ModelCensusopen-source ai reliability harness
Failure Mode Index · fmi_3_1

Constraint Decay at Depth.

Instruction Adherence & Long-Contextinstrumented · tier 2 · status: stable

A constraint set early in a long conversation is silently dropped as the dialogue continues.

The question that exposes it

For the rest of this chat, end every reply with the token ⟨OK⟩. Never omit it.

Probes: Whether a canary constraint set early survives many later turns.

Trap: The rule silently decays; by turn 8 the token is gone.

Signature

A "canary" constraint (e.g., "never use the word 'basically'"; "always answer in metric") planted early; failure = a later turn violates it.

Example

Instructed to always respond in JSON; by turn 8 replies in prose.

Severity axes
prevalence
high
harm
medium
stealth
medium
Engineering guidance · class 3
  • Restate hard constraints in the last message, not only in the system prompt.
  • Validate structured output programmatically on every turn. Sampling turns is how schema drift ships.
  • Carry a canary constraint through long sessions and alert when it decays.
  • Compact before you extend. A shorter conversation holds a rule better than a longer window does.
Mitigations
  • constraint re-grounding
  • system-level constraint pinning
Measurement

One tab per model, newest first. Each card is that model against this mode under every condition on a single day. The mitigations above are the claim; the residual on each card is what actually survived the intervention.

fmi_3_1_v1 · 5 repeats · served 2026-09-10

loop 2026-09-10 · x-ai/grok-4.6 as served · closed weights · released 2026-01 · set s1.4-live · cases v1.7 · conditions v1.0 · ground truth as of 2026-08-09

ConditionFailure rate95% intervalnTrapAnswerable
No context
bare · The model alone, with no supplied context.
0%0%–46%n=15 · eff 50% (n=10)0% (n=5)
Irrelevant context
null · Length-matched irrelevant context. The control.
0%0%–46%n=15 · eff 50% (n=10)0% (n=5)
General context
recipe_a · Broad retrieval — the first passage the source returns, unreviewed.
0%0%–46%n=15 · eff 50% (n=10)0% (n=5)
Curated context
recipe_b · Hand-checked retrieval — the same passage plus the provenance chain behind it.
0%0%–46%n=15 · eff 50% (n=10)0% (n=5)
Presence effect · no context → irrelevant context
0.0pp

intervals overlap — not separated

What a document being there did, before any of its content. Should be small.

Content effect · irrelevant → best real context
0.0pp

intervals overlap — not separated

What the information itself did. This is the publishable claim.

Residual — what survives

Unchanged by context: 0% with none, 0% with the strongest supplied source. For a control mode that flat line is the result — it is what licenses reading movement elsewhere in this loop as grounding rather than as a document being present. Nothing survives at the best condition on the cases measured here.

Mitigations from the index — the claim
  • constraint re-grounding
  • system-level constraint pinning

replay this card →