probity. evidence before compression.

probity is the light: it seals the source identity and makes measured truth inspectable. inspector gadget is the instrument stack. forge acts only on verified evidence, and refuses to guess where the evidence goes dark.

4 models indexed · 32 content-addressed claims · 1.04T parameters in catalog
the optrenium evidence chain

probity is the light. forge refuses the dark.

one source crosses four accountable boundaries: content-addressed truth → instrumented measurement → read-only planning and bounded calibration → transactional verification.

01 recorded in dossier

probity · measured + signed truth

probity opens the gguf read-only, inventories every tensor, and signs canonical claim payloads into content-addressed ids bound to the source fingerprints.

source artifacts
4
claim ids
32
parameters
1.04T

“signed” means deterministic content integrity here, not a certificate-backed author signature.

02 measured by ig

ig · the instrument stack

registered inspector gadget stages audit tensors, attention heads, experts, outliers, spectra, numerical hygiene, and stage lineage.

tensor audits
4202
head records
8576
expert records
45568
ig stages
1002
static findings
254

success-only stage records preserve index, class, kind, energy, entropy, and output shape.

03 separate forge artifact

forge · fit first, measure next

forge can fit a cloned plan to a complete-file, tensor-payload, or hardware ceiling without opening or writing a gguf. it then measures native codec reconstruction per eligible tensor across deterministic seeds. missing evidence stops a decision instead of becoming a guess.

  • audited non-tensor overhead plus output reserve
  • separate runtime reserve and safety margin
  • content-bound plan hashes and ordered ig moves
  • reconstruction and behavior remain later gates

plan-fit emits only a plan and receipt. it writes no gguf and claims neither reconstruction calibration nor behavioral quality.

04 required release gate

refuse darkness · verify

the plan must match the source sha-256. output is written transactionally, re-opened, structurally checked, hashed, and re-audited before release.

  • source clone remains immutable
  • metadata and tensor inventory verified
  • reconstruction evidence stays labeled static
  • behavioral evaluation remains a separate gate
measured

inside this dossier

bytes, metadata, tensors, heads, experts, tokenizer, anomalies, inspector gadget lineage, fingerprints, and claim ids.

derived

inside forge planning + calibration

content-bound fit receipts, complete-file accounting, ordered inspector gadget moves, sampled reconstruction error, deterministic seed envelopes, and per-role precision selection.

not inferred

behavioral quality

coherence, perplexity, task accuracy, runtime stability, and application fit require an allowed inference evaluation.

how to read “findings”: a static anomaly is a review signal, not automatic corruption. for example, the fresh qwen3.6 35b-a3b dossier records 12 ig structural findings, 1 medium and 11 low, while reporting zero tokenizer-source findings, zero tensor-source findings, and zero nan/inf values.
tier
family
shape
0 shown

GLM 5.2 BF16

advisory · 76 moe · 256/layer split · 33 shards
glm-dsa ·256x22B ·mit · arch fp 752058f5f674ae82…
753.86B
parameters
42.38B active/token
tensors
1809
blocks
79
quant
BF16 +1
static findings
214
214 low
412 ig stages 5056 head records 19456 expert records 214 ig / structural findings
0 tensor-source 0 tokenizer-source 0 nan/inf tensors findings are warnings unless an integrity field says otherwise
9 hampel outliers 594 outlier-channel tensors (worst 637.8×)
9 content-addressed claims · audit v0.0.8 · 1.37 TiB across 33 shards 2026-07-16 17:54 utc

Qwen3.6-35B-A3B Q8_0

clean · 86 moe · 256/layer
qwen35moe ·35B-A3B · arch fp a9d345561e1112b2…
34.66B
parameters
3.45B active/token
tensors
733
blocks
40
quant
Q8_0 +1
static findings
12
1 medium 11 low
179 ig stages 160 head records 10240 expert records 12 ig / structural findings
0 tensor-source 0 tokenizer-source 0 nan/inf tensors findings are warnings unless an integrity field says otherwise
1 collapsed experts 8 hampel outliers 154 outlier-channel tensors (worst 28.7×)
8 content-addressed claims · audit v0.0.6 · 34.37 GiB 2026-07-16 05:51 utc

MiniMax-M2.7-Q8_0

clean · 86 moe · 256/layer split · 7 shards huggingface.co
minimax-m2 ·256x4.9B · arch fp 8d3e148b5c6f1ce1…
228.69B
parameters
11.03B active/token
tensors
809
blocks
62
quant
Q8_0 +1
static findings
9
9 low
199 ig stages 2976 head records 15872 expert records 9 ig / structural findings
0 tensor-source 0 tokenizer-source 0 nan/inf tensors findings are warnings unless an integrity field says otherwise
6 hampel outliers 61 outlier-channel tensors (worst 21.5×)
8 content-addressed claims · audit v0.0.5 · 226.44 GiB across 7 shards 2026-07-16 01:37 utc

Qwen3.6-27B-Q8_0

clean · 87 huggingface.co
qwen35 ·27B · arch fp 98a0dc77c78ec13d…
26.90B
parameters
tensors
851
blocks
64
quant
Q8_0 +1
static findings
19
19 low
212 ig stages 384 head records 0 expert records 19 ig / structural findings
0 tensor-source 0 tokenizer-source 0 nan/inf tensors findings are warnings unless an integrity field says otherwise
6 hampel outliers 222 outlier-channel tensors (worst 25.5×)
7 content-addressed claims · audit v0.0.5 · 26.63 GiB 2026-07-16 00:46 utc

every audit on this page was generated by probity, optrenium.ai's proprietary model-audit platform. each claim is content-addressed; re-running the same pinned probity/operator build against the same gguf bytes should reproduce its claim id exactly.

open-source gguf models are audited and the reports hosted on this site free of charge*.

* probity itself is a commercial product. private & closed-source providers can license probity for in-house use: licensing@optrenium.ai

glossary · hover any underlined term for its definition