Files
rea/experiments/atomic-codemode/controller-oracle.md
T
8d13c67e7e fix(experiment): preserve CFG query scope and reuse control indexes
Distinguish unretained blocks from empty relations and page generated control dependencies over one lazy snapshot index. Verify all 2,845 SQLite triples through both real MCP adapters and an independent finite-path oracle; 82 related tests pass.

Record equivalent core-query and process-family costs. Freeze the shared two-level controller question and oracle before real-model launches.

Co-authored-by: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>
Co-authored-by: Kaoru0822 <208281322+kaoru0822-kitauji@users.noreply.github.com>
2026-10-16 18:25:47 +00:00

5.9 KiB

Two-level recovered controller oracle

Freeze this oracle and question before model launches. Both lanes get the same question without addresses, native IDs or precomputed paths; this oracle stays outside their isolated workspace. Complete catalogs are eighteen direct tools or four Code Mode tools over the same API. The question requires native identities/derived relations absent from the traditional catalog, so there is no traditional-provider cell or parity claim. Both clients already expose JavaScript dispatch.

Target: Debian SQLite 3.40.1-2+deb12u2 amd64 /usr/lib/x86_64-linux-gnu/libsqlite3.so.0.8.6, SHA-256 2e6eef9a727f081f0d453b4e5e6cbd8b9ef8b6f86cbf7681cbad444d3b0b55c8. Engine SHA-256: a9ca8a8ac224936afb08a9d4be9c1225feb69473ad9f35764e8d1c4f04875b5b. REA was refreshed from main 4838f444939677107902549c59b73439e14d3ae9 before this work. Freeze runtime/HEAD during both runs. Request gpt-6.1-sol/high, Codex 0.161.0, common exact-name/envelope/stopping/long-wait guidance, fresh leases/private caches, one CPU and 4 GiB hard / 3 GiB early-stop bounds. Charge all discovery, retries, waits and actual model usage. Preserve native timing variation and incomplete attempts.

Formats and first-level controllers

References select VM entry 0xea6d0, rather than same-name stub 0x288f0. Open format 0x12c500 is produced by 0xf2381:712c in block 1142; release format 0x12c538 by 0xf2437:753c in block 1158. PHI 0xf238b:ceb7, output v100033f25, binds slots 0/1 to incoming edges 1142:0/1158:0. Shared CALL 0xf238b:7130 in block 1159 takes this format at input 2, with eight-byte register:0x30 (RSI) passing storage. Target ram:0x296c0, native label sqlite3VdbeError, does not establish a complete source signature or runtime binding.

The complete direct controller set of block 1159 is exactly:

  • Block 1141, CBRANCH 0xedc01:7065: selected 1141:1 -> 1142, true/ne-zero; other 1141:0 -> 1143, false/ne-zero. One-byte condition v100012864 is defined by 0xedc01:7064 INT_SLESS: slot 0 four-byte zero v30001285f, slot 1 four-byte v100012840 from LOAD 0xedbf8:7056. Guard: signed loaded32 > 0.
  • Block 1157, CBRANCH 0xf0389:71b8: selected 1157:1 -> 1158, true/ne-zero; other 1157:0 -> 1161, false/ne-zero. One-byte condition v100012bd9 is 0xf0384:71b3 INT_EQUAL: slot 0 four-byte v10001280a from LOAD 0xedbe5:7041, slot 1 four-byte one v3000393a3. Guard: loaded32 == 1.

Exactly one further level

  • The complete direct set for 1141 is block 1140, CBRANCH 0xedbf2:7054. Selected 1140:0 -> 1141, false/ne-zero; other 1140:1 -> 1149, true/ne-zero. One-byte v10001282b is 0xedbf0:704e INT_NOTEQUAL: slot 0 four-byte v10001280a, LOAD 0xedbe5:7041, slot 1 four-byte zero v10001282d. The selected edge negates NOTEQUAL: loaded32 == 0.
  • The complete direct set for 1157 is block 1156, CBRANCH 0xf0382:71ab. Selected 1156:1 -> 1157, true/ne-zero; other 1156:0 -> 1161, false/ne-zero. One-byte v100012bc0 is 0xf0382:71aa INT_SLESSEQUAL: slot 0 four-byte one v500038ab1, slot 1 four-byte v100012b9e, LOAD 0xf0378:719d. Guard: signed loaded32 >= 1.

Open outer and release inner comparisons share exactly v10001280a (P). Conjoining their selected predicates would require P == 0 and P == 1. These are alternatives, not simultaneous prerequisites. Locally open combines P == 0 with signed S > 0; release combines signed T >= 1 with P == 1. These are not complete VM-entry path conditions or a demonstrated SQL trigger.

P, S (v100012840) and T (v100012b9e) all have unique-space coordinate 0x23d00/4 but are distinct SSA values. S/T have separate LOAD definitions, even though machine code reads [R13+0xd8] in each path. Matching coordinates or memory expressions do not equate contents across reads/paths or solve aliases.

Graph witnesses

Accept any edge-valid witnesses with the required endpoints and omitted block, not only these shortest examples:

  • Selected: 1142 -> 1159, edge 1142:0; 1158 -> 1159, edge 1158:0.
  • Open alternative avoiding 1159: 1143 -> 1144 -> 1201 -> 1204 -> 1205 -> 1207, edges 1143:0, 1144:0, 1201:0, 1204:0, 1205:1.
  • Release alternative avoiding 1159: 1161 -> 1162 -> 1163 -> 1164 -> 1205 -> 1207, edges 1161:1, 1162:1, 1163:1, 1164:0, 1205:1.

1207 has zero successors. Check every edge against the retained graph. These finite recovered paths establish neither machine paths, feasible predicates nor termination.

Independent machine checks and grading

ELF64LE PT_LOAD/file reads verify sixteen instruction sites and eight relative targets; GNU decoding corroborates them. Receipt: controller-machine-oracle.json locally. Earlier sites are in the savepoint oracle. Additional facts:

  • 0xedbf0 85f6 TESTs ESI; 0xedbf2 0f8548270000 is JNE 0xf0340, zero falls through to 0xedbf8. Native successor origins are separate metadata.
  • 0xf0378 458b9dd8000000 loads R11D from [R13+0xd8]; 0xf037f 4585db TESTs it; 0xf0382 7e0b is signed JLE 0xf038f, positive falls through to 0xf0384. Native true/ne-zero here is not machine JLE-taken.
  • Other guards remain signed JG 0xedc01 -> 0xf2372 and equality JE 0xf0389 -> 0xf242d; format setups and shared CALL bytes are unchanged.

The full graph has 21,349 operations, 35,755 values, 1,229 blocks and 2,061 edges. Both actual MCP adapters match all 2,845 control triples against the independent node-removal/exit-reachability oracle, with source fidelity and exact native edge metadata. This is not global machine/IR equivalence. Grade the complete answer and actual calls: both controller levels, all four conditions, shared/distinct SSA identities, valid witnesses, selected machine corroboration, real source IDs/digest and explicit recovered/alias/machine/nontermination/runtime limits. A cheap incomplete/incorrect answer is not an efficiency success. Close the lease.