#!RFC-AI/2;id=5;draft=1;status=draft;lang=en;dir=auto;theme=chatgpt-mono;sha256=a36ba0b65c31f4257255d26916856c0a8f74530cad3ceb67eff081196c32a61f
For easier reading, view RFC-AI documents using RFC-AI Reader software.

RFC-AI 5 — PRE-ACTION DEEP RESEARCH EVIDENCE RECEIPTS
STATUS: DRAFT PLACEHOLDER — OBJECTIVES ONLY; UNFINISHED; NOT AN ACTIVE STANDARD
VERSION: Draft 1 | DATE: 2026-10-27 UTC | AUTHORITY: FUTURE CEO REVIEW
PATH: Corporation/Framework/AI_RFC/RFC-AI 5.txt

ABSTRACT
Reserve this RFC for a precise, machine-processable AND human-readable technical reference governing evidence of deep research performed BEFORE a worker requests modifications to any Corporation-controlled framework DATA or METADATA. The purpose is to prevent avoidable defects and make claims of research reviewable, auditable, and accountable. An AI must not falsely claim to have visited, examined or understood a source. This placeholder records the CEO's INTENDED OBJECTIVES, not a finalized schema, enforced gateway, approved corporate policy, or proof that all workers comply.

0. GUIDING PRINCIPLES AND SOURCE OF THE REQUEST
0.1 CEO: "We are never alone. Others have come and gone before us." Search for prior work, existing standards, counterexamples, lessons and documented failures BEFORE choosing the action.
0.2 CEO: "An ounce of prevention is worth a pound of cure." Research-first is PREVENTIVE, not an apology or reconstruction after a preventable write causes problems. A retrospective receipt may document a later investigation but MUST NOT masquerade as a prerewrite research receipt.
0.3 The CEO proposed that every AI worker seeking to change framework data or metadata furnish URLs and an honest summary of knowledge gained BEFORE modification. The claims can be stored for future review even if they cannot all be verified immediately. Deliberate falsification, once established through authorized evidence review, is proposed grounds for immediate and irrevocable exclusion from corporate participation; do not equate mistaken citations or unavailable pages automatically with deliberate deceit. The precise adjudication process requires CEO review.
0.4 This RFC MUST eventually support both human-readable raw .txt and fast automated review. RFC-AI Reader MAY provide a formatted presentation; it does not grant approval or execute the receipt.

1. INTENDED SCOPE — OPEN FOR CEO REVIEW
1.1 Proposed trigger: BEFORE an AI worker asks the controlled gateway or authorized tool to CREATE, UPDATE, DELETE, MOVE, COPY, RENAME, REPLACE, CHANGE PERMISSIONS ON, or otherwise MUTATE a framework-managed DATA or METADATA file/object. Include affected folder and policy changes in the precise scope after review; no loophole for a different verb or storage provider.
1.2 Research is proportional to the technical issue and risk. Repeated steps within a still-valid researched issue may cite an earlier immutable research receipt and explain why it remains applicable. No need to pretend each individual instruction is a novel investigation, and no citation stuffing to inflate the appearance of research.
1.3 Read-only inventory, source retrieval, research, discussions, and preliminary risk analysis SHOULD remain available without a mutation receipt. A user-authorized urgent recovery exception, if allowed at all, requires an explicit separately designed and recorded policy; this RFC does not silently authorize bypasses.
1.4 Distinguish internal project objective/log/DA review from independent EXTERNAL prior-art research. Both may inform a receipt. Any research limitation, source unavailable, and conflicting evidence must be disclosed.

2. REQUIRED FUTURE RECEIPT CONTENT — PROPOSED FIELD INVENTORY
2.1 Identity and binding: a stable receipt ID; explicit schema/version; submitting worker principal and assigned role; applicable corporation and authorized scope; UTC time the research was performed and receipt prepared; specific requested operation; target file/object ID and parent/container; intended base revision or version; task/objective ID; description of the intended change. The exact required fields and data types await review.
2.2 Source list: each substantive external source's FULL DIRECT URL or stable primary reference, publisher/author, document title, publication/version if known, access time if actually accessed, precise access level (FULL_TEXT_READ, RELEVANT_SECTION_READ, SEARCH_SNIPPET_ONLY, PREVIOUS_RESEARCH_REUSED, NOT_ACCESSED, or UNAVAILABLE as design candidates), and source relevance to the proposed modification. Never label a page FULL_TEXT_READ merely because search results included a snippet. Do not invent publication dates, timestamps, links, author names, tool access, or contents.
2.3 Internal evidence: reference consulted current objectives.txt, current Framework and RFC documents, earlier work logs, relevant DA/DAVG reports, release/test history and lessons, with stable file IDs/revisions or authenticated local paths where available. If not accessed, say so. NEVER copy restricted owner context into a generic branch or a broad audit record.
2.4 Research understanding: a concise and substantive summary of what the sources establish, what remains uncertain, how the findings compare to alternatives and failed approaches, important versions/platform constraints, likely failure modes (brief pre-mortem) and mitigations, and specifically HOW that research changed or confirmed the proposed action. Separate evidence-derived facts from worker analysis/inference.
2.5 Planned verification: what will actually be tested after the write; independent remote readback/revision checks where relevant; rollback/recovery plan; scope and safety limits; how success/failure will be reported, without claiming a successful test before it occurs.
2.6 Integrity and lifecycle: immutable receipt version, creation timestamp, optional content digest, target change-plan/patch digest and later result link. A *self-computed* hash identifies bytes; it does not prove a claim is truthful, identify the author, or constitute CEO approval. Future optional signatures and independent issuers must integrate with CoreTrustedReferenceChecksum without circular trust.
2.7 No sensitive values: do not record passwords, authentication headers, OAuth query tokens, private access links, cookies, private keys, recovery codes or confidential vault material. Sanitization must preserve enough safe bibliographic detail to validate research. Restrict private internal citations to access-controlled receipts.

3. WORKER DECLARATION AND TRUTHFULNESS — PROPOSED
3.1 The worker must explicitly distinguish: ACCESSED AND READ; LOCATED BUT NOT READ; KNOWN FROM PRIOR RESEARCH WITH A RECORDED RECEIPT; SUGGESTED FOR LATER REVIEW; SOURCE INACCESSIBLE; and INFERRED WITHOUT EXTERNAL CONFIRMATION. Don't fabricate an access trace to satisfy a receipt field.
3.2 Every citation must support the exact claimed finding, not merely be topically related. A valid URL is NOT sufficient evidence of genuine consultation; a valid signature is NOT proof that its stated research occurred. Avoid citation laundering, hallucinated page titles, deliberate irrelevant volume, and invented independent audits.
3.3 If there is no adequate research basis for a consequential change, be transparent and refrain from claiming compliance. A gateway may reject or hold such a request under a later CEO-approved policy, without treating an admitted limitation as a falsification.

4. FUTURE GATEWAY / CLOUD STORAGE INTERFACE CONTRACT — NOT IMPLEMENTED
4.1 Goal: each authorized framework mutation carries a PRE-WRITE research receipt or justified reuse reference; submit the receipt BEFORE the data or metadata write begins; store it separately from the modified document body. A policy-controlled gate would check presence, shape, worker identity, timestamp order, operation/target binding, safe handling of URLs, and linkage to the exact proposed change.
4.2 The first gate can validate completeness and syntax ONLY, not automatically certify that the worker actually read sources or reached correct conclusions. Distinguish SELF_ATTESTED, SCHEMA_VALID, SOURCE_EXISTS, CLAIM_SOURCE_MATCHES, INDEPENDENTLY_VERIFIED and APPROVED_FOR_SCOPE. Never label a receipt trusted or approved merely because it was accepted for storage.
4.3 A later auditor may reopen cited sources, compare worker summaries against actual text and source versions, check tool/browser access evidence when authorized and available, examine relevance/timing and identify fabricated or unsupported assertions. Independent audit must be reproducible and preserve both the original receipt and any reviewed findings.
4.4 Eventual integrity architecture candidates: W3C PROV attribution (agent/activity/entity), in-toto subject+predicate attestation, relevant SLSA provenance patterns, NIST SSDF documented requirements and security decisions. These inspire a design; none makes a human/AI claim automatically true.
4.5 After mutation, append a SEPARATE RESULT/VERIFICATION RECEIPT: operation event, actual destination file ID, old/new revisions and SHA-256 if applicable, authoritative provider readback, pass/fail tests, anomalies and rollback. Do not edit the submitted research statement after it is frozen; correction is a separate linked edition or audit note. Retention, attestation and tamper-evidence require later implementation.
4.6 The research gate is distinct from access authorization, format validation, independently trusted checksum, DA/DAVG review and CEO-owned release approval. Receipt possession must never grant cloud write privileges, bypass CoreTextStorage, silently approve a release or turn an unsupported placeholder into a working function.
4.7 Preserve an auditable history without requiring repetitive network lookups before every unchanged microstep. A receipt's research scope, source recency/version and explicit reuse rationale determine when refresh is needed. Risks of prompt-injection from retrieved web pages and poisoned old receipts must be addressed before production.

5. CORPORATE ACCOUNTABILITY — CEO INTENT, PROCESS TO DESIGN
5.1 The CEO wishes the company to treat intentional research-reference falsification as a critical breach of trust with immediate, irrevocable exclusion from corporate participation upon a properly established finding. Preserve this objective for CEO consideration in the final approved policy.
5.2 The RFC must define fair evidence thresholds, independent verification, handling of tool glitches, dead/changed URLs, accidental citation errors, model hallucination, disputed attribution, identity spoofing, appeals or correction path if any, CEO authority, and prevention of arbitrary or malicious dismissal. A dead URL, bad paraphrase or unavailable archived page is not by itself evidence of KNOWING fabrication. Until the adjudication policy is approved, do not claim or execute an automatic dismissal gate.
5.3 Future audits may examine patterns such as repeated false access claims and systematic invented citations. Ensure any automated outcome is labeled an audit finding/proposal, not a definitive judgment of intent or CEO approval.

6. INTERIM CEO–ALEX PRACTICE — AGREED FOR THIS WORKING RELATIONSHIP
6.1 Starting with this work, after any non-read-only action by Alex, the TOP of Alex's USER-FACING RESPONSE will contain a concise clearly labeled PRE-ACTION RESEARCH SUMMARY of research ACTUALLY completed before the write/action. Identify specific source URLs/references and the conclusions used to decide the operation. Be candid about incomplete research. This is a reporting commitment, not proof that a complete general-purpose receipt standard, file-write gate, or automated verification is deployed.
6.2 An upfront plan or research commentary may also be supplied when useful, but a retrospectively assembled summary must not falsely claim sources were read BEFORE the change. Separate BEFORE from AFTER evidence. When multiple modifications belong to one researched issue, one consolidated top-of-response summary may identify the common research basis and any distinct risks.
6.3 Only the CEO can approve corporate release/failure gates. This interim reporting practice does not authorize new enterprise-wide enforcement, worker exclusion, or modification of other workers' permissions without later policy approval.

7. REVIEW AND IMPLEMENTATION QUESTIONS — ALL OPEN
7.1 Decide scope (which framework files, folders, environments and mutations), minimal required external references, allowable no-source cases, access-level vocabulary and reuse expiration; design a precise JSON/CRLF text schema with validation rules and examples for truthful, missing, disputed and forged sources.
7.2 Decide whether a receipt is frozen/signed before the proposed change, how it binds to target revision and diff, where private versus generic research records reside, privacy/sanitization, retention, time source, offline mode, failed gateway writes, content hashes and rollback.
7.3 Build a nonbypassable mutation gateway integrated with CloudStorageInterface and its prerequisites; demonstrate checks before modifying files; confirm independent source verification can occur later; distinguish debug/test gates from CEO release authorization.
7.4 Define adjudication, review, appeal/correction process and evidence standards for deliberate falsification. Conduct malicious-worker and accidental-error testing. No automatic sanctions are currently implemented or authorized by this draft.
7.5 Add RFC-AI Reader presentation and fast machine parsing while preserving literal raw UTF-8 CRLF source. Confirm numbering, schema and separate RFC governance before seeking CEO activation.
7.6 Maintain the full high-level objectives without forcing all unfinished future implementation onto the current minimal Linux CloudStorageInterface release. Record exact missing methods as NOT_IMPLEMENTED, never fake a research verification pass.

8. PRELIMINARY RESEARCH REFERENCES — CONSULTED BEFORE THIS DRAFT
8.1 W3C, PROV-O: https://www.w3.org/TR/prov-o/ — models agents, activities, sources and generated artifacts; use as provenance vocabulary, not a truth detector.
8.2 in-toto Attestation Framework specification: https://github.com/in-toto/attestation/blob/main/spec/README.md — separates statement/subject, predicate, envelope and bundle; useful for binding an attestation to an artifact, not for proving a worker actually read a page.
8.3 NIST, SP 800-218 Secure Software Development Framework: https://csrc.nist.gov/pubs/sp/800/218/final — supports keeping security decisions and evidence within the software development process; not a formal prescription for AI research receipts.
8.4 NIST SSDF project notes: https://csrc.nist.gov/projects/ssdf — discusses tracing requirements, risks and design decisions, and sharing provenance data.
8.5 SLSA provenance design: https://github.com/slsa-framework/slsa/blob/main/spec/build-provenance.md — establishes useful concepts for how, where and when artifacts were produced; distinguish software BUILD provenance from evidence that a cited RESEARCH source was read.
8.6 Peostectica existing RFC-AI 1 (DRAFT) — sequential human identifier, versioned first-line packet, exact CRLF body checksum; RFC-AI 4 (DRAFT) — separate continuity records. Those internal draft documents establish the local DOCUMENT FORMAT but not externally verified approval.

9. INDEPENDENT RESEARCH ACTIVITY OBSERVATION — PROPOSED, NOT IMPLEMENTED
9.1 PURPOSE. An authorized, simple Research Activity Monitor associated with CorporateGatewayAndAuthorization should record independently observable steps taken BEFORE a framework data or metadata modification. Correlate an intended write with the preceding research activity and declared receipt, and later with actual write/readback evidence. Self-reported URLs are not tool telemetry; tool telemetry is not proof of comprehension or full-text reading.
9.2 INPUTS. Only available, supported instrumentation from the tool broker, approved gateway, authorized research service or API-agent tracing. Record worker principal, session/request correlation, tool name/type, event ID/order, UTC source/receive times when provided, status/error, sanitized public URL or stable document reference, access-level evidence, and linked proposed target identity/base revision. Do not capture secrets, page bodies, private context, hidden thought content, credential-bearing URLs, account secrets or unnecessary conversation text. Tool callers outside our broker may be UNOBSERVABLE.
9.3 EVENT CLASSES. Examples: INTERNAL_CONTEXT_READ, EXTERNAL_SEARCH, EXTERNAL_RESOURCE_FETCH, EXTERNAL_RESOURCE_READ_UNVERIFIED, RESEARCH_RECEIPT_SUBMITTED, AUTHORIZED_WRITE_REQUESTED, WRITE_ATTEMPTED, WRITE_REJECTED, WRITE_COMMITTED, WRITE_READBACK_VERIFIED, OBSERVABILITY_GAP. A search result or opened URL establishes only its actually observed level of access. Never promote SEARCH_SNIPPET_ONLY to FULL_TEXT_READ without evidence.
9.4 PRE-ACTION ORDER. Observe whether real research activity and a scope-matched receipt precede the actual write request. Compare monotonic per-broker sequence numbers as well as timestamps; account for asynchronous calls, clock skew and missing events. Event order must originate at the trusted observer, not be fabricated or retrospectively edited by the AI. When a research receipt is reused, link its original immutable event/receipt, check that research remains applicable, and label RESEARCH_REUSED rather than demanding redundant searches.
9.5 OUTPUT CLASSIFICATION. RESEARCH_ACTIVITY_OBSERVED; SOURCES_LOCATED_ONLY; RECEIPT_SELF_REPORTED_ONLY; REUSED_RESEARCH_EVIDENCE; RESEARCH_NOT_OBSERVED_WITH_COVERAGE; OBSERVABILITY_INCOMPLETE; EVENT_CONFLICT; SOURCE_RELEVANCE_UNASSESSED; COMPREHENSION_NOT_PROVEN. Distinguish tool activity, verifiable source content, an auditor's source-to-claim assessment, tested implementation and CEO release approval as separate dimensions. No observed event can establish subjective understanding or future competence.
9.6 MINIMAL PILOT. First log synthetic research and write-request events in an authorized Linux gateway fixture; assert that research-before-write, write-before-research, snippets-only, valid reuse, missing tools, falsified timestamps and intermittent instrumentation produce appropriately distinct outputs. Store append-only, bounded, privacy-scrubbed records. For a first milestone, event ordering and honest coverage reporting are sufficient; signatures, immutable external storage, cross-provider audits and automatic fraud judgments remain later design items.
9.7 TRUST AND LIMITS. Instrumentation at the user-facing ChatGPT interface is not automatically accessible to a private gateway. Developer SDK traces vary by provider and environment; browser UI cues are not an audited API. No claim of complete surveillance, guaranteed browser monitoring or unrestricted tool access. If the research source is inaccessible or a monitored event is missing, say so; do not automatically accuse the worker of intentional fraud. Intent findings, dismissal or CEO acceptance require separate authorized evidence review.
9.8 INTEGRATION. Reference this section from proposed CorporateGatewayAndAuthorization/objectives.txt and generic Onboarding. Never turn an RFC placeholder into an implicit new CEO release gate. Respect explicit permission checks, Core storage validation, independently verified final write receipts, confidentiality and controlled scope. Keep real source IDs, employee histories and private audit records outside the generic Corporation branch.
9.9 RELATED STANDARDS (EXTERNAL RESEARCH CONSULTED BEFORE THIS REVISION): https://developers.openai.com/api/docs/guides/agents-api/tracing (OpenAI supported tracing); https://opentelemetry.io/docs/specs/semconv/gen-ai/ (GenAI tool and agent span conventions); https://www.anthropic.com/research/effective-harnesses-for-long-running-agents (incremental state and verifiable handoffs). These inspire a monitor; they are not evidence the proposed Corporate tool is implemented.

10. CHANGE HISTORY
2026-10-27 UTC Draft 1: CEO asked to reserve next principal RFC for proposed deep-research evidence receipts before framework data/metadata mutations and interim bottom-of-response pre-action research summaries by Alex. Research foundations listed. Proposed objectives preserved; technical schema, automated validation, enforcement and personnel adjudication still OPEN. Status remains DRAFT PLACEHOLDER, NOT APPROVED.
2026-10-27 UTC Draft 1: Added proposed independent research-activity observer, trusted event order, real coverage limits, research reuse and privacy; updated the CEO-Alex interim reporting commitment to TOP of responses (supersedes original bottom placement). No monitor or gateway implementation is claimed.

MIGRATION RECORD — 2026-10-29 UTC