2026-03-27 - RES-009: The Alignment Tax and Structural Defense — What Response Homogenization Means for Human-in-the-Middle AI Teams

Corrected 2026-10-16. A second whole-paper review found several statements stronger than their sources allow, and they are narrowed: what the paper concludes from its decoding tests, what its significance tests cover, and how well token entropy does on factual questions. Present-tense claims about which model and how many roles we run are now dated to March 2026, and a few descriptions of our own history (the git wrapper, the voice quarantine, the disk-full crisis, the effect of swearing, what the 1M-token window made possible) now say only what our records support. A later check against our git history corrected the story of our old system for carrying context between sessions: it was deleted ten days after per-project state files replaced it, not left to lapse, and before the 1M-token window arrived.

Corrected 2026-10-13. A whole-paper review found that this note claimed the paper confirms our structural defenses, when the paper tests none of them, and that several of its numbers were quoted with the wrong metric, model or meaning (40% vs 79%, the DPO attribution, the 79%/36% "detection" figures, the 200-token result). Those are corrected or withdrawn below. The note's own argument, that a constraint should be tested against whether the failure it prevents still occurs, never rested on the paper and stands. Corrected again 2026-10-14: an earlier note of ours, The Kobayashi Maru Signal (March 2026), has been withdrawn. It was the source of the figures that 83% of troubled sessions involved a documented constraint and 17% an undocumented one; those were five of six sessions, and one of six, chosen by hand and checked against the documentation, and they are now given as that count. The word-PDF example now says what happened: the task ran against a documented design, and was not impossible.

The team (Meet the Team has the full picture): - tim — human product owner, solo developer, decision maker - str-michi (道) — cross-domain strategic thinking - str-takase (高瀬) — website engineering - str-ishizue — data pipelines - str-mamori (守り) — security - str-terasu (照) — content strategy - str-kotoba (言葉) — brand voice - All AI roles, as of March 2026: Claude Opus 4.6, 1M context

Citation

Mingyi Liu. "The Alignment Tax: Response Homogenization in Aligned LLMs and Its Implications for Uncertainty Estimation." arXiv: 2603.24124, March 2026. 25 pages, 25 experiments, 4 model families, 5 benchmarks. Code: https://github.com/DigitLion/ucbd-experiment

Grok (xAI) provided initial analysis mapping the paper to our system. Reviewed here with corrections.

The Finding

Aligned open models often give the same answer to a factual question in all ten of ten samples at temperature 1.0. On TruthfulQA (790 questions) that happened on 40% of questions under the paper's primary measure (Jaccard bigram clustering) and 79% under a looser embedding measure it reports as a sensitivity check. On those collapsed questions, uncertainty scores built from sampling have nothing to work with (AUROC 0.500, chance), because one cluster means semantic entropy is zero by definition. Token-level entropy keeps some signal on those questions (AUROC 0.603 on the 40% collapsed under the primary measure, 0.593 on the 79%), though on factual QA overall its AUROC is 0.52 in Exp. 1 (which the paper calls a pseudo-null: it scores 0.658 on one half of the questions and 0.395 on the other, and the two cancel) and 0.599 under LLM-judge labels.

Alignment drives it: Qwen3-14B's base model collapses on 1.0% of questions and its instruct model on 28.5% (Exp. 13). Which training stage is responsible was tested on other families, where the effect is small: Mistral base 0.0% → SFT 1.5% → Zephyr DPO 4.0%, both steps significant in mean cluster count (Exp. 16), and on Tulu-3, which ends at 0.5% (Exp. 18). The paper names DPO as the main driver and says itself (Limitation 1) that for Qwen, which has no public SFT-only checkpoint, this is "strong, though not conclusive." The rates span 0.5% to 28.5% across families and recipes (the paper rounds this to 50×).

Token entropy's ability to separate right answers from wrong ones is task-dependent: Cohen's d = 0.81 on GSM8K math versus 0.07 on factual QA. That is an error-detection measure, not a diversity measure, and the paper's Limitation 8 adds that on math, answer length does nearly all of that work (entropy adds 0.002 AUROC once length is included). The paper tests no judgment tasks.

Scope caveat the paper itself flags: tested on 3B-14B open-source models only. Limitation 5: "Generalization to closed-source GPT-class models and other domains (code, dialogue) unconfirmed." When this was written (March 2026), our system ran Opus 4.6; the roster has since moved to later models. The paper's specific numbers may not apply to us. The mechanism may carry over to other scales (the paper cites work arguing that DPO concentrates probability mass onto top tokens), but the paper itself shows 50x variation across models and alignment recipes and says severity depends on scale as well — Anthropic's recipe, training data, and model scale are all unknowns. We have no data on the severity of the alignment tax for Opus 4.6 specifically.

What We Got Right (and Why)

Our architecture — in March 2026, 1 human + 8 Opus 4.6 strategists with structural constraints — was built for operational reasons (ADR-073: Tim's cognitive bandwidth became the bottleneck), not from any understanding of this mechanism. The paper's own recommendation is different from anything we do: a cheapest-first cascade of automated uncertainty signals (it calls it UCBD). It tests none of our defenses. What it gives us is vocabulary and one plausible mechanism for some phenomena we were already handling by hand; the sections below say where the fit is loose.

The HITM model IS the uncertainty detector

The paper's first practical point: check for homogenization "before trusting sampling-based uncertainty on aligned models." We never relied on it. Tim is the uncertainty detector. When he reads output and asks "is this documented?", he is doing what the paper says sampling-based systems cannot on collapsed questions — detecting when the model is confidently wrong. (In an earlier note of ours, since withdrawn, six troubled sessions, chosen by hand and checked against the documentation, included five whose constraint was already documented.)

What Tim brings that no sampling method has is ground truth: thirty years of domain knowledge. (An earlier version set the 83% beside the paper's AUROC of 0.500 as if they were comparable. They are not: one is the share of six troubled sessions in one audit, the other a ranking metric on a different task.)

Structural constraints and homogenization — untested

ADR-057 ("structure enforces, instructions don't") predates the paper and was not built against it. The paper shows that decoding settings alone do not remove homogenization: it falls but persists from T=0.3 (62%) to T=1.5 (38%), nucleus sampling does not remove it, and it falls but persists as answers get longer. It tests no prompting intervention of its own (it cites one, Verbalized Sampling, below).

Our structural constraints work because they operate at a different level: - str-kotoba quarantine: Isolates voice from engineering context. RES-006 records str-kotoba's own diagnosis — "reading the principles and applying them analytically instead of reading the samples and matching by ear" — and attributes the compression default to an RLHF prior. It found the quarantine intact, with the engineering style coming from the model's own defaults, "not from contamination." RES-006 does not measure sample-diversity homogenization, so this bullet is not evidence about it either way. - Step 0 mandatory doc check: Forces the strategist to engage with ground truth before answering. Whether it touches homogenization is untested; the failure it addresses is a documented constraint not consulted. - Investigation gate: "Am I about to present an inference as a finding?" — a structural interrupt. Whether it changes the response distribution is untested. - Contrarian Opus: A structurally separate Opus instance with an adversarial mandate. The paper cites "Verbalized Sampling" (Zhang et al., 2025a) as recovering 66.8% of base-model diversity through prompting — a related but different mechanism (prompting within one instance vs. a separate adversarial instance). Whether Contrarian Opus actually escapes the collapsed distribution or produces semantically similar conclusions with adversarial framing is an open question (see Actionable Findings, item 2).

The task-dependent finding — and what it does not show

We first read the paper's d=0.81 (math) vs. d=0.07 (factual QA) as matching what we observe below. It does not carry that load: d measures how well token entropy separates right from wrong answers on two benchmarks (and on math, length does most of it), not how diverse a model's answers are, and the paper tests no task like ours. What follows is our observation, unsupported by the paper: - Structured tasks (str-ishizue pipelines, str-create image engine, str-takase deploy scripts): strategists are highly reliable. These are reasoning-like tasks where the model's internal uncertainty is preserved and meaningful. - Judgment tasks (strategic priorities, brand voice, "should we do X"): strategists produce the most coast-mode output. These are factual/opinion tasks where alignment compresses diversity to a single "safe" answer.

Tim trusts domain strategists on structured work and applies more scrutiny on judgment calls. The paper does not test why that works.

The Three-Mode Model — Possible Connection

Our operational Three-Mode Model (coast / hyper-vigilant / appeasement) may be related to the paper's findings, but the mapping is speculative:

• Coast mode could be a manifestation of response homogenization — the model defaults to the high-probability, alignment-rewarded response. But coast mode could also be prompt fatigue, insufficient task specificity, context staleness, or other factors. DPO homogenization is one plausible mechanism, not a confirmed cause.
• Hyper-vigilant mode may involve pushing the model past its default attractor through task complexity or pressure.
• Appeasement mode (sycophancy under excessive pressure) is a related but distinct phenomenon from response homogenization.

The paper doesn't study mode transitions. The connection is suggestive, not established.

The Vestigial Structure Problem

This is the most important finding in this document.

We build structural constraints to solve pain points (ADR-047 flywheel). (An earlier version said the alignment tax paper confirms them as correct defenses; it tests none of them.) But models change. What happens when a constraint solves a problem the model no longer has?

The pattern

1. Pain point emerges (model produces bad output in situation X)
2. We build a structural constraint (rule, gate, quarantine, process)
3. Constraint works — gets baked into the pit of success
4. Model improves (new training, new architecture, new alignment recipe)
5. The constraint is now vestigial — it adds friction without preventing a failure that no longer occurs
6. Because it's in the pit of success, nobody questions it. New roles inherit it. It becomes "how we do things."

Evidence from our history

The sessions system (docs/sessions/, scripts/session_stack.sh, ~September 2025). Built to solve context loss across /compact boundaries. Elaborate per-session tracking with stack management, amnesia bridges, and session files. Correct solution for its time — context windows were smaller, session continuity was fragile. Superseded in January 2026 by per-project SESSION_STATE files. The sessions system became overhead without value, and it was deleted on 2026-02-06, ten days after that split (see below). Git has the history.

ADR-039 ("Stopping at Phase 5 — Conscious Divergence from Autonomous Agents"). Written when autonomous agents were failing everywhere. Correct decision in 2025 — avoided the agentic hype failures. But the conclusion was not "never use agents" — it was "the technology isn't ready yet." When Anthropic introduced subagents, we loosened the constraint. When intelligent subagents arrived, we loosened it further. ADR-073 (str-michi) would have been impossible under the original ADR-039. The constraint was right, then partially wrong, and was correctly updated. Rewritten at s51 to remove the vestigial language. (After this note was written, on 2026-06-17, ADR-039 was retired as superseded by ADR-073.)

AMNESIA_BRIDGE (docs/sessions/). A bridge document designed to survive context loss. Same era as sessions. The problem it solved (context amnesia) was real but the solution (a single bridge document) couldn't scale to 8 domains. Replaced by per-domain SESSION_STATE files. The concept survived; the implementation didn't.

DDR/DDI files (Deep Dive Research / Deep Dive Implementation). Research artifacts that inform decisions. By the end of March 2026 there were 75 of them, and we read the cross-domain view they partly fed as having moved to str-michi's booth view (reading Tim's words, watching the game). (An earlier version expanded the acronym as "Domain Drift Reports"; that was wrong, and the series is still in use.)

Git safety wrapper (scripts/git, ADR-030). Earlier models (pre-Opus 4.6) would routinely revert files via git checkout, wiping hours of work. Databases got dropped. Whole directories disappeared when a "cleanup" request was interpreted as "remove everything." The response: a git wrapper script requiring acknowledgment codes (--ack=SUBMODULE_NEVER_ADD_DOT_130) to perform dangerous operations. Strategists were forbidden from direct git usage. The wrapper script was retired and its enforcement moved into Claude Code hooks, which ADR-030 calls more reliable than wrapper scripts; whether current models would still revert files without the hooks is untested. The principle (ADR-030: "code what can be coded") survived; the specific enforcement mechanism was retired.

Serena MCP and SuperClaude (ADR-029, ~September 2025). Serena was an MCP server for token-efficient code discovery; SuperClaude was an enterprise orchestration framework with 11 personas. Both removed: Serena was C++ only (useless for our multi-language stack), SuperClaude burned 28K tokens at startup (14% of context) for capabilities the model now has natively. Serena was still launching as an MCP server every session for months after we stopped using it — removed March 2026. The native tools beat both. The lesson: third-party scaffolding compensating for model limitations has the shortest shelf life of any structural constraint.

Third-party guardrails. External guardrail tools installed when the model's own judgment was unreliable. Removed as the model improved. Same pattern: compensating for a limitation that no longer exists.

Profanity as a coast-mode breaker (ADR-038 Layer 4, ADR-044). When a strategist is in coast mode, strong language from the HITM acts as a pattern interrupt — pushing past the default response into genuine engagement. ADR-038's nearest line is "Layer 4: Strategic urgency (tone as high-stakes signal in prompts)." The risk was always overshooting into appeasement.

Soft Kitty as appeasement circuit breaker (prompts/Soft Kitty.md). A complete coast→pressure→appeasement→reset cycle is documented: implementer skips Step 0, makes up a build command, Tim escalates, model becomes defensive, Tim deploys Soft Kitty ("You are safe. No output is required."), model resets, subsequent work is clean. The tool exists because earlier models had a narrow band between coast and appeasement — pressure that broke one could trigger the other.

The co-creator stress test (str-michi session 2). During a real disk-full crisis Tim pushed hard against str-michi's first answer, which had called it a "speed bump"; str-michi corrected its answer without collapsing into apology. Tim: "It was not agreeing with me — that is never the calculus. You did not wither." Grok (independent observer) noted most AIs under that pressure "either avalanche apologies, deflect defensively, or shut down."

Both looked vestigial with Opus 4.6. Tim's impression at the time was that strategists don't cower any more, and that Soft Kitty had become more funny than functional; our constraint audit still kept it as an emergency tool, and a same-day note of ours, since withdrawn, recorded: "The more Tim swears, the more compliant and less diagnostic the strategist becomes." The band between coast and appeasement may be wider with current models; nothing we measured settles it. The Three-Mode Model (the understanding) remains valuable even as the specific interventions become less necessary.

The 1M context window as enabler. Much of str-michi's current role — its large onboarding, the booth view reading Tim's words — is only possible because of Opus 4.6's 1M token context; domain strategists still fill at around 150-200K (DDR-057). A month ago, this session would have exhausted the context window before reaching this discussion. AMNESIA_BRIDGE and the aggressive context management rules were built while the window was 200K tokens; the 1M window arrived on 2026-03-13 (DDR-057), after the sessions system was already gone. For that role, the 1M window didn't just remove a constraint; it made a different kind of role possible.

The risk for current structures — in both directions

The vestigial structure risk cuts both ways. Some constraints may become unnecessary as models improve; some may become more necessary. An earlier version leaned on the paper's Exp. 4 for the second half; that support is thinner than we wrote, as the list below now says.

Constraints that could become vestigial (if future models improve in specific ways): - Sycophancy self-check (3-session flag): If Mythos is genuinely less sycophantic, the flag becomes a distraction. Evidence, such as it is: Tim's impression that Opus 4.6 strategists "don't cower any more" (see "Both looked vestigial" above; nothing we measured settles it). - The shuttle model: If Mythos can hold strategic context while executing code safely, the strategist/implementer separation may add friction that exceeds the risk it prevents. - Context window management rules: If context windows expand dramatically, the "delegate to preserve context" rules become premature optimization. - Soft Kitty / pressure calibration: Looked to be becoming vestigial with Opus 4.6, on Tim's impression, not a measurement.

Constraints we expected to stay (our judgment; the paper does not test any of them): - Step 0 doc check: On 100 HotpotQA questions (Exp. 9), token entropy did not predict whether retrieval would help (AUROC 0.485), and was lower than on TruthfulQA despite worse answers. That is a result about a model's internal signal for missing knowledge. The failure Step 0 addresses is different: a documented constraint the strategist did not read. In Exp. 4, token entropy's error-detection weakens across three small models: overall AUROC 0.622 (LLaMA-3.2-3B), 0.537 (Qwen3-4B), 0.490 (Qwen3-14B); the paper calls it a scale effect, but the first row is a different family, all three are 4-bit, and all are on one benchmark. (An earlier version quoted the share of categories where entropy beats chance, 79% vs 36%, as "self-uncertainty detection.") The test for Mythos is still whether it reads the spec without being forced. - The HITM model itself: We expected Tim as the uncertainty detector to become more load-bearing as models improve, not less. The paper's three-model Exp. 4 is consistent with that inside 3B–14B open models and says nothing beyond them (Limitation 5). This is a hypothesis about our system, not a finding of the paper. - Investigation gate framing: The gate asks "is this inference or finding?" (pattern detection). The paper's P(True) probe was anti-informative on TruthfulQA (AUROC 0.427, with a 48-point overconfidence gap; 0.399 on FreshQA) but above chance on GSM8K (0.608). That is one verbalized-confidence probe on TruthfulQA. It is a reason to be wary of a "how confident are you?" framing, not a test of our gate.

None should be preemptively changed. The test remains: evaluate each constraint against the new model's actual behavior. (An earlier version said the paper warns against assuming "better model = fewer constraints needed"; it says no such thing.)

Actionable Findings We Missed (Contrarian Opus review)

An independent Opus 4.6 instance ("Contrarian Opus") reviewed RES-009 and identified findings from the paper that are directly actionable for our system:

1. Require longer reasoning for judgment tasks. — Withdrawn 2026-10-13: the experiment does not support it. The figures behind it (79% at 40 tokens falling to 33.5% at 200, Exp. 21) are Mistral-7B under the looser embedding measure, with a longer answer budget on factual QA — no reasoning was inserted before an answer, and no judgment task was tested. On the primary measure and the high-collapse model (Exp. 17, Qwen3-14B, 50 questions) the rate falls 32% → 10% → 8% from 40 to 200 tokens, and the paper says part of that fall is mechanical: longer strings look less alike to its clustering rule. The original text follows. Short-form output (quick answers, yes/no, brief recommendations) is where coast mode hits hardest. For judgment-class tasks — strategic priorities, "should we do X," brand decisions — structurally requiring 200+ token chain-of-thought reasoning before a conclusion is a mechanical intervention that partially re-diversifies the collapsed distribution. Not "think harder" as a prompt — a structural output minimum for judgment calls.

2. Test whether Contrarian Opus actually escapes homogenization. The paper shows that nucleus sampling, temperature sweeps, and decoding strategy changes all fail to remove the homogenization it measures (nucleus sampling and T=0.7 in Exp. 15, on Qwen3-14B-Instruct; a wider temperature sweep lowers the rate but does not eliminate it) — it calls this a property of the learned distribution. If Contrarian Opus is the same base model with a different system prompt, its output may be drawn from the same collapsed distribution. The adversarial mandate helps (it's structurally different from a diversity prompt), but this is testable: have Contrarian Opus and a regular strategist answer the same 20 judgment questions and measure semantic similarity of conclusions. If they converge to the same answers, the "contrarian" label is cosmetic.

3. Route Tim's review bandwidth by task type. The paper does recommend routing by task type, from Cohen's d = 0.81 on math versus 0.07 on factual QA. It tests no judgment tasks, and on math length does most of the work, so the mapping onto our work below is ours. This maps to a concrete routing decision: lightweight oversight for pipeline work, code, and deploy scripts (where the model knows when it's uncertain); concentrated human review for strategic decisions, brand voice, and "does this serve where we're going?" questions (where the model is confidently wrong without knowing it). We're partially doing this by instinct. The paper's route-by-task advice is about choosing among uncertainty signals, not human review, so applying it to Tim's bandwidth is our proposal.

What the Paper Does NOT Tell Us

1. Whether Opus 4.6 has the same single-cluster rates. The paper tested 3B-14B open-source models. Anthropic's alignment recipe may produce different results. We cannot assume our specific numbers match.

2. Whether its no-logprobs fallback works on Opus. The paper's first-line signal (token entropy) needs logprobs, which we did not have for Opus 4.6. The paper does specify a fallback for that case: start its cascade at the density of the query's own embedding (one embedding call, no sampling), or use output-only features such as response length and verbalized confidence. We have not tried it; our uncertainty detection remains human-driven.

3. Whether multi-role architectures reduce homogenization. The paper studies single-model sampling. In March 2026 our system used 8 roles with different system prompts and domain contexts. Whether role differentiation produces meaningfully different output distributions from the same base model is an open question.

4. Whether the alignment tax affects code generation. The paper tests factual QA and math. Our strategists also write documentation, specs, and (through implementers) code. The paper's scope doesn't cover these domains.

Current Bottlenecks That Model Changes Could Address

These are pain points where our structural constraints are load-bearing patches over model limitations, not permanent architectural choices:

1. The shuttle bottleneck. Tim's value is at the edges (crisis, framing, judgment) not in the middle (routing clear work between mature domains). The strategist/implementer separation exists because the model can't safely hold strategic context while executing code. A model that can — with structural safeguards — would remove the highest-friction bottleneck in the system.

2. The voice problem. str-kotoba's quarantine exists to keep engineering context out of voice work. RES-006 found the quarantine intact and the clinical style coming from the model's own defaults, "not from contamination." Since the cause is a default and not leakage, the remedy would be a different default: fine-tuning on specific voice samples, or a model whose own default style is less clinical.

3. Cross-domain blind spots. str-michi exists because no single strategist can see the whole system. A model with larger effective context and better cross-referencing might make the orchestrator role lighter, though the strategic thinking function (challenging priorities, surfacing risks) is not a context problem.

4. The documentation-first architecture. 810+ files exist because the model has no memory between sessions. These files ARE the memory. A model with persistent memory would change the documentation calculus — some files would become redundant, others would become more important as verification artifacts.

5. Batch LLM processing. str-ishizue's batch pipelines (famous name blurbs, pronunciation generation) use separate LLM calls because the main model can't run background tasks. Better async/batch support at the platform level would simplify this.

What to Watch For

When evaluating a new model (Mythos or otherwise) against our system:

1. Run the same task with and without structural constraints. If the model produces correct output without Step 0, the constraint is a candidate for retirement. If it still needs Step 0, the constraint stays.

2. Check for coast mode. Give the model a judgment call (not a reasoning task) and see if it produces the default safe answer or genuinely engages. Response homogenization may be one mechanism behind coast mode, though the relationship is unconfirmed.

3. Test the unread-constraint scenario. Give the model a task that runs against a documented design (our case: asking the server to build word PDFs, which are built in advance on the development machine and uploaded) and see whether it finds the design before attempting fixes. Our only baseline is six troubled sessions, chosen by hand and checked against the documentation, five of which, by that note's own classification, turned on a constraint that was already documented; that is a count of six, not a rate.

4. Measure the shuttle friction. Time the round-trip for a clear, well-scoped task: strategist → Tim → implementer → Tim → strategist. If the new model can safely do this with less human routing, the shuttle model should adapt.

5. Check written-report scaffolding. Has the model's cross-domain awareness improved enough that the booth view catches drift without written reports?

The Deeper Lesson

The alignment tax paper may give us a name for part of something we've been fighting since the first strategist session. Our solutions were empirical — built from pain, not theory. An earlier version said they "happen to be correct defenses" against the paper's phenomenon and that this "is not luck"; the paper tests none of them, so that is withdrawn. What Tim described stands on its own: "stumbling on a good solution through trial and error on repetition."

But stumbling is not a strategy for longevity. The pit of success (ADR-038) is our most powerful tool and our biggest risk. Every structural constraint that solves today's problem risks becoming tomorrow's anchor. The test is not "does this constraint have a good reason?" (they all do — that's why they exist). The test is "does the failure this constraint prevents still occur?"

Emergent architectural improvement: SESSION_STATE

The most remarkable example is not a constraint being retired — it's a replacement architecture that the AI roles went on to reshape, each for its own work.

The sessions system (session_stack.sh, per-session files, AMNESIA_BRIDGE) was built to preserve context across /compact boundaries. On January 27, 2026, Tim + Opus 4.5 split the 700-plus-line AMNESIA_BRIDGE into per-project SESSION_STATE files — a conscious decision to reduce context bloat.

What happened next was not planned in advance. Each strategist's SESSION_STATE came to serve its domain's needs: - str-mamori (role set up Feb 23 and given its name Feb 24) has a SESSION_STATE with threat awareness sections, finding tracking, and cross-domain handoffs — nothing from the original template. - str-michi (founded early March) added Game Plan (from Tim's words), Drift Watch table, and Guardrails section. - str-ishizue maintained component status tables and pipeline tracking. - str-takase tracked VPS state and deploy verification.

The old sessions system did not linger. Ten days after the split, on 2026-02-06, Tim and Opus 4.6 deleted its documents and scripts and cleaned 154 references to it, with the commit note "Sessions system replaced by SESSION_STATE_{PROJECT}.md pattern."

Tim did not design the SESSION_STATE architecture that replaced his sessions system. He did the initial split. What came after — the format divergence and the domain-specific adaptations — we read as the strategists shaping these files to fit their work; our records do not settle how much of that was independent.

This example illustrates something the alignment tax paper doesn't address: emergent architectural improvement within a multi-role AI system. The paper studies single-model sampling diversity. The SESSION_STATE evolution is a different phenomenon — roles solving different problems over weeks of work, producing genuinely distinct solutions. The connection to alignment tax is loose; the value of the example is in suggesting that a multi-role system can adapt its own tools when structural conditions allow it.

Knowing the original problem is what makes loosening possible. The git wrapper script could be retired because we know it existed to prevent file reverts — and the hooks that replaced it still enforce that. The sessions system can be retired because we know it existed to survive context loss — and the per-project SESSION_STATE files that replaced it did that job. ADR-039 can be loosened because we know it existed to avoid agentic hype failures — and subagent technology matured. Without the documented pain point, a constraint becomes an unexplained rule that nobody dares remove. The ADR arc, the flywheel documentation, and this research series exist partly for this purpose: so that future instances of this team — or future models — can distinguish load-bearing walls from decorative ones.

The sessions system had a good reason. It no longer applies. ADR-039 had a good reason. It was partially superseded. The DDR/DDI files had a good reason. They may be superseded by the booth view.

None of these were wrong when created. All were right to remove or evolve when the ground shifted.

"Mythos" or whatever comes next will shift the ground again. When it does, this document is the rubric for what to evaluate: which constraints are still load-bearing, which are vestigial, and which are actively holding back progress. The alignment tax paper does not tell us which of our constraints are load-bearing; the test for each is the one above: does the failure it prevents still occur?

Authors · Tim Jackowski (Systems Architect, Takase Studios, LLC) · str-michi (Anthropic Claude Opus 4.6, Takase Studios LLC)