You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardExpand all lines: BITCODE_SPEC_V41.md
+3Lines changed: 3 additions & 0 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -123,6 +123,9 @@ The current source-safe contract covers 12 rows, 65 passing source predicates, P
123
123
124
124
Gate 4 must establish benchmark baselines for Reading prompt surfaces before rewrites.
125
125
It covers `ReadNeedComprehensionSynthesis`, `ReadFitsFindingSynthesis`, Need review, Finding Fits, depository query synthesis, many-candidate ranking, AssetPack preview, quote explanation, telemetry summaries, and failure/repair prompts.
126
+
Gate 4 is closed by the package-backed `buildV41ReadingPromptBenchmarkBaselines` source, deterministic `.bitcode/v41-reading-prompt-benchmark-baselines.json` artifact, `generate:v41-reading-prompt-benchmark-baselines`, `check:v41-reading-prompt-benchmark-baselines`, `check:v41-gate4`, protocol tests, and gate/canon workflow wiring.
127
+
The current source-safe baseline covers 10 rows, 120 passing source predicates, all five Reading UX steps, both Reading pipeline ids, Gate 2 PromptPart/Prompt inventory roots, Gate 3 registry/interpolation roots, V38 benchmark fixture roots, V40 prompt smoke roots, 36 parser target ids, 9 registry contract ids, deterministic baseline scores, and source-safe disclosure tiers.
128
+
The artifact may include ids, hashes, counts, fixture ids, metric ids, parser target ids, baseline scores, source path roots, and verdicts, but no raw prompt text, interpolated prompts, raw provider responses, protected source, private context, credentials, or unpaid AssetPack source.
126
129
127
130
## V41 Gate 5 ReadNeedComprehensionSynthesis Prompt Rewrite And Return-Type Hardening
- Source parity state: V41 source parity now includes prompt inventory plus registry/interpolation contracts; prompt rewrites are explicitly deferred until catalogue, contract, and benchmark gates close
- Source parity state: V41 source parity now includes prompt inventory, registry/interpolation contracts, and Reading benchmark baselines; prompt rewrites are explicitly deferred until catalogue, contract, and benchmark gates close
12
12
- Scope: V41 draft delta for Prompt and PromptPart excellence over promoted V40
13
13
14
14
## Why V41 exists
@@ -45,6 +45,8 @@ It preserves V42 as subsequent reliable MVP experience planning while keeping V4
- Source parity state: V41 notes track prompt-program planning, catalogue, benchmark, rewrite, telemetry, and promotion parity over active V40
13
13
- Draft target posture: V41 Gate 1 creates the full SPEC, DELTA, NOTES, and PARITY family.
14
14
- Scope: Prompt and PromptPart excellence over the promoted V40 exhaustive testing base.
@@ -39,6 +39,12 @@ Gate 3 makes the prompt registry call chain source-checkable before prompt rewri
39
39
The package-backed contract currently emits 12 rows and 65 source predicates for Prompt registry totality, TemplatedPromptPart interpolation, PTRR agent and step prompt composition, FailsafeGenerationSequence context handling, ThricifiedGeneration final prompt resolution, execution ancestry overlays, tool doc-code prompt injection, Reading parser targets, Finding Fits search and embedding contracts, AssetPack synthesis and finishing parser targets, and Gate 2 inventory binding.
40
40
The artifact records ids, source paths, hashes, predicate verdicts, parser target ids, interpolation key ids, and ancestry frame ids only; it does not serialize raw prompt text, provider responses, private context, credentials, or unpaid AssetPack source.
41
41
42
+
## Gate 4 Reading benchmark baseline note
43
+
44
+
Gate 4 makes Reading prompt quality source-checkable before prompt rewrites.
45
+
The package-backed baseline currently emits 10 rows and 120 source predicates for Read Request-to-Need synthesis, Need review/resynthesis, Need measurement and quote posture, Finding Fits query synthesis, many-candidate ranking, AssetPack synthesis context, source-safe AssetPack preview, settlement/rights/delivery, telemetry summaries, and failure/repair prompts.
46
+
The artifact binds both Reading pipelines, all five Reading UX steps, Gate 2 prompt inventory, Gate 3 registry/interpolation contracts, V38 benchmark fixtures, V40 prompt smoke readiness, parser target ids, and source-safe score metadata only; it does not serialize raw prompt text, interpolated prompts, provider responses, private context, protected source, credentials, or unpaid AssetPack source.
It may not serialize raw prompt text, raw provider responses, protected source, private context, credentials, settlement-private payloads, or unpaid AssetPack source.
72
72
It closes prompt rewrites' precondition that prompt composition and interpolation contracts are observable before semantic PromptPart or Prompt rewrite work proceeds.
73
73
74
+
V41 Gate 4 may catalogue Reading benchmark baseline ids, fixture ids, metric ids, parser target ids, source path roots, predicate verdicts, deterministic baseline scores, inventory bindings, registry/interpolation bindings, and source-safe disclosure tiers for `ReadNeedComprehensionSynthesis` and `ReadFitsFindingSynthesis`.
75
+
It may not serialize raw prompt text, interpolated prompts, provider responses, protected source, private context, credentials, settlement-private payloads, or unpaid AssetPack source.
76
+
It closes Reading prompt rewrites' benchmark precondition by proving source-safe baselines for all five Reading UX steps before semantic PromptPart or Prompt rewrite work starts.
77
+
74
78
## V41 completion condition
75
79
76
80
V41 closes when prompt-program quality is fully specified, implemented, benchmarked, source-safe, tested, documented, generated, workflow-checked, and promotion-ready across Reading, Conversation, tool-definition, and interface inference surfaces.
0 commit comments