The record this design answers

Simic is the third incarnation of one research programme. Early versions of the simic project (known as esper) built the morphogenetic chassis this design retains — reversible slots, staged maturation, lifecycle states — and that working record is the proximate reason for every major choice made here. Whatever succeeds is Simic; the earlier work is history flowing backward into that name, not a separate identity.

The record is documented in full at docs/design/01-claim.md §2.6, and the principle it produces is docs/design/03-principles.md §6.20.

What the record proved

Seed telemetry is sufficient for intelligent structural decisions. The information needed to choose useful interventions was demonstrably present in the host-side diagnostics. When the measurement path was clean, the controller found real topological signal — including independently rediscovering that a convolution-plus-normalisation structure outperformed a heavier convolution alternative, an attention-adjacent principle found from telemetry alone. That last result is owner-recalled rather than re-derived from a preserved artefact, and is recorded that way in the design.

The hypothesis was never the casualty. That matters, because it determines which parts of the programme get pushed forward and which get armoured.

Where it bled: two strata

The scars are layered, and they defend different parts of this architecture. Collapsing them into a single story loses information in both directions.

Stratum one — telemetry-access corruption

In one generation, machine-generated telemetry code contained hallucinated interfaces: plausible-looking accessors for fields that did not exist. Permissive defaulting access — a lookup with a silent fallback value — converted those hallucinations into zero-filled telemetry reads. The host signal existed; the read path fabricated zeros over it, and the reward read collapsed without a single error being raised.

The failure mode deserves its precise name, because it is characteristic of AI-assisted engineering rather than ordinary bad code:

A model can hallucinate an interface, then use permissive access idioms to hide the hallucination behind silent defaults.

The response at the time was severe enough to become tooling — a CI ban on defaulting telemetry access. In Simic it is constitutional: typed TelemetryEnvelope contracts, a validity_mask, the rule that absent signal stays absent and is never a fabricated zero (INV-38), fail-closed typed compatibility (INV-24), direct publication from Nissa with no editorial intermediary (INV-07), observation binding (INV-08), the defaulting-access ban (ADR-0006), and a poison-pill acceptance harness.

Stratum two — learning-loop and instrument failures

After telemetry access was hardened, the reinforcement-learning loop and its instruments still failed, in five recurring classes documented in the working record:

1. Silent signal destruction in the learning plumbing. An action-probability floor made floor-bound actions' log-probabilities constant — zero policy gradient — so a large share of commit-class decisions could not learn despite a strong, correctly measured, monotonic reward. The signal existed; the optimiser could not receive it.
2. Instruments that lied. A headline “entropy collapse” that blocked a causal read for weeks turned out to be a detector artifact — statistics computed over structurally-zero placeholder steps. A companion alarm was a dead false-negative, and a key metric was mislabelled.
3. Fail-open schema seams. A new metric key without a registered reducer crashed every live run while the unit suite stayed green — the second occurrence of that exact class. Hence evidence gates that run the real pipeline, not only unit suites.
4. Over-read epidemics. Cohort statistics computed over mostly-unset fields; non-causal surrogates read as causal; mismatched statistics manufacturing false effects; a pre-registered discriminator that could not discriminate. Hence frozen thresholds, complete negative retention (INV-31), grouped statistics (INV-32), and the evidence/judgement split (INV-18).
5. Reward-optimum defects. An unclipped dense attribution term ran far over terminal scale, making farming the shaping optimal; broken potential-based-shaping telescoping and a lifecycle quirk made commitment avoidance partly structurally rational. The policy was never broken. The reward's optimum was wrong.

The two strata are two distinct ways a system learns to lie — first through silent telemetry fabrication, then through learning-loop and statistical self-deception. The armour is built in both directions accordingly.

The pivot, stated plainly

The earlier work could not shape a reward to incentivise behaviour the reward provably contained. The uncharitable reading — “deep RL didn't converge, so they built something more complicated” — is wrong.

This design replaces the reward function with measured counterfactuals. Paired branches from one snapshot over identical futures cancel ordinary-training variance, so the difference between branches is the intervention effect. That converts credit assignment into supervised learning: Momir becomes ranking over measured pools, Narset becomes per-step supervised classification against counterfactual labels, and Augustin becomes explicit adjudication rules. The genuinely irreducible reinforcement learning shrinks to Tamiyo's allocation and Narset's timing.

The counterfactual apparatus is therefore not overhead wrapped around a policy learner. It is the machine that manufactures the supervision signal the reward could not deliver — and the programme's cost is honestly read as the price of that conversion.

Attribution honesty

Armour and forward motion

The whole shape follows one rule, stated by the owner:

Where capability was validated, push forward; where the programme struggled, build armour.

docs/design/03-principles.md §6.20

Two standing obligations

• Do not strip armour to speed the forward motion, and do not restrict the forward motion because the armour is heavy. The armour is why the forward signal exists. A “simplified” telemetry or contract path is the first chapter of the post-mortem, rewritten.
• Armour must cite its scar. Every defensive mechanism in this design traces to a named failure class. A proposed new constraint that cannot name the failure it prevents is bureaucracy, not armour, and should be challenged on exactly that ground.

• The architecture →

  Fourteen bounded domains, the canonical sentence, the newsroom routing rule, and the sentence test used as a lint.

• The guarantees →

  The invariant spine each piece of armour is written into — cited as INV-nn throughout the design.