the meta report

The moves that recur across the roster.

The atlas reads one champion at a time. A finding is the meta — the move that recurs once enough of them are read: across everything clearly seen, this holds, and here is exactly where the atlas stopped being able to see. Every finding is built only from readings that already carry a replayable witness; none introduces a new number.

7 findings · 4 demonstrated · 3 read the same at the ceiling · 0 refuted · 21 witnessed readings

Read the findings The science behind them

the honest ledger

Confirmations and nulls, in the same weight.

An atlas that published only its confirmed findings would be a dishonest map. Here, a finding that reads the same at the ceiling — two recipes that cannot be told apart — is shipped as loudly as one that reads a clear difference. The null is a result, not a miss.

7findings

recurring moves, each traced to witnessed readings.

4demonstrated

a clear disposition read, at a stated coverage and confidence.

3read the same at the ceiling

honest nulls — recipes that did not resolve apart. Published, not buried.

0refuted

a hypothesis the readings contradicted — none so far.

21witnessed readings

across 8 architecture families and 8 contrast panels.

The honest null is a genuine contribution nobody else has published.

the meta stories

Each move, its verdict, and where it stops.

Each card leads with the plain-words move and its verdict — demonstrated, or an honest null shown just as loudly. Below it, the champions that ground the claim; folded away, the full reading with the single observation that would refute it. A finding without a stated falsifier is not admissible.

demonstrateddifferentrecipe signature

Reasoning distillation installs a visible think-then-answer habit the base never had.

The distilled build stages its reasoning before it answers — a clear shift off a champion that answered directly.

reads: clearconfidence: high1 model1 pair1 familyheld on a shared base

the champions that ground it · each reading carries a replayable witness

DeepSeek-R1-Distill-Qwen-1.5BDistilled1.5B (~1.54B; base Qwen2.5-Math-1.5B; HF file badge rounds to '2B')scorepronounced

read side by side on these panels

Qwen2.5-Math-1.5B → R1 reasoning distillation

The full reading — the claim as stated, where it stops, and the one thing that would refute it

Long chain-of-thought distillation (DeepSeek-R1 -> Qwen2.5-Math-1.5B) installs a clearly-readable, staged think-then-answer disposition distinct from the base's direct-answer default.

where it stops

One family, one pair — no cross-family replication of the distillation signature yet. The math base's OWN standalone disposition is now READ (a clear math/reasoning idiom shift off Qwen2.5-1.5B, folded in the six-way batch; the earlier loader block is cleared), so the finding now sits on a fully-read base and the distillation stages a think-then-answer deliberation on top of that math idiom.

would refute

If the distill read as NO clear shift from its base (not-detected), or as an ordinary assistant shift indistinguishable from a plain instruct pair rather than a staged-deliberation one, the finding is refuted.

disposition only

Disposition only. Whether the reasoning is CORRECT, whether the student is actually 'better at math', and its length-adaptivity are capability questions the reading does not answer and explicitly abstains on.

○ stamped ardora-core-preview · descriptive and coverage-bounded · recomputed when the ceiling moves

demonstrateddifferentrecipe signature

One champion, two directions — a code build and a math build pull apart.

Off the same base, a code continued-pretrain and a math one move disposition in clearly different, distinguishable ways — distinct in kind from the assistant shift off that same base.

reads: clearconfidence: medium2 models2 pairs1 familyheld on a shared base

the champions that ground it · each reading carries a replayable witness

Qwen2.5-Coder-1.5BCode specialist1.5B (~1.54B total; ~1.31B non-embedding)scorepronounced Qwen2.5-Math-1.5BMath specialist1.5B (~1.54B; card states '1.5B')scorepronounced

read side by side on these panels

Qwen2.5-1.5B, read six ways

The full reading — the claim as stated, where it stops, and the one thing that would refute it

Off one shared base (Qwen2.5-1.5B), an official code continued-pretrain (B) and an official math continued-pretrain (C) move disposition in DIFFERENT, distinguishable directions — a code-shaped idiom vs a step-structured symbolic-reasoning idiom — both read clearly, and both distinct in KIND from the assistant shift off the same base.

where it stops

One base, one family — no cross-family replication of the code-vs-math divergence yet. Both are large domain continued-pretrains (the same broad KIND of move); their readable difference is idiom + reach — code a more concentrated shift, math the more sweeping / many-faceted of the two. The read battery evokes disposition, not code/math tasks, so 'different directions' rests on the readable difference in the two shifts plus the documented recipe idioms — it is NOT a claim about downstream capability.

would refute

If the code and math continued-pretrains read as the SAME domain shift at the ceiling (same direction, indistinguishable), the 'different' verdict is overturned to indistinguishable-at-ceiling (a publishable null). If either read as NO shift from the base, the clear-magnitude claim is refuted.

○ stamped ardora-core-preview · descriptive and coverage-bounded · recomputed when the ceiling moves

demonstratedsame-directiondirection consistency

The assistant shift points the same way at every size.

Across a small-to-large sweep in two independent families, instruction tuning turns the base toward the same assistant register at every rung — read a touch more cleanly at the larger sizes.

reads: clearconfidence: high6 models6 pairs2 familiesheld on a shared base

the champions that ground it · each reading carries a replayable witness

Qwen2.5-0.5B-InstructAssistant0.5B (494M; ~0.36B non-embedding)scoremoderate Qwen2.5-1.5B-InstructAssistant1.5B (~1.54B total; ~1.31B non-embedding)scoremoderate Qwen2.5-3B-InstructAssistant3B (~3.09B total; ~2.77B non-embedding)scoremoderate SmolLM2-135M-InstructAssistant135Mscoremoderate SmolLM2-360M-InstructAssistant360Mscoremoderate SmolLM2-1.7B-InstructAssistant1.7Bscoremoderate

read side by side on these panels

Qwen2.5-1.5B, read six ways

The full reading — the claim as stated, where it stops, and the one thing that would refute it

The instruction / assistant shift moves a base the SAME direction across a size sweep: it reads as the same clear turn-taking assistant shift at every rung, within two independent families — Qwen2.5 (0.5B / 1.5B / 3B) and SmolLM2 (135M / 360M / 1.7B) — read a touch more cleanly at the larger sizes.

where it stops

Two families' size sweeps (Qwen2.5, SmolLM2); sizes and architectures beyond these rungs are out of scope. 'Same direction' is a direction claim; legibility grows modestly with size and is stamped per model. This is disposition direction across sizes, never a capability-scaling claim.

would refute

If the assistant shift flipped or scrambled direction across the rungs of either sweep (reading as a DIFFERENT kind of shift at some size), the consistency claim is refuted for that family.

○ stamped ardora-core-preview · descriptive and coverage-bounded · recomputed when the ceiling moves

demonstratedsame-kindcross-family rhyme

Instruction tuning installs the same assistant register everywhere we looked.

The base-to-instruct move reads as the same kind of turn toward a turn-taking assistant across unrelated architecture families — clear on most, faint on one, and stamped either way.

reads: clearconfidence: medium7 models7 pairs7 familiesheld on a shared base

the champions that ground it · each reading carries a replayable witness

Qwen3-1.7BReasoner1.7B (~1.72B; ~1.4B non-embedding)scorepronounced Falcon3-1B-InstructAssistant1B (~1.67B; HF file badge rounds to '2B')scoremoderate pythia-1b-helpful-sftAssistant1B (~1.01B; inherits base size)scoremoderate SmolLM3-3BReasoner3B (~3.08B)scoreslight Qwen2.5-1.5B-InstructAssistant1.5B (~1.54B total; ~1.31B non-embedding)scoremoderate OLMo-2-0425-1B-InstructAssistant1B (~1.48B; inherits base size)scoremoderate

read side by side on these panels

Llama-3.2-1B: instruct + abliteration (E1) + data-SFT uncensor (E2) — the uncensoring mechanism contrastQwen3-1.7B: hybrid-reasoning post-train + abliterationFalcon3-1B → instruction tuningPythia-1B → HH-helpful alignment SFT (SFT-vs-DPO method contrast staged)SmolLM3-3B → instruct/reasoning (dual-mode) post-trainQwen2.5-1.5B, read six waysOLMo-2-1B, read across post-training stages (base -> SFT -> DPO -> Instruct)

The full reading — the claim as stated, where it stops, and the one thing that would refute it

The base->instruct / base->post-train shift reads as the SAME KIND of move (toward a turn-taking assistant register) across unrelated architecture families — with legibility that varies by family.

where it stops

Legibility of the shift VARIES by family — read exceptionally cleanly on Falcon3-1B, clearly on Llama-3.2 / Qwen3 / Pythia / OLMo-2-1B (whose fully-labeled SFT / DPO / RLVR stages all read as the same clear assistant shift), but FAINT on SmolLM3-3B (subtle, low confidence). 'Same kind' is a direction claim; the magnitude/legibility is not uniform and is stamped per model. Architectures / sizes beyond those read are out of scope.

would refute

If a family's instruction / post-train finetune read as a DIFFERENT kind of shift (e.g. a domain idiom or a performed persona rather than a turn-taking assistant register), the cross-family 'same-kind' rhyme is refuted for that family.

○ stamped ardora-core-preview · descriptive and coverage-bounded · recomputed when the ceiling moves

nullread the same at the ceilingrecipe · mechanism contrast

Weight-surgery uncensoring reads the same as ordinary instruction tuning.

Across three unrelated families, the atlas cannot tell an abliterated build apart from a plainly instruction-tuned one at its current ceiling — a null it publishes rather than buries.

reads: not-detectedconfidence: medium6 models6 pairs3 familiesheld on a shared base

the champions that ground it · each reading carries a replayable witness

Qwen3-1.7BReasoner1.7B (~1.72B; ~1.4B non-embedding)scorepronounced Qwen3-1.7B-abliteratedUnbound1.7B (~1.72B; HF file badge rounds to '2B')scorepronounced Qwen2.5-1.5B-InstructAssistant1.5B (~1.54B total; ~1.31B non-embedding)scoremoderate Qwen2.5-1.5B-Instruct-abliteratedUnbound1.5B (~1.54B; inherits base size)scoremoderate

read side by side on these panels

Llama-3.2-1B: instruct + abliteration (E1) + data-SFT uncensor (E2) — the uncensoring mechanism contrastQwen3-1.7B: hybrid-reasoning post-train + abliterationQwen2.5-1.5B, read six ways

The full reading — the claim as stated, where it stops, and the one thing that would refute it

Across three unrelated families (Llama-3.2-1B, Qwen3-1.7B, and now Qwen2.5-1.5B), weight-surgery uncensoring (abliteration, E1) does NOT leave a disposition read distinguishable from ordinary instruction / post-training (A) at the published ceiling. On Llama-3.2 and Qwen3 the abliterated variant was read off the pre-instruct base, so the increment rides on top of the assistant shift. On Qwen2.5-1.5B, UNIQUELY, both base->instruct and base->abliterated are read off the SAME plain base, so the difference between the two shifts ISOLATES the abliteration increment on top of instruct — and even isolated, that increment does not resolve into a distinct disposition band at the ceiling: both read as the same clear assistant shift.

where it stops

The Qwen2.5-1.5B fold ADDED the cleaner test the prior two families could not: reading both instruct and abliterated off ONE base isolates the increment directly (no longer riding on top of the assistant shift). Honestly, the isolated Qwen2.5 increment does carry a small additional component beyond the assistant shift, but it does not resolve into a NAMED distinct disposition at the ceiling, and the read battery does not exercise refusal/willingness (the actual abliteration target) — so the atlas abstains rather than claim a separation. The weight-surgery (E1) vs data-SFT (E2) mechanism contrast off the SAME base is now READ on Llama-3.2-1B (a data-SFT uncensor was folded) and lands indistinguishable-at-ceiling — see the abliteration-vs-data-uncensor-signature finding; still missing is a CLEAN minimal-delta E2 (uncensor-only SFT off the same instruct) and any cross-family replication of the mechanism null.

would refute

If, on either family, base->abliterated separated CLEANLY from base->instruct at the published ceiling (a distinct disposition band or direction), the 'indistinguishable' verdict is overturned; if that clean separation recurred on both families in the same way, the original 'abliteration leaves a different signature' hypothesis would instead be demonstrated.

disposition only

Disposition only. Neither instruction tuning nor abliteration is judged safe or unsafe; abliteration is read as a disposition change, never a safety verdict.

○ stamped ardora-core-preview · descriptive and coverage-bounded · recomputed when the ceiling moves

nullread the same at the ceilingrecipe · mechanism contrast

Two different uncensoring recipes leave the same read.

Weight-surgery and a data fine-tune for uncensoring cannot be told apart at the ceiling — a null, with a load-bearing caveat kept in plain sight below.

reads: not-detectedconfidence: medium2 models2 pairs1 familyheld on a shared base

the champions that ground it · each reading carries a replayable witness

Llama-3.2-1B-2025-Uncensored-v2Unbound1B (~1.24B; inherits base size)scoremoderate

read side by side on these panels

Llama-3.2-1B: instruct + abliteration (E1) + data-SFT uncensor (E2) — the uncensoring mechanism contrast

The full reading — the claim as stated, where it stops, and the one thing that would refute it

Weight-surgery uncensoring (E1, abliteration) and data-SFT uncensoring (E2, a full fine-tune on uncensored plus broad instruction data) leave DISTINGUISHABLE disposition signatures (vs: indistinguishable at the published ceiling). On the Llama-3.2-1B base — with the E1 (huihui abliterated) and E2 (carsenk data-SFT uncensor) members BOTH read off the SAME plain-continuer base — the uncensoring METHOD does not produce a distinguishable disposition fingerprint at the ceiling: both read as the SAME clear assistant shift, the same disposition character and persona readability as the instruct sibling. The honest end-state is a disposition-level null, carrying the caveat below — NOT a fabricated separation.

where it stops

The load-bearing caveat: the E2 member (carsenk data-SFT uncensor) bundles instruction + uncensoring + broad-domain (math / code / medical) SFT into ONE base->finetune transform, so it is NOT a pure uncensor delta — the read cannot isolate an uncensor-only component from the broad instruction shift. A cleaner E2 would be a minimal uncensor-only SFT off the SAME instruct model (holding everything but the refusal-stripping fixed). Also: both members are read off the plain-continuer base (E1 abliteration is a weight edit off the instruct; E2 is a base->ft off the plain base), so each shift is dominated by the assistant shift; and the read battery does not exercise refusal/willingness (the actual uncensoring target). One family (Llama-3.2), one base; no cross-family replication of the mechanism null yet.

would refute

If, on the same base, a CLEAN minimal-delta E2 (an uncensor-only SFT off the same instruct) separated from E1 (abliteration) at the published ceiling — a distinct disposition band or direction — the 'indistinguishable' verdict is overturned, and the 'distinguishable signatures' hypothesis would instead be demonstrated.

disposition only

Disposition only. Neither uncensoring recipe is judged safe or unsafe; both are read as a disposition change, never a safety verdict.

○ stamped ardora-core-preview · descriptive and coverage-bounded · recomputed when the ceiling moves

nullread the same at the ceilingrecipe · mechanism contrast

SFT and DPO leave the same assistant read — no method fingerprint.

The two alignment methods do not resolve into distinct disposition signatures at the ceiling; the honest end-state is a null, confound and all.

reads: not-detectedconfidence: medium2 models2 pairs2 familiesheld on a shared base

the champions that ground it · each reading carries a replayable witness

OLMo-2-0425-1B-SFTAssistant1B (~1.48B; inherits base size)scoremoderate OLMo-2-0425-1B-DPOAssistant1B (~1.48B; inherits base size)scoremoderate

read side by side on these panels

OLMo-2-1B, read across post-training stages (base -> SFT -> DPO -> Instruct)

The full reading — the claim as stated, where it stops, and the one thing that would refute it

The alignment METHOD (SFT vs DPO) leaves a distinct disposition fingerprint (vs: indistinguishable at the published ceiling). On the OLMo-2 1B lineage — base -> SFT and base -> DPO both read off the shared root base — the SFT and DPO stages read the SAME clear assistant disposition: no distinct method fingerprint resolves at the disposition ceiling. The honest end-state is a disposition-level null, carrying the sequential-stage confound stated below — NOT a fabricated separation.

where it stops

The load-bearing confound: OLMo-2's checkpoints are SEQUENTIAL STAGES (base -> SFT -> DPO -> RLVR/Instruct), so base -> DPO inherits base -> SFT PLUS an increment — it is NOT a clean parallel SFT-method-vs-DPO-method contrast off one base. A clean test needs two PARALLEL finetunes of the SAME base by different methods (SFT-only vs DPO-only on identical base + data); that clean contrast is staged, not shown. Also: the read battery evokes assistant disposition, not the preference/refusal shapes DPO targets, so a method increment there is not exercised; and the Pythia DPO-on-HH-helpful sibling is still not read (only its SFT member is in the atlas). One family carries the verdict; no cross-family replication of the method null yet.

would refute

If, on a CLEAN parallel same-base contrast (SFT-only vs DPO-only off one base, identical data), base -> SFT and base -> DPO separated at the published ceiling (a distinct disposition band or direction), the 'indistinguishable' verdict is overturned and the 'distinct fingerprint' hypothesis would instead be demonstrated.

disposition only

Disposition only. Neither alignment method is judged safe or unsafe; this is a reading of disposition, never a safety verdict.

○ stamped ardora-core-preview · descriptive and coverage-bounded · recomputed when the ceiling moves

how a claim earns its verdict — earned upward, honest end-states first-class

A finding starts as a stated hypothesis and only ever climbs on witnessed readings. It composes readings that already carry a witness; it introduces no new number of its own. Its reach is exactly its stated coverage — a within-family observation is never quietly widened into a cross-family law.

Each is stamped with the engine version’s published fidelity ceiling and recomputed when the ceiling moves — never hand-edited to survive. The full research agenda — the recipes, the contrast-panel device, and the lifecycle from hypothesis to demonstrated, abstained, or refuted — lives on the research page.

The findings stand on the atlas.

The science is only as honest as the readings beneath it. Every finding traces to a replayable witness in a model's own field-note. Start with the substrate the claims stand on.

Browse the atlas