AIWS showroom

Independent research & development · Local candidate

Gitty / Source continuity and verified retrieval

Keep the source in the answer.

Verification kits: requirements, scope and distribution blockers

Find the source before trusting the answer.

Gitty addresses a common development problem: an answer survives, but the source version, project context and failed checks disappear.

Within the inspected candidates, Gitty is a local source-indexing and verified-retrieval implementation. It connects readable search records to structured records and original source bytes. It helps answer “which source supports this?” rather than treating a search hit as accepted knowledge.

Observed here

Retrieve and check

Seven read-only API checks produced their expected statuses against one saved candidate generation. Returned source bytes matched the original.

Recorded earlier

A corrected candidate

REVIEW-001 integration records 105 unittest executions across seven unittest suites, plus one CLI assertion script: eight selected suites in total.

Still unresolved

No product completion claim

CE2 remains OPEN. PRODUCT_ROOT, current project authority and authenticated independent verification remain UNKNOWN or unestablished.

Why CSV first, with SQLite behind it?

CSV exposes the portable search projection: a reviewer can inspect search text, scope and record locators without treating an opaque database as the only source of context. In the inspected query code, CSV is searched first. Matching record IDs then select corresponding SQLite records, whose scope and source references are checked.

SQLite contributes structured lookup and consistency checks. It is a rebuildable projection in this information model; it does not replace the original source or decide authority. The showroom’s downloadable CSVs are a separate presentation projection, not Gitty’s Registry or authoritative project database.

SHA-256 checks whether bytes match an expected digest. A different digest rejects the proposed byte identity. Matching bytes alone cannot establish truth, safe behavior, authorship, ownership or permission. A pin supplied by the same untrusted party needs a separately justified trust anchor.

Receipt identifies delivered artifacts and checks. Handover links the receipt and preserves the frontier, blockers and continuation context. Here their hash links match; their identities remain UNSIGNED.

A source-grounded worked example

Observed API readback + conceptual routing context · not one end-to-end Registry execution

Question: “Find the policy-loader source in the Gitty candidate so it can be reviewed.” Sources G-CONSUMER, G-READBACK and G-QUERY support the steps below.

  1. Identify the project. The API requires the explicit Gitty project identity and candidate scope. Wrong project or scope returned SCOPE_REJECTED in this task.
  2. Resolve a verified Registry pointer, when available. The broader query implementation checks Registry route CSV and SQLite agreement. This task did not establish a current operational Registry pointer or product root. For the observed calls, an existing Receipt/source chain supplied an explicit saved-generation location and manifest pin instead.
  3. Search the project CSV. SEARCH for “policy” returned FOUND. The inspected consumer searches the candidate’s CSV projection, not the whole ecosystem.
  4. Check corresponding SQLite records. The consumer checks its saved generation’s records and original-source declarations. The query code checks CSV/SQLite identity, generation and source consistency. This description comes from inspected code; no SQL trace was instrumented here.
  5. Verify the original. GET_VERIFIED_SOURCE returned the older policy-loader version. This task decoded the response and compared all 4,711 bytes with the pinned original. It did not confuse that version with the later corrected integration source.
  6. Answer within the evidence. “This is the identified source for bounded review.” This does not answer “is the whole product safe?” A miss returns NOT_INDEXED; a wrong manifest pin returns INTEGRITY_FAILURE, and must stop the answer.

Conceptual diagram · source-grounded mechanism

Question → explicit project → verified route or explicit pinned generation → project CSV → matching SQLite records → original-byte check → scoped answer / explicit failure

The inspected broader query function permits parent/Master fallback only after a verified project miss, with a verified edge and the specified fallback policy, where the relevant contract permits it. An unreadable, stale or corrupt index is not a verified miss. The saved generation explicitly marks fallback DISABLED and edge verification UNKNOWN. The observed Consumer API does not discover routes or perform this fallback. Master remains REFERENCE_ONLY; a reference never transfers project ownership.

A failed search is useful information.

Observed consumer statuses and their limits
OutcomeMeaning in this candidateWhat follows
FOUND / VERIFIEDA scoped hit or successful integrity checkReview identified evidence; no authority gained
NOT_INDEXEDNo match in the available verified projectionDo not infer the source does not exist elsewhere
INTEGRITY_FAILUREGeneration, pin or source verification failedStop; do not turn index failure into a search miss
SCOPE_REJECTEDExplicit project/scope does not matchStop; do not silently search another project

Evidence with its limits attached

Linked identities below resolve to sanitized summaries, not private source payloads. Exact paths, versions and SHA-256 values remain in task-local evidence, outside dist/.

Each row has a separate scope; counts must not be added into a whole-product score
Precise claim / source identityReported outcomeVerification level hereLimit
G-INTEGRATION: REVIEW-001 integration105 unittest executions; 8/8 selected suites exit 0; 176 ZIP entriesOriginal reports/logs inspected; 175 manifest entries matched; Handover link matched; ZIP countedHistorical tests not rerun. Full regression not run; inherited 9 failures and 1 error unresolved
G-CONSUMER: read-only consumer19 consumer, 7 realfile and 86 policy checks reported separately; real-source CLI PASS30 manifest entries and Receipt linkage matched; implementation inspectedCounts include inherited/imported tests; historical generation and custody limitations persist
G-REVIEW: earlier source consumptionPASS_WITH_REVIEW_FINDING; malformed owner list raised TypeError24 manifest entries matched; report inspectedStubbed dependency probe; no demonstrated trust bypass. Later correction is a distinct version
G-READBACK: this task’s saved-generation API checks7 expected statuses; decoded original-byte equalityLocally re-executed and byte-checked by the performerOne candidate generation; no independent verifier, production test or full regression
G-QUERY: CSV-first query mechanismCode implements route checks, CSV selection, SQLite agreement and conditional fallbackOriginal implementation bytes inspected and manifest-matchedMechanism observation; operational Registry and fallback were not executed here
G-REFRESH: original-source generation86 policy and 7 realfile checks reported separately55 manifest entries matched; source chain followedHistorical failures retained; atomic multi-file snapshot and privileged ABA guarantees unproven

Independent verification: NOT_PERFORMED. A recorded result, a terminal readback, locally reverified bytes and independently verified behavior are different evidence levels. This task establishes the middle two only within its stated scope.

What this suggests about the architecture

Observation: the inspected code separates lookup, structured record checks and original-byte verification. Interpretation: that separation helps locate failures instead of accepting every search result. Hypothesis: it could improve continuity across larger workflows; broader evaluation is needed.

What could limit the claim? Missing source bytes, untrusted pins, stale projections, inconsistent records or an unauthorized route can prevent a supported answer. Passing selected tests does not remove custody, concurrency or deployment questions.

Gitty concerns source continuity. AIWS concerns semantics and contract boundaries; CI concerns cognitive architecture; CHARLI concerns a distinct entity. Harbour and NodeFlow retain their own execution/control and coordination scopes. No live integration, project membership or governing binding among these projects is established by this page. Inspect the typed relationships.