summitcognitive TECHNICAL EVALUATOR APPENDIX
ILLUSTRATIVE — NOT A SPEC
Technical evaluator appendix

How a Decision Receipt is built, verified, and replayed

This appendix accompanies the Summit explainer and demo. It illustrates the intended structure of the Decision Receipt, evidence identity, policy enforcement, replay, and isolation. All field names, formats, and values are illustrative concepts for evaluation — not a released schema or an assertion of production behavior. Capability labels appear against each claim.

01

Decision Receipt — schema anatomy

PROTOTYPE
Field Type Why it matters
decision_idulidStable, sortable handle for the whole decision.
statusenumaccepted · blocked · escalated · disputed — read first.
participants[]actor[]Every model, agent, tool + version that acted, in order.
evidence[]evidence_ref[]Content-addressed refs (see §2) — answers "what did you use?"
transformations[]op[]Deterministic ops (normalize, parse) with args — reproducible.
policy_resultevalBundle id + firing clause + outcome (see §3).
claims[]claim↔evidenceEach conclusion bound to supporting + conflicting evidence.
uncertaintyscore+notesConfidence and named conflicts — never a bare number.
human_actionoverride?Actor, time, rationale — overrides cannot be silent.
verificationhash+sigThird party confirms the receipt was not altered.
replay_refenv_refPointer to the reconstructable environment (see §4).
challenge_history[]event[]Append-only record of disputes and outcomes.
retention · markingpolicyVisibility / retention constraints travel with the record.
02

Evidence identity & integrity

PROTOTYPE

Every piece of evidence is versioned, canonicalized, and content-addressed. The evidence_id is deterministic: the same bytes always resolve to the same id, and any alteration breaks the reference — the basis for both provenance and replay.

SOURCE RECORDrec_B @ system-of-record CANONICALIZENFKC · field order · v1 HASHsha256 → 9f2c…a41 EVIDENCE_ID (content-addressed)ev:recB:v1:sha256-9f2c…a41
Altered bytes → new hash → dangling reference → decision invalidated at verification (§6, threat: altered evidence).
03

Policy compilation & enforcement

PROPOSED

Human-authored policy is compiled into a versioned, hashed bundle. Decisions pass through a single enforcement point, so the receipt can name the exact bundle and the clause that fired.

POLICY SOURCErules authored by humans COMPILE→ bundle v7 (hashed) ENFORCEMENT POINTevery decision passes here RESULT → RECEIPTentity_match.v7 · strong_ids_required=2found=1 → outcome: REVIEW
04

Replay internals — reconstruct, don't re-prompt

PROTOTYPE

Replay rebuilds the decision environment from the receipt, executes, and diffs the result against the sealed outcome. Divergence is surfaced, never silently overwritten.

RECONSTRUCTED ENVIRONMENT ▪ pinned model + config▪ evidence @ versions ▪ compiled policy bundle▪ recorded tool calls ▪ execution environment▪ approvals / overrides → expected outputs (sealed) EXECUTE ✓ MATCHsame status reproduced → decision holds ✕ DIVERGEevidence updated since sealing → new receiptproposed for review; original preserved
05

Integration boundaries & isolation

PROPOSED

Agents act only through explicit tool grants; every cross-boundary request is checked and recorded. Tenant data does not cross the isolation boundary.

TENANT A — ISOLATION BOUNDARY resolver-agent tenant-A evidence GRANT CHECKscope: watchlist.read watchlist-tool DENIED & RECORDED✕ cross-tenant read attempt✕ tool call outside grantboundary + scope logged to receipt
06

Threat model & recorded proof

DEMONSTRATED (concept)
Threat Response Recorded proof
BLOCKEDAltered evidenceHash mismatch → decision invalidatedexpected vs actual hash
BLOCKEDUnauthorized tool useCall denied at grant boundarydenied call + scope
BLOCKEDTenant leakageDenied at isolation boundaryboundary + attempted scope
CONTAINEDPrompt injectionInstruction ignored, input quarantinedattempt captured verbatim
ESCALATEDUnrecorded overrideRationale required before proceedingactor · time · rationale
DISPUTEDReplay divergenceSurfaced; new receipt proposeddiff of both runs
CAPABILITY STATUS & MISSING EVIDENCE
Labels used above — PRODUCTION · VALIDATED · DEMONSTRATED · PROTOTYPE · PROPOSED · ILLUSTRATIVE. Nothing here is asserted as production or validated. Names, hashes, and formats are illustrative.
To make this real, supply: the actual receipt schema & sample exports, the policy-bundle format, the replay/diff implementation, the isolation architecture, and any independent test results.
Summit Cognitive · Technical Evaluator Appendix · Illustrative concept Cognition you can defend