Compiler integrity: no more wasted paid calls, real validation on shipped packets, v14 - #132
Merged
Merged
Conversation
…ating what it ships
Integrity scan (compiler/sutta-studio slice), all fixes red-proven:
- P1: a FIFTH uncached LLM call per phase — a fully-cached recompile still
billed one large call per phase to buy a title, and that call's failure
discarded four successful billed passes into a degraded view. The
phaseView call now runs only when a pass output is actually missing;
its failure degrades the TITLE, not the phase. Cost comment counts
honestly. Red-proof: full-cache compile = 0 LLM calls (pre-fix: N).
- P1: the "RIPPLES — CRITICAL" feature was structurally forbidden by the
schema sent on the same request (prompt/type/renderer/example all use a
token-id map; the schema demanded an array of {tradition, rendering}
under additionalProperties:false — under strict mode the model COULD
NOT comply). Both schema blocks now use the map shape; the OpenAI
strict transform drops open-map properties with a disclosure (the
dialect cannot express them); the worked example's impossible key
fixed. SUTTA_STUDIO_PROMPT_VERSION → v14 (v13's missing history entry
backfilled) so caches invalidate and future benchmark runs are marked
incomparable.
- P1: production-compiled packets were invisible to the rich validator —
the compiler never set canonicalSegmentIds, and every deep check keys
on it (vacuous pass on every compiled packet; the ledgered
validator-weaker-than-the-system class). rehydratePhase and the
degraded builder now populate it from sourceSpan; a new ERROR fires
when a non-degraded phase has neither. Red-proof: 7/8 tests fail
pre-fix, including a pali-corruption case the vacuous check missed.
- P2: the weak validatePacket renamed validatePacketIds; the RICH
validator now runs (log-only) at compile end; validator version stamps
unified ('1.0.0', was a fictional 'v1').
- P2: repair layers now agree — validatePhase delegates to the render
backstop repairEnglishStructure, so dangling-linked tokens are dropped
instead of surviving as undetectable empty pills.
- P2: TRANSLATOR_DEBATE no longer instructs per-sense fields the schema
forbids; the retired sourceCitationIds wire-back docstrings now name
citationIds honestly.
- P3: the false "dead skeleton.ts" comment fixed (it IS production, the
"canonical" passes are not — all six now say so); zombie
morphologyCache deleted; packet.compiler.provider recorded honestly;
final progress no longer counts skipped/degraded phases as ready; the
OpenAI strict-schema transform MOVED to services/ai and applied in the
production adapter's chatJSON (the benchmark keeps its import via a
shim) — production is now immunized against the 400 class the
benchmark already survived.
tsc clean; full suite green; build green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0117Ksuo6oeSBw4iTq3yzfYi
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
…ts) + pilot-scoped rich validation readySegments now counts unique canonical ids landing in non-degraded phases (per-phase += overshot totalSegments when wordRange split one segment across phases); rich validation now runs against the segments the retained skeleton actually covers, so pilot mode (phaseLimit) no longer flags every deliberately-uncompiled segment as missing. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0117Ksuo6oeSBw4iTq3yzfYi
Owner
Author
|
Codex findings adjudicated — both real, both fixed: readySegments counts unique canonical ids (wordRange splits no longer overshoot totalSegments) and rich validation is scoped to the retained skeleton's covered segments (pilot mode no longer self-flags). Compiler + sutta-studio suites green. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Compiler-slice integrity fixes, all red-proven: the fifth uncached billed call per phase (fully-cached recompiles billed N calls for N titles; failures destroyed 4 successful paid passes), the ripples schema/prompt self-contradiction (feature impossible under strict mode — now map-shaped, v14 prompt bump, OpenAI-dialect degradation disclosed), canonicalSegmentIds so the rich validator actually validates compiled packets (was vacuous), repair-layer agreement, retired-field prompt cleanup, and the strict-schema transform moved into the production adapter. Full suite green.
🤖 Generated with Claude Code
https://claude.ai/code/session_0117Ksuo6oeSBw4iTq3yzfYi