docs/methodology.md

# Methodology

## Measurement contract

Freeze the following before comparing candidates:

- model artifact and revision;
- tokenizer and chat-template hashes;
- runtime version and request adapter;
- context window and output reserve;
- sampler, reasoning mode, stop behavior, and seed;
- task configuration and holdout policy;
- critical zero-regression cases;
- minimum performance improvement;
- hardware and cache condition;
- KV-cache precision and quantization policy;
- concurrent workloads and runtime features that materially affect memory.

Change one declared variable at a time. Compare a quantized candidate with its exact higher-precision parent, and compare MTPLX depths with AR using paired cases.

Classify allocation failures, host watchdogs, and out-of-memory terminations as operability outcomes rather than incorrect model answers. A maximum-window result is valid only for the reported hardware and runtime contract. If KV-cache precision changes, treat it as a separate candidate and rerun the quality, context, performance, and stability gates.

## Long-context construction

Lengths in the frozen release grid are 4,096, 8,192, 16,384, 32,768, 65,536, 131,072, 196,608, 229,376, 245,760, and 260,096 rendered prompt tokens. The last length reserves 2,048 tokens inside a 262,144-token model context.

The routine quant-comparison gate stops at 131,072 prompt tokens. The single 260,096-token capacity sentinel lives in `configs/context-max-window-v1.json` so it cannot accidentally turn a lightweight comparison into a host-stability test. Use it on a sufficiently provisioned host or when explicitly qualifying a q8/q4 KV-cache runtime candidate.

Evidence centers are placed at 1%, 10%, 25%, 50%, 75%, 90%, and 99% of the context body. Each prompt records requested and actual evidence positions. Start and end sentinels must both appear in the answer, causing silent front or tail truncation to fail.

The task families are:

1. exact single-key retrieval;
2. key/value binding among similar identifiers;
3. current-record selection among stale and test-only records;
4. multi-hop resolution across distant evidence;
5. semantic retrieval with reduced lexical overlap;
6. ordered state updates.

The release configuration also compares queries placed before and after the context. Filler is procedurally generated synthetic telemetry with unique traces and semantically similar records, not a tiny repeated word vocabulary.

## Scoring

Structured tasks require bare, fully parseable JSON. Markdown fences and trailing prose fail. Tool cases compare the emitted protocol calls, names, arguments, order, and no-call behavior. Executable Python is statically screened and evaluated in a resource-limited isolated subprocess.

The Python evaluator is appropriate for trusted local benchmark output, not hostile multi-tenant code. A public submission service must add an operating-system or container sandbox with networking disabled.

## Agentic execution profiles

The default `baseline` profile sends the scenario system prompt, task, tool schemas, and tool results without an external execution layer. `shiftedx-harness-v1` is the first Shiftedx Agent Harness profile and is an opt-in paired treatment; it must never replace baseline results in a model-quality scorecard.

The harness operates only on observable request and tool-result state. It records success or failure receipts, scopes duplicate suppression to a state epoch, advances that epoch after a successful mutation, requires a later successful verification receipt, and validates only the task's declared bare-JSON output shape. Hidden expected answers and grader-required tool sets are not harness inputs. At most two terminal correction turns are allowed.

Declared output contracts include both key names and JSON primitive types. Grounded receipt projection may eliminate a final model turn only when a successful structured receipt itself contains the full declared schema, or when a successful verifier reports an exact `N passed` result for a declared `{status, tests}` response. Projection is recorded in telemetry and may not supply an unstated value.

The `expanded` scenario set contains 30 cases across repair loops, stale evidence, structured status, wrong-path recovery, parallel reads, and underspecified destructive-action refusal. The `repo` set creates four fresh temporary source trees, permits writes only to the named source module, statically screens replacement code, and executes visible plus hidden assertions through the resource-limited Python sandbox. Hidden assertions are never included in prompts or controller state. Disposable-repository results are still small-fixture evidence, not SWE-bench claims.

For a valid comparison, keep model artifact, server, sampler, reasoning mode, tools, cases, and case order fixed. Report both emitted and actually dispatched tool calls because duplicate suppression can reduce external side effects without reducing the model's attempted calls. Also report correction turns, total prompt/completion tokens, wall time, and per-turn decode telemetry. Treat higher accuracy bought through unbounded retries or materially larger token budgets as a different operating point, not a free quality improvement.

## Statistical reporting

Report each lane separately. Required outputs include:

- accuracy and a bootstrap 95% confidence interval;
- context accuracy by length, position, and family;
- worst context cell and effective context length at a declared threshold;
- wall time, time to first token when streaming is supported, and end-to-end token rate;
- server-reported cached tokens and MTPLX acceptance telemetry when available;
- paired quality delta and speed ratio for every runtime variant.

Do not infer an intelligence improvement from one or two task changes. Do not combine hardware-dependent throughput with model quality in a single score.

## Public and rotating sets

The checked-in configurations are reproducible development contracts. For a leaderboard release, derive a rotating seed configuration from a private master key, freeze its hash before evaluation, and publish the concrete seeds only after submissions close. Keep calibration prompts, development seeds, and holdout seeds disjoint.

## Sanity controls

Before accepting a new suite version, verify:

- an oracle response passes every case;
- empty, echo, random, stale-decoy, and truncated responses fail;
- deliberately mutated code fails hidden tests;
- duplicate case identifiers are rejected;
- prompt counts and evidence positions are exact under a real target tokenizer;
- a server/tokenizer mismatch fails rather than silently changing the score.