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
docs: normalize remaining project naming in legacy docs
Normalize remaining human-facing project-name references in API and research docs.
Scope: documentation-only; no runtime, fixture, validator, workflow, artifact, evidence-index, package, README, test, LLM, embedding, fuzzy, or taxonomy changes.
Copy file name to clipboardExpand all lines: docs/research_positioning.md
+15-15Lines changed: 15 additions & 15 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -2,9 +2,9 @@
2
2
3
3
CompText V7 is the deterministic replay-integrity layer for compressed operational agent traces. It asks whether compact or reconstructed operational state can preserve the evidence, constraints, blockers, dependencies, recovery paths, capability boundaries, and tool-order signals needed to replay a safe operational trajectory after compression. The project is complementary to learned context-compression research, RAG evaluation, vector-memory systems, serving-layer cache optimization, and durable workflow infrastructure, but it does not replace those systems or claim solved AI memory.
4
4
5
-
## What CompTextv7 measures
5
+
## What CompText V7 measures
6
6
7
-
CompTextv7 measures fixture-bound replay survivability with deterministic artifacts. Current metrics and labels are intended to show whether explicitly encoded operational fields survive replay pressure, not whether a model answer is useful or semantically complete.
7
+
CompText V7 measures fixture-bound replay survivability with deterministic artifacts. Current metrics and labels are intended to show whether explicitly encoded operational fields survive replay pressure, not whether a model answer is useful or semantically complete.
8
8
9
9
Measured signals include:
10
10
@@ -20,15 +20,15 @@ Measured signals include:
20
20
21
21
## Replay-survivability evaluator brief
22
22
23
-
CompTextv7 evaluates replay survivability of compact operational state: whether fixture-authored operational fields can be compacted, reconstructed, replayed, and audited without relying on an LLM judge. The current prototype measures field survival, evidence survival, operational drift, and deterministic failure labels against checked-in fixtures. Its claims are therefore fixture-bound and prototype-scoped: it can show what the current validators detect under replay/compression pressure, not whether a deployed agent or memory product will succeed in the wild.
23
+
CompText V7 evaluates replay survivability of compact operational state: whether fixture-authored operational fields can be compacted, reconstructed, replayed, and audited without relying on an LLM judge. The current prototype measures field survival, evidence survival, operational drift, and deterministic failure labels against checked-in fixtures. Its claims are therefore fixture-bound and prototype-scoped: it can show what the current validators detect under replay/compression pressure, not whether a deployed agent or memory product will succeed in the wild.
24
24
25
-
Adjacent benchmark ecosystems include long-term memory benchmarks, RAG evaluation, long-horizon agent evaluation, software-agent/task benchmarks, and context-compression evaluation. Those ecosystems often evaluate task success, retrieval, answer quality, memory recall, or downstream performance. CompTextv7 is complementary: it evaluates whether compact operational state remains replayable and auditable, and it identifies which blockers, constraints, evidence, dependencies, recovery paths, or tool-order signals fail under compression/replay pressure.
25
+
Adjacent benchmark ecosystems include long-term memory benchmarks, RAG evaluation, long-horizon agent evaluation, software-agent/task benchmarks, and context-compression evaluation. Those ecosystems often evaluate task success, retrieval, answer quality, memory recall, or downstream performance. CompText V7 is complementary: it evaluates whether compact operational state remains replayable and auditable, and it identifies which blockers, constraints, evidence, dependencies, recovery paths, or tool-order signals fail under compression/replay pressure.
26
26
27
-
Why this matters: fluent summaries can lose blockers, constraints, evidence, dependencies, or recovery paths while still reading well. CompTextv7 treats that as deterministic replay degradation, not subjective text quality. The review path is the current trust chain: fixtures, generators, committed artifacts, Markdown summaries, README/doc values, artifact drift validation, and CI checks. See [Iterative Replay Degradation](iterative_replay_degradation.md), [Benchmark Explanation](BENCHMARK_EXPLANATION.md), the committed [iterative replay degradation summary](../artifacts/iterative_replay_degradation_results.summary.md), and [`scripts/validate_replay_artifact_drift.py`](../scripts/validate_replay_artifact_drift.py).
27
+
Why this matters: fluent summaries can lose blockers, constraints, evidence, dependencies, or recovery paths while still reading well. CompText V7 treats that as deterministic replay degradation, not subjective text quality. The review path is the current trust chain: fixtures, generators, committed artifacts, Markdown summaries, README/doc values, artifact drift validation, and CI checks. See [Iterative Replay Degradation](iterative_replay_degradation.md), [Benchmark Explanation](BENCHMARK_EXPLANATION.md), the committed [iterative replay degradation summary](../artifacts/iterative_replay_degradation_results.summary.md), and [`scripts/validate_replay_artifact_drift.py`](../scripts/validate_replay_artifact_drift.py).
28
28
29
-
## What CompTextv7 does not measure
29
+
## What CompText V7 does not measure
30
30
31
-
CompTextv7 does not measure general intelligence, answer quality, production readiness, or universal memory. It intentionally avoids:
31
+
CompText V7 does not measure general intelligence, answer quality, production readiness, or universal memory. It intentionally avoids:
32
32
33
33
- LLM judges or subjective scoring;
34
34
- embeddings, vector databases, graph stores, and external APIs;
@@ -50,23 +50,23 @@ The core contribution is a small deterministic review layer for operational repl
50
50
51
51
## Operational state vs raw chat history
52
52
53
-
CompTextv7 focuses on operational state, not raw chat-history retention. Rather than preserving every dialogue turn, it extracts, compacts, reconstructs, and verifies the fields that fixtures declare operationally relevant: tasks, constraints, blockers, evidence, dependencies, tool order, recovery actions, and continuation requirements.
53
+
CompText V7 focuses on operational state, not raw chat-history retention. Rather than preserving every dialogue turn, it extracts, compacts, reconstructs, and verifies the fields that fixtures declare operationally relevant: tasks, constraints, blockers, evidence, dependencies, tool order, recovery actions, and continuation requirements.
54
54
55
55
This framing is intentionally narrower than semantic memory. A replay can pass only for the fields represented in the fixture and checked by the deterministic validator.
56
56
57
57
## How deterministic replay validation differs from adjacent categories
58
58
59
-
| Category | What that category usually evaluates or provides |CompTextv7 boundary |
59
+
| Category | What that category usually evaluates or provides |CompText V7 boundary |
60
60
| --- | --- | --- |
61
-
| RAG evaluation | Retrieval quality, answer grounding, citation coverage, or generated-answer quality. |CompTextv7 does not retrieve documents or judge generated answers. It checks whether fixture-defined operational state survives compact/replay cycles. |
62
-
| Vector memory | Embedding-based recall and similarity search over stored memories. |CompTextv7 does not use embeddings or vector databases. It compares explicit fixture IDs, fields, attachments, and normalized values. |
63
-
| KV-cache compression | Serving-layer efficiency for model attention/cache reuse. |CompTextv7 does not optimize model internals or inference caches. It emits reviewable replay artifacts and field-survival metrics. |
64
-
| Workflow orchestration | Durable execution, retries, scheduling, state machines, and tool execution. |CompTextv7 does not run autonomous workflows. It validates whether replayed operational state still contains fixture-defined continuation requirements. |
65
-
| Learned context compression | Model-learned summaries or compressed prompts optimized for downstream performance. |CompTextv7 does not train or evaluate a learned compressor. It measures deterministic replay preservation under controlled fixtures. |
61
+
| RAG evaluation | Retrieval quality, answer grounding, citation coverage, or generated-answer quality. |CompText V7 does not retrieve documents or judge generated answers. It checks whether fixture-defined operational state survives compact/replay cycles. |
62
+
| Vector memory | Embedding-based recall and similarity search over stored memories. |CompText V7 does not use embeddings or vector databases. It compares explicit fixture IDs, fields, attachments, and normalized values. |
63
+
| KV-cache compression | Serving-layer efficiency for model attention/cache reuse. |CompText V7 does not optimize model internals or inference caches. It emits reviewable replay artifacts and field-survival metrics. |
64
+
| Workflow orchestration | Durable execution, retries, scheduling, state machines, and tool execution. |CompText V7 does not run autonomous workflows. It validates whether replayed operational state still contains fixture-defined continuation requirements. |
65
+
| Learned context compression | Model-learned summaries or compressed prompts optimized for downstream performance. |CompText V7 does not train or evaluate a learned compressor. It measures deterministic replay preservation under controlled fixtures. |
66
66
67
67
## Artifact-backed JSON and CI checks
68
68
69
-
CompTextv7 uses artifact-backed JSON and deterministic Markdown summaries so reviewers can inspect the exact replay evidence for a commit. CI artifacts are evidence records for tested fixtures; they are not universal guarantees.
69
+
CompText V7 uses artifact-backed JSON and deterministic Markdown summaries so reviewers can inspect the exact replay evidence for a commit. CI artifacts are evidence records for tested fixtures; they are not universal guarantees.
0 commit comments