str) -> None: | ||
| leaks = _RAW_CONFIDENCE_FLOAT.findall(prompt) | ||
| assert not leaks, f"raw confidence float leaked into prompt-bound text: {prompt}" |
There was a problem hiding this comment.
🧩 Analysis chain
🏁 Script executed:
#!/bin/bash
set -euo pipefail
python - <<'PY'
import re
current = re.compile(r"(?<![\w.])(?:0(?:\.\d+)?|1(?:\.0+)?)(?![\w.])")
proposed = re.compile(r"(?<![\w.])(?:0\.\d+|1\.0+)(?![\w.])")
samples = [
"<brain-rules>\n[RULE] Use 1 sentence.\n</brain-rules>",
"<brain-rules>\n[RULE] Keep 0 padding.\n</brain-rules>",
"<brain-rules>\n[RULE:0.95] Prefer concrete dates.\n</brain-rules>",
"<brain-rules>\n[RULE:1.00] Prefer concrete dates.\n</brain-rules>",
]
for s in samples:
print("TEXT:", s.replace("\n", "\\n"))
print(" current:", current.findall(s))
print("proposed:", proposed.findall(s))
print("-" * 60)
PYRepository: Gradata/gradata
Length of output: 701
Regex matches plain integers and will cause false positives.
The optional decimal portion in the current pattern ((?:\.\d+)? and (?:\.0+)?) makes it match standalone 0 and 1. This means valid text like "use 1 sentence" or "keep 0 padding" will incorrectly trigger the gate.
Proposed fix
-_RAW_CONFIDENCE_FLOAT = re.compile(r"(?<![\w.])(?:0(?:\.\d+)?|1(?:\.0+)?)(?![\w.])")
+_RAW_CONFIDENCE_FLOAT = re.compile(r"(?<![\w.])(?:0\.\d+|1\.0+)(?![\w.])")The proposed regex requires an explicit decimal point and digits, matching only actual floats like 0.95 and 1.00 while ignoring plain integers.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@Gradata/tests/test_score_obfuscation_gate.py` around lines 8 - 13, The regex
_RAW_CONFIDENCE_FLOAT currently also matches integers because the decimal
portion is optional; update the pattern used by the _RAW_CONFIDENCE_FLOAT
constant so it requires an explicit decimal point and digits (i.e., only match
floats like 0.95 or 1.00), then keep _assert_no_raw_confidence_float unchanged
so it will only detect actual float leaks and not plain integers like "1" or
"0".
| def test_apply_brain_rules_prompt_does_not_leak_raw_confidence(tmp_path) -> None: | ||
| brain = Brain.init( | ||
| tmp_path / "brain", | ||
| name="ObfuscationGate", | ||
| domain="Testing", | ||
| embedding="local", | ||
| interactive=False, | ||
| ) |
There was a problem hiding this comment.
🧩 Analysis chain
🏁 Script executed:
#!/bin/bash
set -euo pipefail
echo "Looking for test conftest and isolation hooks..."
fd -i 'conftest.py' tests
echo
echo "Searching for BRAIN_DIR and _paths cache refresh logic in test infra..."
rg -n -C3 'BRAIN_DIR|_paths|Brain\.init|reload|refresh' tests
echo
echo "Reference call site in this test file:"
rg -n -C2 'test_apply_brain_rules_prompt_does_not_leak_raw_confidence|Brain\.init\(' Gradata/tests/test_score_obfuscation_gate.pyRepository: Gradata/gradata
Length of output: 219
🏁 Script executed:
#!/bin/bash
set -euo pipefail
# Find all conftest.py files in the repository
echo "=== Locating conftest.py files ==="
git ls-files | grep -i conftest.py
echo
echo "=== Checking Gradata/tests directory structure ==="
git ls-files | grep "^Gradata/tests/" | head -20
echo
echo "=== Looking for the target test file ==="
git ls-files | grep "test_score_obfuscation_gate.py"Repository: Gradata/gradata
Length of output: 953
🏁 Script executed:
#!/bin/bash
set -euo pipefail
echo "=== Reading conftest.py for test isolation setup ==="
cat -n Gradata/tests/conftest.py
echo
echo "=== Checking for test_score_obfuscation_gate.py ==="
git ls-files | grep "test_score_obfuscation_gate.py"
echo
echo "=== Reading test_score_obfuscation_gate.py around lines 16-23 ==="
head -30 Gradata/tests/test_score_obfuscation_gate.pyRepository: Gradata/gradata
Length of output: 9867
Use the fresh_brain fixture or init_brain() from conftest.py instead of calling Brain.init() directly.
This test calls Brain.init() directly without setting BRAIN_DIR or refreshing the _paths.py module cache. The conftest.py provides the init_brain() helper (and fresh_brain fixture) specifically for this purpose. Either replace the direct Brain.init() call with brain = init_brain(tmp_path, name="ObfuscationGate", domain="Testing") or use the fresh_brain fixture and customize it as needed. See conftest.py lines 26-92.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@Gradata/tests/test_score_obfuscation_gate.py` around lines 16 - 23, Replace
the direct Brain.init(...) invocation in
test_apply_brain_rules_prompt_does_not_leak_raw_confidence with the test helpers
from conftest: either call brain = init_brain(tmp_path, name="ObfuscationGate",
domain="Testing", embedding="local", interactive=False) or switch the test to
use the fresh_brain fixture and adjust its parameters; this ensures BRAIN_DIR
and the _paths cache are handled the same way as other tests instead of calling
Brain.init directly.
* cleanup(enhancements): move retrieval_fusion to scoring/, flip Beta-LB gate, add invariant + obfuscation tests Council v4 verdict (council_2026-05-02T11-59-00.md) and v4-rerun (council_2026-05-02T11-59-00.md) flagged a small set of production-readiness items that don't depend on the larger dual-write work. This commit lands those independently so dual-write atomicity can ship as its own reviewable PR. What - Move src/gradata/enhancements/retrieval_fusion.py into enhancements/scoring/retrieval_fusion.py and update importers. Council vote 5/7 — RRF is a ranking primitive, lives more naturally with scoring/ than as a sibling. - Flip GRADATA_BETA_LB_GATE default ON in enhancements/self_improvement/_graduation.py. The 2026-04 ablation documented in the file showed ~15-20% of RULE-tier graduations miscalibrated by format-not-content; shipping the fix opt-in was textbook silent regression (council 5/7). GRADATA_BETA_LB_GATE=0 preserves the override-off escape hatch. - New tests/test_initial_confidence_invariant.py — locks the INITIAL_CONFIDENCE / PATTERN_THRESHOLD = 0.60 boundary that almost promoted every fresh lesson before strict-> was wired in. - New tests/test_score_obfuscation_gate.py — CI gate that fails the build if any raw confidence float in [0,1] leaks into the <brain-rules> prompt-bound payload. middleware/_core.py build_brain_rules_block() updated to obfuscate. Why - Each item is independently testable, low-risk, and clears the runway for the dual-write atomicity PR. - Beta-LB default-on closes a known correctness hole that ships every release until flipped. - Obfuscation gate converts a comment-level guarantee (security/score_obfuscation.py) into an enforced one. Test plan - pytest tests/test_initial_confidence_invariant.py tests/test_score_obfuscation_gate.py tests/test_retrieval_fusion.py tests/test_rule_pipeline.py tests/test_middleware_core.py — 63 passed. - pyright src/ — 0 errors, 27 warnings (unchanged baseline). - ruff on changed files — clean. Layering check - No Layer 0 → 2 imports introduced. Risk - Beta-LB flip changes the default for graduation calibration. Users relying on miscalibrated PATTERN→RULE behavior will see fewer graduations until they set GRADATA_BETA_LB_GATE=0. This is the intended fix. Council references - council_2026-05-02T11-08-25.md (initial v4 review) - council_2026-05-02T11-59-00.md (v4 rerun, all 7 lenses through fallback chain) - council_2026-05-02T12-24-08.md (PR sequencing decision) * fix(events): JSONL canonical, SQLite projection, reconcile-on-init, doctor --reconcile Council v4 (council_2026-05-02T11-59-00.md) ranked dual-write atomicity the #1 production blocker. Crash mid-write between events.jsonl append and SQLite INSERT could leave the brain in silent split-brain state with no recovery path. What - src/gradata/_events.py - JSONL is the canonical source of truth. Append + fsync FIRST, SQLite INSERT is now an idempotent projection derived from JSONL. - Added reconcile_jsonl_to_sqlite() that scans JSONL past the SQLite watermark and replays missing rows. - Single SQLite projection helper used by both the live write path and the retain orchestrator. - Env-gated crash-window delay for deterministic kill-9 testing only (no production effect). - src/gradata/brain.py - Brain.__init__ runs JSONL → SQLite reconciliation after migrations. - Brain() resolves BRAIN_DIR / cwd when no explicit path is supplied. - observe(text, kind="correction") public event API used by the PR2 spec. - src/gradata/cli.py + src/gradata/_doctor.py - New `gradata doctor --reconcile`: scans for drift, reports the count, replays missing JSONL rows into SQLite, exits non-zero on inconsistency that can't be healed. - tests/test_dualwrite_atomicity.py - Path-agnostic public-API tests covering: happy path, kill-9 mid batch (JSONL must lead SQLite, never trail), reconcile replay, idempotency, doctor CLI drift report, concurrent-writer JSONL line integrity. Why - Before: dual-write claimed atomic in CLAUDE.md, no two-phase commit, no recovery. Crash → silent data loss or duplicate-on-replay. - After: JSONL is the log, SQLite is the projection. Every reopen reconciles. doctor --reconcile is the operator escape hatch. Property: jsonl_count >= sqlite_count, always. Test plan - pytest tests/test_dualwrite_atomicity.py — 6 passed. - Full focused regression on changed surface — 42 passed. - Non-integration suite (excluding socket-bound daemon/plugin tests blocked by sandbox) — 4130 passed, 4 skipped. - pyright src/ — 0 errors, 27 warnings (unchanged baseline). Layering check - _events.py is Layer 0. Brain.__init__ in Layer 2 calls into it. No upward imports introduced. Risk - Reconcile-on-init runs on every Brain open. For a brain with 100k events this adds ~50ms-200ms one-time at startup. Watermark is incremental so subsequent opens are O(drift) not O(total). - Concurrent writers serialize via JSONL append + advisory lock. Throughput trade-off is acceptable for correctness. Council references - council_2026-05-02T11-59-00.md (v4 RISK class, all 7 lenses) - council_2026-05-02T12-24-08.md (PR sequencing — TDD-first) Stacks on #163. * fix(cloud/client): coerce float ts and non-int session for /sync The backfill script and incremental sync both crashed on real-world events.jsonl rows that contain: - float ts (epoch seconds, e.g. 1776803751.89) — broke str-vs-str comparison against the watermark cursor with TypeError. - float or string session values (e.g. 4.5, UUID strings) — server schema rejects non-ints with HTTP 422. Coerce ts to str and session to int|None at the format-event boundary. Also surface the response body in HTTPError so 4xx/5xx debugging is not opaque. Discovered while running scripts/backfill_to_cloud.py against a brain with ~28k events accumulated since 2026-03-22. --------- Co-authored-by: Oliver Le <oliverle94@gmail.com>
2 participants
Summary
Council-validated cleanup work, split out so the larger dual-write PR can be reviewed independently.
retrieval_fusion.pyunderenhancements/scoring/(council 5/7 — RRF is a ranking primitive).GRADATA_BETA_LB_GATEdefault ON. Documented 2026-04 ablation showed ~15-20% of RULE-tier graduations miscalibrated by format-not-content; shipping the fix opt-in was the textbook silent-regression pattern.GRADATA_BETA_LB_GATE=0preserves the override.test_initial_confidence_invariant.pylocks the INITIAL_CONFIDENCE / PATTERN_THRESHOLD = 0.60 boundary.test_score_obfuscation_gate.pyis a CI gate that fails the build if any raw confidence float in [0,1] leaks into the<brain-rules>prompt.middleware/_core.pyupdated to obfuscate.Test plan
pytest tests/test_initial_confidence_invariant.py tests/test_score_obfuscation_gate.py tests/test_retrieval_fusion.py tests/test_rule_pipeline.py tests/test_middleware_core.py— 63 passed.pyright src/— 0 errors, 27 warnings (unchanged baseline).Layering check
No Layer 0 → 2 imports introduced.
Risk
Beta-LB flip changes default graduation calibration. Users relying on the miscalibrated path will see fewer PATTERN→RULE promotions until they set
GRADATA_BETA_LB_GATE=0. Intended.Council references
council_2026-05-02T11-08-25.mdcouncil_2026-05-02T11-59-00.md(all 7 lenses via fallback chain)council_2026-05-02T12-24-08.md(PR sequencing)