Index source identity
Sort portable paths. Record exact byte lengths and BLAKE3 file digests under one declared inclusion policy.
symgliph.manifest/v4
Source-verifiable
context, compiled.
Symgliph gives a corpus a deterministic identity, selects candidate source evidence, and emits ordinary text under an explicit model-token budget. Every included excerpt is verified before it reaches a model.
Context compiler
01 / 03
Corpus indexed
Canonical files retain their exact byte identity.
Measured, not generalized 30 Rust tasks · one localizer · one response per arm
97.58%
fewer prompt tokens
−11.56 pp
fix-path recall tradeoff
Read the scope
Claim boundary
Verification has
a precise boundary.
Retrieval predicts which evidence may help. Verification proves that every selected excerpt matches the indexed source. Symgliph guarantees identity, provenance, and budget compliance—not relevance, completeness, or answer correctness.
query
→
candidates → verified excerpts → bounded prompt
Research protocol
Four capabilities.
One deterministic core.
SCGP separates protocol guarantees from empirical performance. Context and Glyph are sibling profiles: model-facing context does not depend on SVG rendering.
-
I
Identity
Canonical files, corpus root, manifest, verification.
required -
S
Semantics
Versioned analyzers, nodes, references, and semantic root.
extends I -
C
Context
Source-verifiable text packets under a declared token budget.
extends S -
G
Optional glyph
Deterministic SVG fingerprint of corpus and semantic identity.
extends S
Context compiler
From corpus to bounded prompt.
Four stages index supported structure, select candidate evidence, and emit source-backed text that an ordinary model can consume.
Compilation frame
01 / 04
Canonical source
Deterministic
Derive addressable structure
Extract nodes, typed references, source spans, and analyzer versions. Ambiguous targets remain explicitly unresolved.
symgliph.semantics/v3
Commit independent roots
Create separate corpus and semantic roots. Source identity stays stable while declared analyzers are free to evolve.
ρ = BLAKE3(transcript)
Select, verify, and pack
Rank candidate nodes, expand graph neighbors, re-hash selected source files, and fit excerpts beneath a hard token ceiling.
tokens(P) ≤ B
Measured result
Efficiency improved.
Recall did not hold.
Thirty pinned Multi-SWE-bench Rust tasks across ripgrep, fd, bytes, and tracing. The original four-arm study plus a separately recorded adaptive follow-up used one localization model and one response per task and arm. Cold embedding overhead is included where applicable.
| Measure | All files | BM25 | Embedding | Fixed | Adaptive |
|---|---|---|---|---|---|
| Mean paired prompt reduction | 0.00% | 95.82% | 95.78% | 97.58% | 97.80% |
| Gold-path coverage in context | 100.00% | 58.55% | 56.49% | 51.85% | 51.85% |
| Exact predicted fix-path recall | 53.32% | 46.87% | 43.34% | 41.76% | 45.09% |
| Recorded API spend | $4.5755 | $0.0974 | $0.3237 | $0.0668 | $0.0573 |
Limits
What the protocol establishes—and what remains open.
Protocol guarantees
- Included corpus identity
- Analyzer-versioned semantic identity
- Selected source and excerpt digests
- Portable paths and exact byte ranges
- A declared model-token ceiling
Open research questions
- Retrieval completeness across domains
- Answer-quality preservation
- Repeated-model variance
- Enterprise policy enforcement at scale
- Operational cost beyond provider API spend
Reproduce
Run each claim through its own gate.
SCGP conformance, ECEB policy checks, and both comparative study records are independent claims with independent commands. All four are checked in and run in continuous integration.
Four independent gates
$ spec/validate.sh
$ scripts/eceb-conformance.sh
$ scripts/eceb-study-check.sh
$ scripts/eceb-adaptive-study-check.sh Independent gates checked in
Enterprise direction
A research foundation for governed context.
Today, Symgliph is a protocol and context compiler—not an enterprise gateway. If broader evaluation establishes quality preservation and policy behavior, this architecture can support budget controls, provenance, access rules, and auditable AI spend. The geometric mark remains a corpus fingerprint, never model input.