DIVERGE-SC1 Raw-Source Object Compiler

Status: closed; read-only failure localization complete

Decision date: 2026-08-05

1. Decision

DIVERGE-v0 is closed as a broad architecture promotion because its corrected factorized representation misses the frozen >=2x storage advantage at four worlds. Its exact high-ambiguity (>=8 worlds) delayed-recovery mechanism is retained. The next experiment does not alter that executor. It removes the successful component pilot's hidden source scaffolding.

The token-role/source-copy pilot received each gold physical record and each gold option as a separate model call. It proved that frozen SmolLM2 residuals can classify finite roles and copy complete options, but not that a model can construct those objects from one raw source. The earlier whole-mention span gate did read a complete source, but decoded fields independently and failed under lexical shift through duplicate and missing roles.

DIVERGE-SC1 tests one hypothesis:

A source-local occurrence ledger plus a globally scored complete-record parser can turn one unsegmented language source into the same exact sealed DIVERGE packet more reliably than independent token decisions, because the legal object is a coherent pair of complete options rather than a bag of locally likely fields.

This is a compiler hypothesis. It is not a DIVERGE promotion, a general reasoning claim, or permission for continuation pretraining.

2. Only architectural change

The model reads the complete raw WORLD source once and emits:

1. source-token role scores for candidate/background cues, nominal alias boundaries, favored/reserve priors, and ordered action identities;
2. boundary scores for every source-token gap; and
3. no record list, option list, alias dictionary, renderer ID, ontology ID, answer, query, state, or execution outcome.

The hard compiler creates two separate ledgers:

• occurrence ledger: every selected physical alias span has a unique source address even when its bytes equal another mention;
• nominal ledger: selected occurrences with exactly equal normalized source bytes share a nominal signature, without fusing their physical addresses.

It globally scores complete nonoverlapping records. Every accepted record has one source-local kind cue and exactly two nonoverlapping complete options. Each option has exactly one alias span, one prior, and one legal ordered action program. The decoder may select a locally second- or third-ranked field only when the total complete parse scores higher. Invalid or overflowing parses fail closed. Source-position canonicalization defines option IDs; generator- private option ordering is never exposed.

The CPU calibration cap is 4,096 complete record proposals per episode. An initial 1,024-proposal calibration failed closed on one 452-token, nine-record composition episode; that report is retained as a pre-freeze capacity negative. The 4,096 cap is frozen before any neural score.

After hard compilation, only canonical occurrence addresses, nominal commitments, priors, programs, record provenance, and accounting receipts enter the existing DIVERGE packet. Raw source bytes, source residuals, and source KV state are deleted before delayed evidence, execution, or query.

3. Grammar firewall

Allowed fixed constraints are:

• nonoverlap and source order of physical spans;
• one alias, one prior, and one legal action program per option;
• exactly two options per accepted record;
• one record-kind cue per record;
• source-local gap consistency and bounded span/candidate counts; and
• exact-byte nominal equality among already selected alias spans.

Forbidden constraints are:

• gold record/option/alias spans at inference;
• a renderer-specific parser, alias dictionary, or action phrase lookup;
• DIVERGE world state, trajectory, delayed evidence, query, answer, or target packet in decoding;
• retry after execution, host semantic repair, or answer-guided selection; and
• merging physical occurrences merely because their nominal bytes match.

The supervisor may use gold spans and packet fields for training and scoring. The candidate runtime may not.

4. CPU nontriviality gate

Before neural training, an independent generated board must cover:

• 2--6 candidate records plus 1--3 background records;
• one- through four-token aliases and repeated nominal aliases at distinct physical occurrences;
• record and option order permutations;
• variable inter-record connectors with no supplied record offsets;
• local prior/action/alias decoys that create rank-two or rank-three fields;
• all four noncommuting finite programs; and
• train, lexical-shift, renderer-shift, and composition-shift cohorts.

The candidate CPU decoder receives only tokens and synthetic unary/boundary scores. An independent exhaustive assessor receives annotations. Freeze neural thresholds only after calibrating this board.

Required CPU outcomes:

1. 100% exact parity between the joint decoder and exhaustive best legal parses on at least 1,000 episodes;
2. 100% packet reconstruction when supplied calibrated gold-plus-decoy scores;
3. at least 25 points exact advantage over independent local decoding using the identical scores;
4. at least 95% of episodes contain a true field locally ranked second or third;
5. shuffling boundary scores or fusing occurrence and nominal identity reduces exact reconstruction by at least 20 points;
6. alpha renaming, consistent source-position reindexing, and post-seal source poisoning change no compiled semantics; and
7. zero accepted overlapping records/options, duplicate option fields, false nominal merges, or cap violations.

A CPU miss closes this decoder before H100 use.

5. Neural matched gate

If CPU mechanics pass, use the frozen SmolLM2-135M parent and the same full- source residual layer, data, updates, seeds, and packet evaluator for:

Arm A is a ceiling, not a fair autonomous model. B--G must share the exact encoder and unary scores where their interface permits. The treatment may not receive additional semantic labels or source views.

The frozen neural pass gate is:

• development and shifted gold-support recall >=99%;
• development and shifted exact packet >=95%;
• shifted strict packet at least 15 points above B and C;
• every held-out cohort exact packet >=90%;
• five of five seeds above B, with at least four of five passing all floors;
• zero distractor records in exact packets and zero false nominal merges;
• source deletion and post-seal poison invariance exactly 100%; and
• at least 20-point collapse for the causally relevant E/F/G controls.

The second and fourth bullets overlap. For adjudication, the stricter written requirement controls: every shifted exact-packet cohort must reach at least 95%. The executable's >=90% shift_packet field is only a coarse process exit condition for dependency dispatch and cannot override this scientific floor. Follow-on seeds remain scheduler-held until the report is checked against the stricter threshold.

Failure closes this raw-source compiler family. Passing authorizes one end-to-end DIVERGE packet/recovery evaluation, not long pretraining.

6. Accounting

Report complete parameters, trainable parameters, charged source tokens, candidate record/option counts, structured-decoder operations, peak activation memory, wall time, and packet bytes. Dynamic-programming work and temporary candidate storage count even when they add no parameters.

7. Novelty boundary

Occurrence ledgers, nominal equality, segmental parsing, structured prediction, object files, and hard packets are not individually novel. The bounded candidate is their source-sealed use as a language-to-factorized-DIVERGE compiler that preserves physical occurrences while committing one coherent complete packet. Any claim must remain at that conjunction and only after the matched neural gate.

8. CPU result

The calibrated CPU gate passes. Six focused unit tests cover exact/reference parity, local-versus-joint separation, noncommuting action order, repeated nominal identities, alpha renaming, post-seal poisoning, and record/option nonoverlap. The final seed 202608056100 board contains 1,000 episodes, 250 in each of the four frozen cohorts.

The decoder considered a mean 59.74 complete options and 305.705 complete records per episode. The retained pre-freeze 1,024-candidate calibration is 999/1,000 exact because one 452-token, nine-record composition episode has 1,124 valid scored record proposals and fails closed. The frozen 4,096 cap admits it without changing any score, data, decoder objective, or gate.

Calibration/final report SHA-256 values are 61dd90839b534f1e934d5bc3a9f2c7e88b1a99be8bc7b84fa65a85de5d48c4c2 / 1626e3fe5fbba89203bf76c5368da8fe5a847d398cec4d2dde02bbadfaa031f0. This establishes exact and nontrivial structured mechanics under synthetic scores only. It authorizes one neural raw-source seed; it is not learned language evidence.

9. Reconciliation with the external architecture proposal

The external proposal's central prescription is retained: premature certainty must be structurally avoidable, incompatible worlds must remain coherent, and commitment must be licensed by evidence or query invariance. Its individual ingredients are not treated as novel. Version spaces, guarded execution, particles, conflict clauses, typed state, and source deletion all have prior art.

The empirical record narrows the proposal in two ways:

1. Resource-corrected DIVERGE-v4 misses its frozen sharing gate at four worlds (1.893x, below 2x). The current claim is therefore restricted to the observed high-ambiguity regime, where sharing first clears the threshold at eight worlds and grows to 27.365x at 64 worlds.
2. The protected token-role compiler was not autonomous because it received gold record/option segmentation. SC1 is the first bounded test of whether a single raw-source pass can construct the coherent packet. A pass qualifies this source boundary; a failure leaves DIVERGE as exact synthetic mechanics without a learned raw-language front end.

This reconciliation also supersedes the review's original FCPT recommendation. PCSD and FCPT remain closed by their measured negatives. Full particles remain the principal matched scaffold/control, while CSDC's protected role-copy result is not replaced or weakened.

10. Neural result

Newton job 742328 completed all 1,200 updates / 9,600 charged episodes, wrote its immutable report and checkpoint, and exited nonzero because the capability gate failed. Training itself was stable and deceptively strong: final role, boundary, and pair accuracies were 99.999994%, 100%, and 98.381054%; loss was 0.061805; throughput was 85.294 episodes/s; peak allocated GPU memory was 796,946,432 bytes.

Autonomous complete-object decoding failed decisively:

Alpha-renamed exactness, independent decode, pair-disabled decode, and boundary-shuffled decode are all zero. Post-seal source-poison invariance is 100%, but that is the only passing executable gate. The four held follow-on seeds 742329--742332 were canceled without allocation.

The immediate diagnosis is combinatorial error amplification, not insufficient local fit: a small false-positive rate in a quadratic pair graph creates enough plausible complete records to erase support or hit the hard proposal cap. One no-gradient audit may substitute the already-accepted CPU oracle role, boundary, and pair scores componentwise to localize the necessary failure. It may not search thresholds, update weights, or reopen SC1.

Report/checkpoint SHA-256 values are 1a23d1aaae3276d54ec8d27abea266b822b0c9f28a058951dd2d942108d59059 / 7b5348cacb1772bf45e34442e94010db71a6be20bd8d689477d037ac5fee2ffd.

11. Read-only failure localization

Job 742457 evaluated 128 frozen episodes (32 per cohort) without gradients, threshold search, or parameter changes. It substituted the accepted CPU calibrated role, boundary, and pair factors into the failed checkpoint one component at a time.

The causal result is exact:

The learned boundary is exact over this audit: 1,494 true positives, 34,064 true negatives, and zero false positives or misses. The pair head has 4,697 true positives, 7,942 false positives, 1,263 misses, and 445,271 true negatives: 37.162% precision and 78.809% recall despite high aggregate accuracy. Active-role detection has zero false positives and 98.873% recall, but record-kind cues and alias starts are confused with other active roles.

Those errors amplify combinatorially. Learned decoding creates about 111--140 option candidates per episode versus 54--63 under calibrated factors, and 1,976--2,875 complete-record candidates versus 273--328: roughly 2x options and 7.2--9.4x records. This explains both support deletion and the hard-cap overflows.

The conclusion is architectural: independent dense edge classification plus thresholded Cartesian proposal generation is not a viable source boundary for support-preserving reasoning. Any future, separately frozen successor must construct bounded whole-record assignments by design (for example, a boundary-first typed matching object with exact-one constraints). It may not be presented as a threshold, seed, width, duration, or loss repair of SC1.

Audit report SHA-256 is e6dd4874029f653f808c426d04b15a98da390c67c83b30429e9d015b00ab9799.